chore: import upstream snapshot with attribution

tcp_server/src/lib/asr/api.py

@@ -0,0 +1,215 @@
+import audioop
+import time
+import torch
+import numpy as np
+from faster_whisper import WhisperModel
+
+from ..constants import ASR_MODEL_PATH
+from ..utils import ThrottledCallback, is_macos, get_settings
+from ..audio_input import AudioInputStream
+
+
+class ASR:
+    def __init__(self,
+                 tcp_server=None,
+                 # @see https://github.com/SYSTRAN/faster-whisper/blob/master/faster_whisper/transcribe.py
+                 # auto, cpu, cuda
+                 device='auto',
+                 interrupt_leon_speech_callback=None,
+                 transcribed_callback=None,
+                 end_of_owner_speech_callback=None,
+                 active_listening_disabled_callback=None):
+        tic = time.perf_counter()
+        self.log('Loading model...')
+
+        if device == 'auto':
+            if torch.cuda.is_available():
+                device = 'cuda'
+                self.log('Using CUDA (Compute Unified Device Architecture)')
+
+        if 'cuda' in device:
+            assert torch.cuda.is_available()
+
+        self.log(f'Device: {device}')
+
+        compute_type = 'float16'
+        if is_macos():
+            compute_type = 'int8_float32'
+
+        if device == 'cpu':
+            compute_type = 'int8_float32'
+
+        self.tcp_server = tcp_server
+        self.wake_word = None
+        self.compute_type = compute_type
+        self.is_recording = False
+
+        """
+        Thottle the interrupt Leon's speech callback to avoid sending too many messages to the client
+        """
+        self.interrupt_leon_speech_callback = ThrottledCallback(
+            interrupt_leon_speech_callback, 0.8
+        )
+        self.transcribed_callback = transcribed_callback
+        self.end_of_owner_speech_callback = end_of_owner_speech_callback
+        self.active_listening_disabled_callback = active_listening_disabled_callback
+
+        self.device = device
+        self.is_voice_activity_detected = False
+        self.silence_start_time = 0
+        self.is_active_listening_enabled = False
+        self.complete_text = ''
+
+        self.buffer = bytearray()
+        self.silence_frames_count = 0
+        self.channels = 1
+        self.rate = 16000
+        self.frames_per_buffer = 1024
+        self.rms_threshold = get_settings('asr')['rms_mic_threshold']
+        # Duration of silence after which the audio data is considered as a new utterance (in seconds)
+        self.silence_duration = get_settings('asr')['silence_duration']
+        """
+        Duration of silence after which the active listening is stopped (in seconds).
+        Once stopped, the active listening can be resumed by starting a new recording event
+        """
+        self.base_active_listening_duration = get_settings('asr')['active_listening_duration']
+        self.active_listening_duration = self.base_active_listening_duration
+
+        self.mic_stream = None
+        self.model = None
+
+        model_params = {
+            'model_size_or_path': ASR_MODEL_PATH,
+            'device': self.device,
+            'compute_type': self.compute_type,
+            'local_files_only': True
+        }
+        if self.device == 'cpu':
+            model_params['cpu_threads'] = 4
+
+        self.open_mic_stream()
+
+        self.model = WhisperModel(**model_params)
+
+        self.log('Model loaded')
+        toc = time.perf_counter()
+
+        self.log(f'Time taken to load model: {toc - tic:0.4f} seconds')
+
+    def open_mic_stream(self):
+        try:
+            self.mic_stream = AudioInputStream(
+                channels=self.channels,
+                rate=self.rate,
+                frames_per_buffer=self.frames_per_buffer
+            )
+            self.mic_stream.open()
+        except Exception as e:
+            self.log('Error to open mic stream:', e)
+
+    def start_recording(self):
+        if self.wake_word:
+            # Make sure to stop the wake word detection before recording
+            # otherwise it will loop for the wake word and create conflict
+            # on the audio stream
+            self.wake_word.stop_listening()
+
+        self.is_recording = True
+        # Convert the silence duration to the number of audio frames required to detect the silence
+        silence_threshold = int(self.silence_duration * self.rate / self.frames_per_buffer)
+
+        try:
+            self.log('Recording...')
+
+            while self.is_recording:
+                data = self.mic_stream.read(self.frames_per_buffer, exception_on_overflow=False)
+                rms = audioop.rms(data, 2)  # width=2 for signed 16-bit PCM
+
+                if rms >= self.rms_threshold:
+                    if not self.is_voice_activity_detected:
+                        self.is_active_listening_enabled = True
+                        self.is_voice_activity_detected = True
+
+                    self.interrupt_leon_speech_callback()
+
+                    self.buffer.extend(data)
+                    self.silence_frames_count = 0
+                else:
+                    if self.is_voice_activity_detected:
+                        self.silence_start_time = time.time()
+                        self.is_voice_activity_detected = False
+
+                    if self.silence_frames_count < silence_threshold:
+                        self.silence_frames_count += 1
+                    else:
+                        if len(self.buffer) > 0:
+                            self.log('Silence detected')
+
+                            audio_data = np.frombuffer(self.buffer, dtype=np.int16)
+                            if self.compute_type == 'int8_float32':
+                                audio_data = audio_data.astype(np.float32) / 32768.0
+                            transcribe_params = {
+                                'beam_size': 5,
+                                'language': 'en',
+                                'task': 'transcribe',
+                                'condition_on_previous_text': False,
+                                'hotwords': 'talking to Leon'
+                            }
+                            if self.device == 'cpu':
+                                transcribe_params['temperature'] = 0
+                            segments, info = self.model.transcribe(audio_data, **transcribe_params)
+
+                            for segment in segments:
+                                self.log('[%.2fs -> %.2fs] %s' % (segment.start, segment.end, segment.text))
+                                self.complete_text += segment.text
+
+                            self.transcribed_callback(self.complete_text)
+                            time.sleep(0.1)
+                            # Notify the end of the owner's speech
+                            self.end_of_owner_speech_callback(self.complete_text)
+
+                            self.complete_text = ''
+                            self.buffer = bytearray()
+
+                        should_stop_active_listening = self.is_active_listening_enabled and time.time() - self.silence_start_time > self.active_listening_duration
+                        if should_stop_active_listening:
+                            self.is_active_listening_enabled = False
+                            self.active_listening_disabled_callback()
+                            # Do not add anything after this line because it will be ignored
+                            # as it loops for the wake word
+                            self.stop_recording()
+        except Exception as e:
+            self.log('Error:', e)
+
+    def stop_recording(self):
+        self.log('Recording stopped')
+
+        if self.wake_word:
+            self.wake_word.reset_model_state()
+            if self.wake_word.is_enabled:
+                # Do not add anything after this line because it will be ignored
+                # as it loops for the wake word
+                self.wake_word.start_listening()
+        else:
+            self.is_recording = False
+            # self.mic_stream.stop_stream()
+            # self.mic_stream.close()
+            # self.log('Stream closed, recording stopped')
+
+        # Make sure to wait for the recording thread to join before starting a new recording.
+        # Only needed when the wake word detection is enabled
+        if (
+            self.wake_word and
+            self.wake_word.is_enabled and
+            self.tcp_server.asr_recording_thread and
+            self.tcp_server.asr_recording_thread.is_alive()
+        ):
+            # The thread is only used when received TCP message from the core,
+            # hence it is not used when triggered by the wake word.
+            # If we do not "join" it, it'll duplicate the recording loop
+            self.log('Join recording thread')
+            self.tcp_server.asr_recording_thread.join()
+
+    @staticmethod
+    def log(*args, **kwargs):
+        print('[ASR]', *args, **kwargs)

tcp_server/src/lib/audio_input.py

@@ -0,0 +1,57 @@
+import numpy as np
+import soundcard as sc
+
+
+class AudioInputStream:
+    def __init__(self, channels=1, rate=16000, frames_per_buffer=1024):
+        self.channels = channels
+        self.rate = rate
+        self.frames_per_buffer = frames_per_buffer
+        self.sample_width = 2
+        self._microphone = None
+        self._recorder = None
+        self._is_open = False
+        self.device_name = None
+
+    def open(self):
+        if self._is_open:
+            return
+
+        microphones = sc.all_microphones(include_loopback=False)
+
+        try:
+            self._microphone = sc.default_microphone()
+        except Exception as error:
+            if not microphones:
+                raise RuntimeError('No capture device found') from error
+
+            self._microphone = microphones[0]
+
+        self.device_name = str(self._microphone)
+        self._recorder = self._microphone.recorder(
+            samplerate=self.rate,
+            channels=self.channels,
+            blocksize=self.frames_per_buffer
+        )
+        self._recorder.__enter__()
+        self._is_open = True
+
+    def read(self, frames, exception_on_overflow=False):
+        del exception_on_overflow
+        audio_data = self._recorder.record(numframes=frames)
+        audio_data = np.clip(audio_data, -1.0, 1.0)
+        pcm_data = (audio_data * 32767.0).astype(np.int16)
+
+        return pcm_data.tobytes()
+
+    def stop_stream(self):
+        return
+
+    def close(self):
+        self._is_open = False
+
+        if self._recorder:
+            self._recorder.__exit__(None, None, None)
+            self._recorder = None
+
+        self._microphone = None

tcp_server/src/lib/constants.py

@@ -0,0 +1,52 @@
+import os
+import sys
+
+IS_RAN_FROM_BINARY = getattr(sys, 'frozen', False)
+
+DEFAULT_LEON_PROFILE = 'just-me'
+
+
+def resolve_leon_codebase_path():
+    configured_codebase_path = os.getenv('LEON_CODEBASE_PATH', '').strip()
+
+    return os.path.abspath(configured_codebase_path or os.getcwd())
+
+
+def resolve_leon_home():
+    configured_leon_home = os.getenv('LEON_HOME', '').strip()
+
+    if configured_leon_home:
+        return os.path.abspath(configured_leon_home)
+
+    return os.path.join(os.path.expanduser('~'), '.leon')
+
+
+EXECUTABLE_DIR_PATH = os.path.dirname(sys.executable) if IS_RAN_FROM_BINARY else '.'
+
+CODEBASE_PATH = resolve_leon_codebase_path()
+LEON_HOME_PATH = resolve_leon_home()
+
+LIB_PATH = os.path.join(CODEBASE_PATH, 'tcp_server', 'src', 'lib')
+if IS_RAN_FROM_BINARY:
+    LIB_PATH = os.path.join(os.path.dirname(sys.executable), 'lib', 'lib')
+
+PYTHON_VERSION = '3.11'
+
+TMP_PATH = os.path.join(LIB_PATH, 'tmp')
+AUDIO_MODELS_PATH = os.path.join(LEON_HOME_PATH, 'models', 'audio')
+SETTINGS_PATH = os.path.join(CODEBASE_PATH, 'tcp_server', 'settings.json')
+
+# TTS
+TTS_MODEL_FOLDER_PATH = os.path.join(AUDIO_MODELS_PATH, 'tts')
+TTS_BERT_FRENCH_MODEL_DIR_PATH = os.path.join(TTS_MODEL_FOLDER_PATH, 'bert-case-french-europeana-cased')
+TTS_BERT_BASE_MODEL_DIR_PATH = os.path.join(TTS_MODEL_FOLDER_PATH, 'bert-base-uncased')
+TTS_MODEL_CONFIG_PATH = os.path.join(TTS_MODEL_FOLDER_PATH, 'config.json')
+IS_TTS_ENABLED = os.environ.get('LEON_TTS', 'true') == 'true'
+
+# ASR
+ASR_MODEL_PATH = os.path.join(AUDIO_MODELS_PATH, 'asr')
+IS_ASR_ENABLED = os.environ.get('LEON_ASR', 'true') == 'true'
+
+# Wake word
+WAKE_WORD_MODEL_FOLDER_PATH = os.path.join(AUDIO_MODELS_PATH, 'wake_word')
+IS_WAKE_WORD_ENABLED = os.environ.get('LEON_WAKE_WORD', 'true') == 'true'

tcp_server/src/lib/tcp_server.py

@@ -0,0 +1,254 @@
+import socket
+import json
+import os
+from typing import Union
+import time
+import re
+import threading
+
+from .utils import get_settings
+from .wake_word.api import WakeWord
+from .asr.api import ASR
+from .tts.api import TTS
+from .constants import (
+    TTS_MODEL_CONFIG_PATH,
+    TTS_MODEL_FOLDER_PATH,
+    WAKE_WORD_MODEL_FOLDER_PATH,
+    IS_WAKE_WORD_ENABLED,
+    IS_TTS_ENABLED,
+    TMP_PATH,
+    IS_ASR_ENABLED
+)
+
+TTS_MODEL_PATH = os.path.join(TTS_MODEL_FOLDER_PATH, get_settings('tts')['model_file_name'])
+
+
+class TCPServer:
+    def __init__(self, host: str, port: Union[str, int]):
+        self.host = host
+        self.port = port
+        self.tcp_socket = socket.socket(socket.AF_INET, socket.SOCK_STREAM)
+        self.conn = None
+        self.addr = None
+        self.wake_word = None
+        self.tts = None
+        self.asr = None
+        self.asr_recording_thread = None
+
+    @staticmethod
+    def log(*args, **kwargs):
+        print('[TCP Server]', *args, **kwargs)
+
+    def send_tcp_message(self, data: dict):
+        if not self.conn:
+            self.log('No client connection found. Cannot send message')
+            return
+
+        self.conn.sendall(json.dumps(data).encode('utf-8'))
+
+    def init_tts(self):
+        if not IS_TTS_ENABLED:
+            self.log('TTS is disabled')
+            return
+
+        if not os.path.exists(TTS_MODEL_CONFIG_PATH):
+            self.log(f'TTS model config not found at {TTS_MODEL_CONFIG_PATH}')
+            return
+
+        if not os.path.exists(TTS_MODEL_PATH):
+            self.log(f'TTS model not found at {TTS_MODEL_PATH}')
+            return
+
+        self.tts = TTS(language='EN',
+                       device=get_settings('tts')['device'],
+                       config_path=TTS_MODEL_CONFIG_PATH,
+                       ckpt_path=TTS_MODEL_PATH)
+
+    def init_asr(self):
+        if not IS_ASR_ENABLED:
+            self.log('ASR is disabled')
+            return
+
+        def transcribed_callback(text):
+            # cleaned_text = clean_up_speech(text)
+            self.log('Cleaned speech:', text)
+            self.send_tcp_message({
+                'topic': 'asr-new-speech',
+                'data': {
+                    'text': text
+                }
+            })
+
+        def interrupt_leon_speech_callback():
+            self.log('Interrupting Leon speech because owner started speaking')
+            self.send_tcp_message({
+                'topic': 'asr-interrupt-leon-speech',
+                'data': {}
+            })
+
+        def end_of_owner_speech_callback(utterance):
+            self.log('End of owner speech:', utterance)
+            self.send_tcp_message({
+                'topic': 'asr-end-of-owner-speech-detected',
+                'data': {
+                    'utterance': utterance
+                }
+            })
+
+        def active_listening_disabled_callback():
+            self.log('Active listening disabled')
+            self.send_tcp_message({
+                'topic': 'asr-active-listening-disabled',
+                'data': {}
+            })
+
+        self.asr = ASR(tcp_server=self,
+                       device=get_settings('asr')['device'],
+                       interrupt_leon_speech_callback=interrupt_leon_speech_callback,
+                       transcribed_callback=transcribed_callback,
+                       end_of_owner_speech_callback=end_of_owner_speech_callback,
+                       active_listening_disabled_callback=active_listening_disabled_callback)
+
+        if not IS_WAKE_WORD_ENABLED:
+            self.log('Wake word is disabled')
+            return
+
+        wake_word_model_name = get_settings('wake_word')['model_file_name']
+        wake_word_model_path = os.path.join(WAKE_WORD_MODEL_FOLDER_PATH, wake_word_model_name)
+
+        self.asr.wake_word = WakeWord(
+            asr=self.asr,
+            model_path=wake_word_model_path,
+            device=get_settings('wake_word')['device'],
+            detection_threshold=get_settings('wake_word')['detection_threshold']
+        )
+        # Do not add anything after this line because it will be ignored
+        # as it loops for the wake word
+        self.asr.wake_word.start_listening()
+
+    def init(self):
+        try:
+            # Make sure to establish TCP connection by reusing the address so it does not conflict with port already in use
+            self.tcp_socket.setsockopt(socket.SOL_SOCKET, socket.SO_REUSEADDR, 1)
+            self.tcp_socket.bind((self.host, int(self.port)))
+            self.tcp_socket.listen()
+        except OSError as e:
+            # If the port is already in use, close the connection and retry
+            if 'Address already in use' in str(e):
+                self.log(f'Port {self.port} is already in use. Disconnecting client and retrying...')
+                if self.conn:
+                    self.conn.close()
+                # Wait for a moment before retrying
+                time.sleep(1)
+                self.init()
+            else:
+                raise
+
+        while True:
+            # Flush buffered output to make it IPC friendly (readable on stdout)
+            self.log('Waiting for connection...', flush=True)
+
+            # Our TCP server only needs to support one connection
+            self.conn, self.addr = self.tcp_socket.accept()
+
+            try:
+                self.log(f'Client connected: {self.addr}')
+
+                while True:
+                    # socket_data = self.conn.recv(1024)
+                    socket_data = self.conn.recv(8096)
+
+                    if not socket_data:
+                        break
+
+                    data_dict = json.loads(socket_data)
+
+                    # Verify the received topic can execute the method
+                    method = data_dict['topic'].lower().replace('-', '_')
+                    if hasattr(self.__class__, method) and callable(getattr(self.__class__, method)):
+                        data = data_dict['data']
+                        method = getattr(self, method)
+                        res = method(data)
+
+                        self.send_tcp_message(res)
+                    else:
+                        self.log(f'Method "{method}" not found')
+            finally:
+                self.log(f'Client disconnected: {self.addr}')
+                self.conn.close()
+
+    def asr_start_recording(self, extra=None) -> dict:
+        # If ASR is not initialized yet, then wait for 2 seconds before starting recording
+        if not self.asr:
+            self.log('ASR is not initialized yet. Waiting for 2 seconds before starting recording...')
+            time.sleep(2)
+
+        if self.asr.is_recording is False:
+            self.asr_recording_thread = threading.Thread(target=self.asr.start_recording)
+            self.asr_recording_thread.start()
+
+        return {
+            'topic': 'asr-started-recording',
+            'data': {}
+        }
+
+    def tts_synthesize(self, speech: str) -> dict:
+        # If TTS is not initialized yet, then wait for 2 seconds before synthesizing
+        if not self.tts:
+            self.log('TTS is not initialized yet. Waiting for 2 seconds before synthesizing...')
+            time.sleep(2)
+
+        """
+        TODO:
+        - Implement one speaker per style (joyful, sad, angry, tired, etc.)
+        - Need to train a new model with default voice speaker and other speakers with different styles
+        - EN-Leon-Joyful-V1; EN-Leon-Sad-V1; etc.
+        """
+        speaker_ids = self.tts.hps.data.spk2id
+        # Random file name to avoid conflicts
+        audio_id = f'{int(time.time())}_{os.urandom(2).hex()}'
+        output_file_name = f'{audio_id}.wav'
+        output_path = os.path.join(TMP_PATH, output_file_name)
+        speed = 1
+
+        formatted_speech = speech.replace(' - ', '.').replace(',', '.').replace(': ', '. ')
+        # Clean up emojis
+        formatted_speech = re.sub(r'[\U00010000-\U0010ffff]', '', formatted_speech)
+        formatted_speech = formatted_speech.strip()
+        # formatted_speech = speech.replace(',', '.').replace('.', '...')
+
+        # TODO: should not wait to finish for streaming support
+        self.tts.tts_to_file(
+            formatted_speech,
+            speaker_ids['EN-Leon-V1_1'],
+            output_path=output_path,
+            speed=speed,
+            quiet=True,
+            format='wav',
+            stream=False
+        )
+
+        return {
+            'topic': 'tts-audio-streaming',
+            'data': {
+                'outputPath': output_path,
+                'audioId': audio_id
+            }
+        }
+
+    def leon_speech_audio_ended(self, audio_duration: float) -> dict:
+        if not self.asr:
+            self.log('ASR is None, cannot update active listening duration')
+
+        if self.asr:
+            if not audio_duration:
+                audio_duration = 0
+            self.asr.active_listening_duration = self.asr.base_active_listening_duration + audio_duration
+            self.log(f'ASR active listening duration increased to {self.asr.active_listening_duration}s')
+
+        return {
+            'topic': 'asr-active-listening-duration-increased',
+            'data': {
+                'activeListeningDuration': self.asr.active_listening_duration
+            }
+        }

tcp_server/src/lib/tts/api.py

@@ -0,0 +1,199 @@
+import re
+import soundfile
+import numpy as np
+import torch.nn as nn
+from tqdm import tqdm
+import torch
+import time
+import wave
+import os
+
+from . import utils
+from .models import SynthesizerTrn
+from .split_utils import split_sentence
+from ..utils import is_macos
+
+# torch.backends.cudnn.enabled = False
+
+class TTS(nn.Module):
+    def __init__(self,
+                language,
+                # auto, cpu, cuda, mps
+                device='auto',
+                use_hf=True,
+                config_path=None,
+                ckpt_path=None):
+        super().__init__()
+
+        tic = time.perf_counter()
+        self.log('Loading model...')
+
+        if device == 'auto':
+            device = 'cpu'
+
+            if torch.cuda.is_available():
+                device = 'cuda'
+                self.log('Using CUDA (Compute Unified Device Architecture)')
+            if torch.backends.mps.is_available():
+                device = 'mps'
+                self.log('Using MPS (Metal Performance Shaders)')
+        if 'cuda' in device:
+            assert torch.cuda.is_available()
+        if 'mps' in device:
+            assert torch.backends.mps.is_available()
+
+        if is_macos():
+            """
+            Temporary fix.
+            Force CPU device for macOS because of the memory leak where cache does not want to clear up on MPS
+            """
+            device = 'cpu'
+
+        self.log(f'Device: {device}')
+
+        hps = utils.get_hparams_from_file(config_path)
+
+        num_languages = hps.num_languages
+        num_tones = hps.num_tones
+        symbols = hps.symbols
+
+        model = SynthesizerTrn(
+            len(symbols),
+            hps.data.filter_length // 2 + 1,
+            hps.train.segment_size // hps.data.hop_length,
+            n_speakers=hps.data.n_speakers,
+            num_tones=num_tones,
+            num_languages=num_languages,
+            **hps.model,
+        ).to(device)
+
+        model.eval()
+        self.model = model
+        self.symbol_to_id = {s: i for i, s in enumerate(symbols)}
+        self.hps = hps
+        self.device = device
+
+        # load state_dict
+        checkpoint_dict = torch.load(ckpt_path, map_location=device)
+        self.model.load_state_dict(checkpoint_dict['model'], strict=True)
+
+        language = language.split('_')[0]
+        self.language = 'ZH_MIX_EN' if language == 'ZH' else language # we support a ZH_MIX_EN model
+
+        self.log('Model loaded')
+        toc = time.perf_counter()
+
+        self.log(f"Time taken to load model: {toc - tic:0.4f} seconds")
+
+        self.log('Warming up model...')
+        speaker_ids = self.hps.data.spk2id
+        self.tts_to_file('This is a test.', speaker_ids['EN-Leon-V1_1'], quiet=True, format='wav')
+        self.log('Model warmed up')
+
+    @staticmethod
+    def audio_numpy_concat(segment_data_list, sr, speed=1.):
+        audio_segments = []
+        for segment_data in segment_data_list:
+            audio_segments += segment_data.reshape(-1).tolist()
+            audio_segments += [0] * int((sr * 0.05) / speed)
+        audio_segments = np.array(audio_segments).astype(np.float32)
+        return audio_segments
+
+    @staticmethod
+    def split_sentences_into_pieces(text, language, quiet=False):
+        texts = split_sentence(text, language_str=language)
+        if not quiet:
+            print(" > Text split to sentences.")
+            print('\n'.join(texts))
+            print(" > ===========================")
+        return texts
+
+    def tts_iter(self, text, speaker_id, sdp_ratio=0.2, noise_scale=0.6, noise_scale_w=0.8, speed=1.0, pbar=None, position=None, quiet=False, stream=False):
+        tic = time.perf_counter()
+        self.log(f"Generating audio for:\n{text}")
+        language = self.language
+
+        texts = self.split_sentences_into_pieces(text, language, quiet)
+
+        if pbar:
+            tx = pbar(texts)
+        else:
+            if position:
+                tx = tqdm(texts, position=position)
+            elif quiet:
+                tx = texts
+            else:
+                tx = tqdm(texts)
+        for t in tx:
+            if language in ['EN', 'ZH_MIX_EN']:
+                t = re.sub(r'([a-z])([A-Z])', r'\1 \2', t)
+            device = self.device
+            bert, ja_bert, phones, tones, lang_ids = utils.get_text_for_tts_infer(t, language, self.hps, device, self.symbol_to_id)
+            with torch.no_grad():
+                x_tst = phones.to(device).unsqueeze(0)
+                tones = tones.to(device).unsqueeze(0)
+                lang_ids = lang_ids.to(device).unsqueeze(0)
+                bert = bert.to(device).unsqueeze(0)
+                ja_bert = ja_bert.to(device).unsqueeze(0)
+                x_tst_lengths = torch.LongTensor([phones.size(0)]).to(device)
+                del phones
+                speakers = torch.LongTensor([speaker_id]).to(device)
+                audio = self.model.infer(
+                        x_tst,
+                        x_tst_lengths,
+                        speakers,
+                        tones,
+                        lang_ids,
+                        bert,
+                        ja_bert,
+                        sdp_ratio=sdp_ratio,
+                        noise_scale=noise_scale,
+                        noise_scale_w=noise_scale_w,
+                        length_scale=1. / speed,
+                    )[0][0, 0].data.cpu().float().numpy()
+                del x_tst, tones, lang_ids, bert, ja_bert, x_tst_lengths, speakers
+
+                audio_segments = []
+                audio_segments += audio.reshape(-1).tolist()
+                audio_segments += [0] * int((self.hps.data.sampling_rate * 0.05) / speed)
+                audio_segments = np.array(audio_segments).astype(np.float32)
+
+                yield audio_segments
+
+        toc = time.perf_counter()
+        self.log(f"Time taken to generate audio: {toc - tic:0.4f} seconds")
+
+        if self.device == 'cuda':
+            torch.cuda.empty_cache()
+        if self.device == 'mps':
+            torch.mps.empty_cache()
+
+    def tts_to_file(self, text, speaker_id, output_path=None, sdp_ratio=0.2, noise_scale=0.6, noise_scale_w=0.8, speed=1.0, pbar=None, format=None, position=None, quiet=False, stream=False):
+        audio_list = []
+        for audio in self.tts_iter(
+            text=text,
+            speaker_id=speaker_id,
+            sdp_ratio=sdp_ratio,
+            noise_scale=noise_scale,
+            noise_scale_w=noise_scale_w,
+            speed=speed,
+            pbar=pbar,
+            position=position,
+            quiet=quiet,
+            stream=stream
+        ):
+            audio_list.append(audio)
+
+        audio = np.concatenate(audio_list)
+
+        if output_path is None:
+            return audio
+        else:
+            if format:
+                soundfile.write(output_path, audio, self.hps.data.sampling_rate, format=format)
+            else:
+                soundfile.write(output_path, audio, self.hps.data.sampling_rate)
+
+    @staticmethod
+    def log(*args, **kwargs):
+        print('[TTS]', *args, **kwargs)

tcp_server/src/lib/tts/attentions.py

@@ -0,0 +1,449 @@
+import math
+import torch
+from torch import nn
+from torch.nn import functional as F
+
+from . import commons
+import logging
+
+logger = logging.getLogger(__name__)
+
+
+class LayerNorm(nn.Module):
+    def __init__(self, channels, eps=1e-5):
+        super().__init__()
+        self.channels = channels
+        self.eps = eps
+
+        self.gamma = nn.Parameter(torch.ones(channels))
+        self.beta = nn.Parameter(torch.zeros(channels))
+
+    def forward(self, x):
+        x = x.transpose(1, -1)
+        x = F.layer_norm(x, (self.channels,), self.gamma, self.beta, self.eps)
+        return x.transpose(1, -1)
+
+
+class Encoder(nn.Module):
+    def __init__(
+        self,
+        hidden_channels,
+        filter_channels,
+        n_heads,
+        n_layers,
+        kernel_size=1,
+        p_dropout=0.0,
+        window_size=4,
+        isflow=True,
+        **kwargs
+    ):
+        super().__init__()
+        self.hidden_channels = hidden_channels
+        self.filter_channels = filter_channels
+        self.n_heads = n_heads
+        self.n_layers = n_layers
+        self.kernel_size = kernel_size
+        self.p_dropout = p_dropout
+        self.window_size = window_size
+
+        self.cond_layer_idx = self.n_layers
+        if "gin_channels" in kwargs:
+            self.gin_channels = kwargs["gin_channels"]
+            if self.gin_channels != 0:
+                self.spk_emb_linear = nn.Linear(self.gin_channels, self.hidden_channels)
+                self.cond_layer_idx = (
+                    kwargs["cond_layer_idx"] if "cond_layer_idx" in kwargs else 2
+                )
+                assert (
+                    self.cond_layer_idx < self.n_layers
+                ), "cond_layer_idx should be less than n_layers"
+        self.drop = nn.Dropout(p_dropout)
+        self.attn_layers = nn.ModuleList()
+        self.norm_layers_1 = nn.ModuleList()
+        self.ffn_layers = nn.ModuleList()
+        self.norm_layers_2 = nn.ModuleList()
+
+        for i in range(self.n_layers):
+            self.attn_layers.append(
+                MultiHeadAttention(
+                    hidden_channels,
+                    hidden_channels,
+                    n_heads,
+                    p_dropout=p_dropout,
+                    window_size=window_size,
+                )
+            )
+            self.norm_layers_1.append(LayerNorm(hidden_channels))
+            self.ffn_layers.append(
+                FFN(
+                    hidden_channels,
+                    hidden_channels,
+                    filter_channels,
+                    kernel_size,
+                    p_dropout=p_dropout,
+                )
+            )
+            self.norm_layers_2.append(LayerNorm(hidden_channels))
+
+    def forward(self, x, x_mask, g=None):
+        attn_mask = x_mask.unsqueeze(2) * x_mask.unsqueeze(-1)
+        x = x * x_mask
+        for i in range(self.n_layers):
+            if i == self.cond_layer_idx and g is not None:
+                g = self.spk_emb_linear(g.transpose(1, 2))
+                g = g.transpose(1, 2)
+                x = x + g
+                x = x * x_mask
+            y = self.attn_layers[i](x, x, attn_mask)
+            y = self.drop(y)
+            x = self.norm_layers_1[i](x + y)
+
+            y = self.ffn_layers[i](x, x_mask)
+            y = self.drop(y)
+            x = self.norm_layers_2[i](x + y)
+        x = x * x_mask
+        return x
+
+
+class Decoder(nn.Module):
+    def __init__(
+        self,
+        hidden_channels,
+        filter_channels,
+        n_heads,
+        n_layers,
+        kernel_size=1,
+        p_dropout=0.0,
+        proximal_bias=False,
+        proximal_init=True,
+        **kwargs
+    ):
+        super().__init__()
+        self.hidden_channels = hidden_channels
+        self.filter_channels = filter_channels
+        self.n_heads = n_heads
+        self.n_layers = n_layers
+        self.kernel_size = kernel_size
+        self.p_dropout = p_dropout
+        self.proximal_bias = proximal_bias
+        self.proximal_init = proximal_init
+
+        self.drop = nn.Dropout(p_dropout)
+        self.self_attn_layers = nn.ModuleList()
+        self.norm_layers_0 = nn.ModuleList()
+        self.encdec_attn_layers = nn.ModuleList()
+        self.norm_layers_1 = nn.ModuleList()
+        self.ffn_layers = nn.ModuleList()
+        self.norm_layers_2 = nn.ModuleList()
+        for i in range(self.n_layers):
+            self.self_attn_layers.append(
+                MultiHeadAttention(
+                    hidden_channels,
+                    hidden_channels,
+                    n_heads,
+                    p_dropout=p_dropout,
+                    proximal_bias=proximal_bias,
+                    proximal_init=proximal_init,
+                )
+            )
+            self.norm_layers_0.append(LayerNorm(hidden_channels))
+            self.encdec_attn_layers.append(
+                MultiHeadAttention(
+                    hidden_channels, hidden_channels, n_heads, p_dropout=p_dropout
+                )
+            )
+            self.norm_layers_1.append(LayerNorm(hidden_channels))
+            self.ffn_layers.append(
+                FFN(
+                    hidden_channels,
+                    hidden_channels,
+                    filter_channels,
+                    kernel_size,
+                    p_dropout=p_dropout,
+                    causal=True,
+                )
+            )
+            self.norm_layers_2.append(LayerNorm(hidden_channels))
+
+    def forward(self, x, x_mask, h, h_mask):
+        """
+        x: decoder input
+        h: encoder output
+        """
+        self_attn_mask = commons.subsequent_mask(x_mask.size(2)).to(
+            device=x.device, dtype=x.dtype
+        )
+        encdec_attn_mask = h_mask.unsqueeze(2) * x_mask.unsqueeze(-1)
+        x = x * x_mask
+        for i in range(self.n_layers):
+            y = self.self_attn_layers[i](x, x, self_attn_mask)
+            y = self.drop(y)
+            x = self.norm_layers_0[i](x + y)
+
+            y = self.encdec_attn_layers[i](x, h, encdec_attn_mask)
+            y = self.drop(y)
+            x = self.norm_layers_1[i](x + y)
+
+            y = self.ffn_layers[i](x, x_mask)
+            y = self.drop(y)
+            x = self.norm_layers_2[i](x + y)
+        x = x * x_mask
+        return x
+
+
+class MultiHeadAttention(nn.Module):
+    def __init__(
+        self,
+        channels,
+        out_channels,
+        n_heads,
+        p_dropout=0.0,
+        window_size=None,
+        heads_share=True,
+        block_length=None,
+        proximal_bias=False,
+        proximal_init=False,
+    ):
+        super().__init__()
+        assert channels % n_heads == 0
+
+        self.channels = channels
+        self.out_channels = out_channels
+        self.n_heads = n_heads
+        self.p_dropout = p_dropout
+        self.window_size = window_size
+        self.heads_share = heads_share
+        self.block_length = block_length
+        self.proximal_bias = proximal_bias
+        self.proximal_init = proximal_init
+        self.attn = None
+
+        self.k_channels = channels // n_heads
+        self.conv_q = nn.Conv1d(channels, channels, 1)
+        self.conv_k = nn.Conv1d(channels, channels, 1)
+        self.conv_v = nn.Conv1d(channels, channels, 1)
+        self.conv_o = nn.Conv1d(channels, out_channels, 1)
+        self.drop = nn.Dropout(p_dropout)
+
+        if window_size is not None:
+            n_heads_rel = 1 if heads_share else n_heads
+            rel_stddev = self.k_channels**-0.5
+            self.emb_rel_k = nn.Parameter(
+                torch.randn(n_heads_rel, window_size * 2 + 1, self.k_channels)
+                * rel_stddev
+            )
+            self.emb_rel_v = nn.Parameter(
+                torch.randn(n_heads_rel, window_size * 2 + 1, self.k_channels)
+                * rel_stddev
+            )
+
+        nn.init.xavier_uniform_(self.conv_q.weight)
+        nn.init.xavier_uniform_(self.conv_k.weight)
+        nn.init.xavier_uniform_(self.conv_v.weight)
+        if proximal_init:
+            with torch.no_grad():
+                self.conv_k.weight.copy_(self.conv_q.weight)
+                self.conv_k.bias.copy_(self.conv_q.bias)
+
+    def forward(self, x, c, attn_mask=None):
+        q = self.conv_q(x)
+        k = self.conv_k(c)
+        v = self.conv_v(c)
+
+        x, self.attn = self.attention(q, k, v, mask=attn_mask)
+
+        x = self.conv_o(x)
+        return x
+
+    def attention(self, query, key, value, mask=None):
+        # reshape [b, d, t] -> [b, n_h, t, d_k]
+        b, d, t_s, t_t = (*key.size(), query.size(2))
+        query = query.view(b, self.n_heads, self.k_channels, t_t).transpose(2, 3)
+        key = key.view(b, self.n_heads, self.k_channels, t_s).transpose(2, 3)
+        value = value.view(b, self.n_heads, self.k_channels, t_s).transpose(2, 3)
+
+        scores = torch.matmul(query / math.sqrt(self.k_channels), key.transpose(-2, -1))
+        if self.window_size is not None:
+            assert (
+                t_s == t_t
+            ), "Relative attention is only available for self-attention."
+            key_relative_embeddings = self._get_relative_embeddings(self.emb_rel_k, t_s)
+            rel_logits = self._matmul_with_relative_keys(
+                query / math.sqrt(self.k_channels), key_relative_embeddings
+            )
+            scores_local = self._relative_position_to_absolute_position(rel_logits)
+            scores = scores + scores_local
+        if self.proximal_bias:
+            assert t_s == t_t, "Proximal bias is only available for self-attention."
+            scores = scores + self._attention_bias_proximal(t_s).to(
+                device=scores.device, dtype=scores.dtype
+            )
+        if mask is not None:
+            scores = scores.masked_fill(mask == 0, -1e4)
+            if self.block_length is not None:
+                assert (
+                    t_s == t_t
+                ), "Local attention is only available for self-attention."
+                block_mask = (
+                    torch.ones_like(scores)
+                    .triu(-self.block_length)
+                    .tril(self.block_length)
+                )
+                scores = scores.masked_fill(block_mask == 0, -1e4)
+        p_attn = F.softmax(scores, dim=-1)  # [b, n_h, t_t, t_s]
+        p_attn = self.drop(p_attn)
+        output = torch.matmul(p_attn, value)
+        if self.window_size is not None:
+            relative_weights = self._absolute_position_to_relative_position(p_attn)
+            value_relative_embeddings = self._get_relative_embeddings(
+                self.emb_rel_v, t_s
+            )
+            output = output + self._matmul_with_relative_values(
+                relative_weights, value_relative_embeddings
+            )
+        output = (
+            output.transpose(2, 3).contiguous().view(b, d, t_t)
+        )  # [b, n_h, t_t, d_k] -> [b, d, t_t]
+        return output, p_attn
+
+    def _matmul_with_relative_values(self, x, y):
+        """
+        x: [b, h, l, m]
+        y: [h or 1, m, d]
+        ret: [b, h, l, d]
+        """
+        ret = torch.matmul(x, y.unsqueeze(0))
+        return ret
+
+    def _matmul_with_relative_keys(self, x, y):
+        """
+        x: [b, h, l, d]
+        y: [h or 1, m, d]
+        ret: [b, h, l, m]
+        """
+        ret = torch.matmul(x, y.unsqueeze(0).transpose(-2, -1))
+        return ret
+
+    def _get_relative_embeddings(self, relative_embeddings, length):
+        2 * self.window_size + 1
+        # Pad first before slice to avoid using cond ops.
+        pad_length = max(length - (self.window_size + 1), 0)
+        slice_start_position = max((self.window_size + 1) - length, 0)
+        slice_end_position = slice_start_position + 2 * length - 1
+        if pad_length > 0:
+            padded_relative_embeddings = F.pad(
+                relative_embeddings,
+                commons.convert_pad_shape([[0, 0], [pad_length, pad_length], [0, 0]]),
+            )
+        else:
+            padded_relative_embeddings = relative_embeddings
+        used_relative_embeddings = padded_relative_embeddings[
+            :, slice_start_position:slice_end_position
+        ]
+        return used_relative_embeddings
+
+    def _relative_position_to_absolute_position(self, x):
+        """
+        x: [b, h, l, 2*l-1]
+        ret: [b, h, l, l]
+        """
+        batch, heads, length, _ = x.size()
+        # Concat columns of pad to shift from relative to absolute indexing.
+        x = F.pad(x, commons.convert_pad_shape([[0, 0], [0, 0], [0, 0], [0, 1]]))
+
+        # Concat extra elements so to add up to shape (len+1, 2*len-1).
+        x_flat = x.view([batch, heads, length * 2 * length])
+        x_flat = F.pad(
+            x_flat, commons.convert_pad_shape([[0, 0], [0, 0], [0, length - 1]])
+        )
+
+        # Reshape and slice out the padded elements.
+        x_final = x_flat.view([batch, heads, length + 1, 2 * length - 1])[
+            :, :, :length, length - 1 :
+        ]
+        return x_final
+
+    def _absolute_position_to_relative_position(self, x):
+        """
+        x: [b, h, l, l]
+        ret: [b, h, l, 2*l-1]
+        """
+        batch, heads, length, _ = x.size()
+        # pad along column
+        x = F.pad(
+            x, commons.convert_pad_shape([[0, 0], [0, 0], [0, 0], [0, length - 1]])
+        )
+        x_flat = x.view([batch, heads, length**2 + length * (length - 1)])
+        # add 0's in the beginning that will skew the elements after reshape
+        x_flat = F.pad(x_flat, commons.convert_pad_shape([[0, 0], [0, 0], [length, 0]]))
+        x_final = x_flat.view([batch, heads, length, 2 * length])[:, :, :, 1:]
+        return x_final
+
+    def _attention_bias_proximal(self, length):
+        """Bias for self-attention to encourage attention to close positions.
+        Args:
+          length: an integer scalar.
+        Returns:
+          a Tensor with shape [1, 1, length, length]
+        """
+        r = torch.arange(length, dtype=torch.float32)
+        diff = torch.unsqueeze(r, 0) - torch.unsqueeze(r, 1)
+        return torch.unsqueeze(torch.unsqueeze(-torch.log1p(torch.abs(diff)), 0), 0)
+
+
+class FFN(nn.Module):
+    def __init__(
+        self,
+        in_channels,
+        out_channels,
+        filter_channels,
+        kernel_size,
+        p_dropout=0.0,
+        activation=None,
+        causal=False,
+    ):
+        super().__init__()
+        self.in_channels = in_channels
+        self.out_channels = out_channels
+        self.filter_channels = filter_channels
+        self.kernel_size = kernel_size
+        self.p_dropout = p_dropout
+        self.activation = activation
+        self.causal = causal
+
+        if causal:
+            self.padding = self._causal_padding
+        else:
+            self.padding = self._same_padding
+
+        self.conv_1 = nn.Conv1d(in_channels, filter_channels, kernel_size)
+        self.conv_2 = nn.Conv1d(filter_channels, out_channels, kernel_size)
+        self.drop = nn.Dropout(p_dropout)
+
+    def forward(self, x, x_mask):
+        x = self.conv_1(self.padding(x * x_mask))
+        if self.activation == "gelu":
+            x = x * torch.sigmoid(1.702 * x)
+        else:
+            x = torch.relu(x)
+        x = self.drop(x)
+        x = self.conv_2(self.padding(x * x_mask))
+        return x * x_mask
+
+    def _causal_padding(self, x):
+        if self.kernel_size == 1:
+            return x
+        pad_l = self.kernel_size - 1
+        pad_r = 0
+        padding = [[0, 0], [0, 0], [pad_l, pad_r]]
+        x = F.pad(x, commons.convert_pad_shape(padding))
+        return x
+
+    def _same_padding(self, x):
+        if self.kernel_size == 1:
+            return x
+        pad_l = (self.kernel_size - 1) // 2
+        pad_r = self.kernel_size // 2
+        padding = [[0, 0], [0, 0], [pad_l, pad_r]]
+        x = F.pad(x, commons.convert_pad_shape(padding))
+        return x

tcp_server/src/lib/tts/commons.py

@@ -0,0 +1,150 @@
+import math
+import torch
+from torch.nn import functional as F
+
+
+def init_weights(m, mean=0.0, std=0.01):
+    classname = m.__class__.__name__
+    if classname.find("Conv") != -1:
+        m.weight.data.normal_(mean, std)
+
+
+def get_padding(kernel_size, dilation=1):
+    return int((kernel_size * dilation - dilation) / 2)
+
+
+def convert_pad_shape(pad_shape):
+    layer = pad_shape[::-1]
+    pad_shape = [item for sublist in layer for item in sublist]
+    return pad_shape
+
+
+def intersperse(lst, item):
+    result = [item] * (len(lst) * 2 + 1)
+    result[1::2] = lst
+    return result
+
+
+def kl_divergence(m_p, logs_p, m_q, logs_q):
+    """KL(P||Q)"""
+    kl = (logs_q - logs_p) - 0.5
+    kl += (
+        0.5 * (torch.exp(2.0 * logs_p) + ((m_p - m_q) ** 2)) * torch.exp(-2.0 * logs_q)
+    )
+    return kl
+
+
+def rand_gumbel(shape):
+    """Sample from the Gumbel distribution, protect from overflows."""
+    uniform_samples = torch.rand(shape) * 0.99998 + 0.00001
+    return -torch.log(-torch.log(uniform_samples))
+
+
+def rand_gumbel_like(x):
+    g = rand_gumbel(x.size()).to(dtype=x.dtype, device=x.device)
+    return g
+
+
+def slice_segments(x, ids_str, segment_size=4):
+    ret = torch.zeros_like(x[:, :, :segment_size])
+    for i in range(x.size(0)):
+        idx_str = ids_str[i]
+        idx_end = idx_str + segment_size
+        ret[i] = x[i, :, idx_str:idx_end]
+    return ret
+
+
+def rand_slice_segments(x, x_lengths=None, segment_size=4):
+    b, d, t = x.size()
+    if x_lengths is None:
+        x_lengths = t
+    ids_str_max = x_lengths - segment_size + 1
+    ids_str = (torch.rand([b]).to(device=x.device) * ids_str_max).to(dtype=torch.long)
+    ret = slice_segments(x, ids_str, segment_size)
+    return ret, ids_str
+
+
+def get_timing_signal_1d(length, channels, min_timescale=1.0, max_timescale=1.0e4):
+    position = torch.arange(length, dtype=torch.float)
+    num_timescales = channels // 2
+    log_timescale_increment = math.log(float(max_timescale) / float(min_timescale)) / (
+        num_timescales - 1
+    )
+    inv_timescales = min_timescale * torch.exp(
+        torch.arange(num_timescales, dtype=torch.float) * -log_timescale_increment
+    )
+    scaled_time = position.unsqueeze(0) * inv_timescales.unsqueeze(1)
+    signal = torch.cat([torch.sin(scaled_time), torch.cos(scaled_time)], 0)
+    signal = F.pad(signal, [0, 0, 0, channels % 2])
+    signal = signal.view(1, channels, length)
+    return signal
+
+
+def add_timing_signal_1d(x, min_timescale=1.0, max_timescale=1.0e4):
+    b, channels, length = x.size()
+    signal = get_timing_signal_1d(length, channels, min_timescale, max_timescale)
+    return x + signal.to(dtype=x.dtype, device=x.device)
+
+
+def cat_timing_signal_1d(x, min_timescale=1.0, max_timescale=1.0e4, axis=1):
+    b, channels, length = x.size()
+    signal = get_timing_signal_1d(length, channels, min_timescale, max_timescale)
+    return torch.cat([x, signal.to(dtype=x.dtype, device=x.device)], axis)
+
+
+def subsequent_mask(length):
+    mask = torch.tril(torch.ones(length, length)).unsqueeze(0).unsqueeze(0)
+    return mask
+
+
+def convert_pad_shape(pad_shape):
+    layer = pad_shape[::-1]
+    pad_shape = [item for sublist in layer for item in sublist]
+    return pad_shape
+
+
+def shift_1d(x):
+    x = F.pad(x, convert_pad_shape([[0, 0], [0, 0], [1, 0]]))[:, :, :-1]
+    return x
+
+
+def sequence_mask(length, max_length=None):
+    if max_length is None:
+        max_length = length.max()
+    x = torch.arange(max_length, dtype=length.dtype, device=length.device)
+    return x.unsqueeze(0) < length.unsqueeze(1)
+
+
+def generate_path(duration, mask):
+    """
+    duration: [b, 1, t_x]
+    mask: [b, 1, t_y, t_x]
+    """
+
+    b, _, t_y, t_x = mask.shape
+    cum_duration = torch.cumsum(duration, -1)
+
+    cum_duration_flat = cum_duration.view(b * t_x)
+    path = sequence_mask(cum_duration_flat, t_y).to(mask.dtype)
+    path = path.view(b, t_x, t_y)
+    path = path - F.pad(path, convert_pad_shape([[0, 0], [1, 0], [0, 0]]))[:, :-1]
+    path = path.unsqueeze(1).transpose(2, 3) * mask
+    return path
+
+
+def clip_grad_value_(parameters, clip_value, norm_type=2):
+    if isinstance(parameters, torch.Tensor):
+        parameters = [parameters]
+    parameters = list(filter(lambda p: p.grad is not None, parameters))
+    norm_type = float(norm_type)
+    if clip_value is not None:
+        clip_value = float(clip_value)
+
+    total_norm = 0
+    for p in parameters:
+        param_norm = p.grad.data.norm(norm_type)
+        total_norm += param_norm.item() ** norm_type
+        if clip_value is not None:
+            p.grad.data.clamp_(min=-clip_value, max=clip_value)
+    total_norm = total_norm ** (1.0 / norm_type)
+    return total_norm

tcp_server/src/lib/tts/modules.py

@@ -0,0 +1,571 @@
+import math
+import torch
+from torch import nn
+from torch.nn import functional as F
+
+from torch.nn import Conv1d
+from torch.nn.utils import weight_norm, remove_weight_norm
+
+from . import commons
+from .commons import init_weights, get_padding
+from .transforms import piecewise_rational_quadratic_transform
+from .attentions import Encoder
+
+LRELU_SLOPE = 0.1
+
+
+class LayerNorm(nn.Module):
+    def __init__(self, channels, eps=1e-5):
+        super().__init__()
+        self.channels = channels
+        self.eps = eps
+
+        self.gamma = nn.Parameter(torch.ones(channels))
+        self.beta = nn.Parameter(torch.zeros(channels))
+
+    def forward(self, x):
+        x = x.transpose(1, -1)
+        x = F.layer_norm(x, (self.channels,), self.gamma, self.beta, self.eps)
+        return x.transpose(1, -1)
+
+
+class ConvReluNorm(nn.Module):
+    def __init__(
+        self,
+        in_channels,
+        hidden_channels,
+        out_channels,
+        kernel_size,
+        n_layers,
+        p_dropout,
+    ):
+        super().__init__()
+        self.in_channels = in_channels
+        self.hidden_channels = hidden_channels
+        self.out_channels = out_channels
+        self.kernel_size = kernel_size
+        self.n_layers = n_layers
+        self.p_dropout = p_dropout
+        assert n_layers > 1, "Number of layers should be larger than 0."
+
+        self.conv_layers = nn.ModuleList()
+        self.norm_layers = nn.ModuleList()
+        self.conv_layers.append(
+            nn.Conv1d(
+                in_channels, hidden_channels, kernel_size, padding=kernel_size // 2
+            )
+        )
+        self.norm_layers.append(LayerNorm(hidden_channels))
+        self.relu_drop = nn.Sequential(nn.ReLU(), nn.Dropout(p_dropout))
+        for _ in range(n_layers - 1):
+            self.conv_layers.append(
+                nn.Conv1d(
+                    hidden_channels,
+                    hidden_channels,
+                    kernel_size,
+                    padding=kernel_size // 2,
+                )
+            )
+            self.norm_layers.append(LayerNorm(hidden_channels))
+        self.proj = nn.Conv1d(hidden_channels, out_channels, 1)
+        self.proj.weight.data.zero_()
+        self.proj.bias.data.zero_()
+
+    def forward(self, x, x_mask):
+        x_org = x
+        for i in range(self.n_layers):
+            x = self.conv_layers[i](x * x_mask)
+            x = self.norm_layers[i](x)
+            x = self.relu_drop(x)
+        x = x_org + self.proj(x)
+        return x * x_mask
+
+
+class DDSConv(nn.Module):
+    """
+    Dialted and Depth-Separable Convolution
+    """
+
+    def __init__(self, channels, kernel_size, n_layers, p_dropout=0.0):
+        super().__init__()
+        self.channels = channels
+        self.kernel_size = kernel_size
+        self.n_layers = n_layers
+        self.p_dropout = p_dropout
+
+        self.drop = nn.Dropout(p_dropout)
+        self.convs_sep = nn.ModuleList()
+        self.convs_1x1 = nn.ModuleList()
+        self.norms_1 = nn.ModuleList()
+        self.norms_2 = nn.ModuleList()
+        for i in range(n_layers):
+            dilation = kernel_size**i
+            padding = (kernel_size * dilation - dilation) // 2
+            self.convs_sep.append(
+                nn.Conv1d(
+                    channels,
+                    channels,
+                    kernel_size,
+                    groups=channels,
+                    dilation=dilation,
+                    padding=padding,
+                )
+            )
+            self.convs_1x1.append(nn.Conv1d(channels, channels, 1))
+            self.norms_1.append(LayerNorm(channels))
+            self.norms_2.append(LayerNorm(channels))
+
+    def forward(self, x, x_mask, g=None):
+        if g is not None:
+            x = x + g
+        for i in range(self.n_layers):
+            y = self.convs_sep[i](x * x_mask)
+            y = self.norms_1[i](y)
+            y = F.gelu(y)
+            y = self.convs_1x1[i](y)
+            y = self.norms_2[i](y)
+            y = F.gelu(y)
+            y = self.drop(y)
+            x = x + y
+        return x * x_mask
+
+
+class WN(torch.nn.Module):
+    def __init__(
+        self,
+        hidden_channels,
+        kernel_size,
+        dilation_rate,
+        n_layers,
+        gin_channels=0,
+        p_dropout=0,
+    ):
+        super(WN, self).__init__()
+        assert kernel_size % 2 == 1
+        self.hidden_channels = hidden_channels
+        self.kernel_size = (kernel_size,)
+        self.dilation_rate = dilation_rate
+        self.n_layers = n_layers
+        self.gin_channels = gin_channels
+        self.p_dropout = p_dropout
+
+        self.in_layers = torch.nn.ModuleList()
+        self.res_skip_layers = torch.nn.ModuleList()
+        self.drop = nn.Dropout(p_dropout)
+
+        if gin_channels != 0:
+            cond_layer = torch.nn.Conv1d(
+                gin_channels, 2 * hidden_channels * n_layers, 1
+            )
+            self.cond_layer = torch.nn.utils.weight_norm(cond_layer, name="weight")
+
+        for i in range(n_layers):
+            dilation = dilation_rate**i
+            padding = int((kernel_size * dilation - dilation) / 2)
+            in_layer = torch.nn.Conv1d(
+                hidden_channels,
+                2 * hidden_channels,
+                kernel_size,
+                dilation=dilation,
+                padding=padding,
+            )
+            in_layer = torch.nn.utils.weight_norm(in_layer, name="weight")
+            self.in_layers.append(in_layer)
+
+            # last one is not necessary
+            if i < n_layers - 1:
+                res_skip_channels = 2 * hidden_channels
+            else:
+                res_skip_channels = hidden_channels
+
+            res_skip_layer = torch.nn.Conv1d(hidden_channels, res_skip_channels, 1)
+            res_skip_layer = torch.nn.utils.weight_norm(res_skip_layer, name="weight")
+            self.res_skip_layers.append(res_skip_layer)
+
+    def remove_weight_norm(self):
+        if self.gin_channels != 0:
+            torch.nn.utils.remove_weight_norm(self.cond_layer)
+        for l in self.in_layers:
+            torch.nn.utils.remove_weight_norm(l)
+        for l in self.res_skip_layers:
+            torch.nn.utils.remove_weight_norm(l)
+
+
+class ResBlock1(torch.nn.Module):
+    def __init__(self, channels, kernel_size=3, dilation=(1, 3, 5)):
+        super(ResBlock1, self).__init__()
+        self.convs1 = nn.ModuleList(
+            [
+                weight_norm(
+                    Conv1d(
+                        channels,
+                        channels,
+                        kernel_size,
+                        1,
+                        dilation=dilation[0],
+                        padding=get_padding(kernel_size, dilation[0]),
+                    )
+                ),
+                weight_norm(
+                    Conv1d(
+                        channels,
+                        channels,
+                        kernel_size,
+                        1,
+                        dilation=dilation[1],
+                        padding=get_padding(kernel_size, dilation[1]),
+                    )
+                ),
+                weight_norm(
+                    Conv1d(
+                        channels,
+                        channels,
+                        kernel_size,
+                        1,
+                        dilation=dilation[2],
+                        padding=get_padding(kernel_size, dilation[2]),
+                    )
+                ),
+            ]
+        )
+        self.convs1.apply(init_weights)
+
+        self.convs2 = nn.ModuleList(
+            [
+                weight_norm(
+                    Conv1d(
+                        channels,
+                        channels,
+                        kernel_size,
+                        1,
+                        dilation=1,
+                        padding=get_padding(kernel_size, 1),
+                    )
+                ),
+                weight_norm(
+                    Conv1d(
+                        channels,
+                        channels,
+                        kernel_size,
+                        1,
+                        dilation=1,
+                        padding=get_padding(kernel_size, 1),
+                    )
+                ),
+                weight_norm(
+                    Conv1d(
+                        channels,
+                        channels,
+                        kernel_size,
+                        1,
+                        dilation=1,
+                        padding=get_padding(kernel_size, 1),
+                    )
+                ),
+            ]
+        )
+        self.convs2.apply(init_weights)
+
+    def forward(self, x, x_mask=None):
+        for c1, c2 in zip(self.convs1, self.convs2):
+            xt = F.leaky_relu(x, LRELU_SLOPE)
+            if x_mask is not None:
+                xt = xt * x_mask
+            xt = c1(xt)
+            xt = F.leaky_relu(xt, LRELU_SLOPE)
+            if x_mask is not None:
+                xt = xt * x_mask
+            xt = c2(xt)
+            x = xt + x
+        if x_mask is not None:
+            x = x * x_mask
+        return x
+
+    def remove_weight_norm(self):
+        for l in self.convs1:
+            remove_weight_norm(l)
+        for l in self.convs2:
+            remove_weight_norm(l)
+
+
+class ResBlock2(torch.nn.Module):
+    def __init__(self, channels, kernel_size=3, dilation=(1, 3)):
+        super(ResBlock2, self).__init__()
+        self.convs = nn.ModuleList(
+            [
+                weight_norm(
+                    Conv1d(
+                        channels,
+                        channels,
+                        kernel_size,
+                        1,
+                        dilation=dilation[0],
+                        padding=get_padding(kernel_size, dilation[0]),
+                    )
+                ),
+                weight_norm(
+                    Conv1d(
+                        channels,
+                        channels,
+                        kernel_size,
+                        1,
+                        dilation=dilation[1],
+                        padding=get_padding(kernel_size, dilation[1]),
+                    )
+                ),
+            ]
+        )
+        self.convs.apply(init_weights)
+
+    def forward(self, x, x_mask=None):
+        for c in self.convs:
+            xt = F.leaky_relu(x, LRELU_SLOPE)
+            if x_mask is not None:
+                xt = xt * x_mask
+            xt = c(xt)
+            x = xt + x
+        if x_mask is not None:
+            x = x * x_mask
+        return x
+
+    def remove_weight_norm(self):
+        for l in self.convs:
+            remove_weight_norm(l)
+
+
+class Log(nn.Module):
+    def forward(self, x, x_mask, reverse=False, **kwargs):
+        if not reverse:
+            y = torch.log(torch.clamp_min(x, 1e-5)) * x_mask
+            logdet = torch.sum(-y, [1, 2])
+            return y, logdet
+        else:
+            x = torch.exp(x) * x_mask
+            return x
+
+
+class Flip(nn.Module):
+    def forward(self, x, *args, reverse=False, **kwargs):
+        x = torch.flip(x, [1])
+        if not reverse:
+            logdet = torch.zeros(x.size(0)).to(dtype=x.dtype, device=x.device)
+            return x, logdet
+        else:
+            return x
+
+
+class ElementwiseAffine(nn.Module):
+    def __init__(self, channels):
+        super().__init__()
+        self.channels = channels
+        self.m = nn.Parameter(torch.zeros(channels, 1))
+        self.logs = nn.Parameter(torch.zeros(channels, 1))
+
+    def forward(self, x, x_mask, reverse=False, **kwargs):
+        if not reverse:
+            y = self.m + torch.exp(self.logs) * x
+            y = y * x_mask
+            logdet = torch.sum(self.logs * x_mask, [1, 2])
+            return y, logdet
+        else:
+            x = (x - self.m) * torch.exp(-self.logs) * x_mask
+            return x
+
+
+class ResidualCouplingLayer(nn.Module):
+    def __init__(
+        self,
+        channels,
+        hidden_channels,
+        kernel_size,
+        dilation_rate,
+        n_layers,
+        p_dropout=0,
+        gin_channels=0,
+        mean_only=False,
+    ):
+        assert channels % 2 == 0, "channels should be divisible by 2"
+        super().__init__()
+        self.channels = channels
+        self.hidden_channels = hidden_channels
+        self.kernel_size = kernel_size
+        self.dilation_rate = dilation_rate
+        self.n_layers = n_layers
+        self.half_channels = channels // 2
+        self.mean_only = mean_only
+
+        self.pre = nn.Conv1d(self.half_channels, hidden_channels, 1)
+        self.enc = WN(
+            hidden_channels,
+            kernel_size,
+            dilation_rate,
+            n_layers,
+            p_dropout=p_dropout,
+            gin_channels=gin_channels,
+        )
+        self.post = nn.Conv1d(hidden_channels, self.half_channels * (2 - mean_only), 1)
+        self.post.weight.data.zero_()
+        self.post.bias.data.zero_()
+
+    def forward(self, x, x_mask, g=None, reverse=False):
+        x0, x1 = torch.split(x, [self.half_channels] * 2, 1)
+        h = self.pre(x0) * x_mask
+        h = self.enc(h, x_mask, g=g)
+        stats = self.post(h) * x_mask
+        if not self.mean_only:
+            m, logs = torch.split(stats, [self.half_channels] * 2, 1)
+        else:
+            m = stats
+            logs = torch.zeros_like(m)
+
+        if not reverse:
+            x1 = m + x1 * torch.exp(logs) * x_mask
+            x = torch.cat([x0, x1], 1)
+            logdet = torch.sum(logs, [1, 2])
+            return x, logdet
+        else:
+            x1 = (x1 - m) * torch.exp(-logs) * x_mask
+            x = torch.cat([x0, x1], 1)
+            return x
+
+
+class ConvFlow(nn.Module):
+    def __init__(
+        self,
+        in_channels,
+        filter_channels,
+        kernel_size,
+        n_layers,
+        num_bins=10,
+        tail_bound=5.0,
+    ):
+        super().__init__()
+        self.in_channels = in_channels
+        self.filter_channels = filter_channels
+        self.kernel_size = kernel_size
+        self.n_layers = n_layers
+        self.num_bins = num_bins
+        self.tail_bound = tail_bound
+        self.half_channels = in_channels // 2
+
+        self.pre = nn.Conv1d(self.half_channels, filter_channels, 1)
+        self.convs = DDSConv(filter_channels, kernel_size, n_layers, p_dropout=0.0)
+        self.proj = nn.Conv1d(
+            filter_channels, self.half_channels * (num_bins * 3 - 1), 1
+        )
+        self.proj.weight.data.zero_()
+        self.proj.bias.data.zero_()
+
+    def forward(self, x, x_mask, g=None, reverse=False):
+        x0, x1 = torch.split(x, [self.half_channels] * 2, 1)
+        h = self.pre(x0)
+        h = self.convs(h, x_mask, g=g)
+        h = self.proj(h) * x_mask
+
+        b, c, t = x0.shape
+        h = h.reshape(b, c, -1, t).permute(0, 1, 3, 2)  # [b, cx?, t] -> [b, c, t, ?]
+
+        unnormalized_widths = h[..., : self.num_bins] / math.sqrt(self.filter_channels)
+        unnormalized_heights = h[..., self.num_bins : 2 * self.num_bins] / math.sqrt(
+            self.filter_channels
+        )
+        unnormalized_derivatives = h[..., 2 * self.num_bins :]
+
+        x1, logabsdet = piecewise_rational_quadratic_transform(
+            x1,
+            unnormalized_widths,
+            unnormalized_heights,
+            unnormalized_derivatives,
+            inverse=reverse,
+            tails="linear",
+            tail_bound=self.tail_bound,
+        )
+
+        x = torch.cat([x0, x1], 1) * x_mask
+        logdet = torch.sum(logabsdet * x_mask, [1, 2])
+        if not reverse:
+            return x, logdet
+        else:
+            return x
+
+
+class TransformerCouplingLayer(nn.Module):
+    def __init__(
+        self,
+        channels,
+        hidden_channels,
+        kernel_size,
+        n_layers,
+        n_heads,
+        p_dropout=0,
+        filter_channels=0,
+        mean_only=False,
+        wn_sharing_parameter=None,
+        gin_channels=0,
+    ):
+        assert n_layers == 3, n_layers
+        assert channels % 2 == 0, "channels should be divisible by 2"
+        super().__init__()
+        self.channels = channels
+        self.hidden_channels = hidden_channels
+        self.kernel_size = kernel_size
+        self.n_layers = n_layers
+        self.half_channels = channels // 2
+        self.mean_only = mean_only
+
+        self.pre = nn.Conv1d(self.half_channels, hidden_channels, 1)
+        self.enc = (
+            Encoder(
+                hidden_channels,
+                filter_channels,
+                n_heads,
+                n_layers,
+                kernel_size,
+                p_dropout,
+                isflow=True,
+                gin_channels=gin_channels,
+            )
+            if wn_sharing_parameter is None
+            else wn_sharing_parameter
+        )
+        self.post = nn.Conv1d(hidden_channels, self.half_channels * (2 - mean_only), 1)
+        self.post.weight.data.zero_()
+        self.post.bias.data.zero_()
+
+    def forward(self, x, x_mask, g=None, reverse=False):
+        x0, x1 = torch.split(x, [self.half_channels] * 2, 1)
+        h = self.pre(x0) * x_mask
+        h = self.enc(h, x_mask, g=g)
+        stats = self.post(h) * x_mask
+        if not self.mean_only:
+            m, logs = torch.split(stats, [self.half_channels] * 2, 1)
+        else:
+            m = stats
+            logs = torch.zeros_like(m)
+
+        if not reverse:
+            x1 = m + x1 * torch.exp(logs) * x_mask
+            x = torch.cat([x0, x1], 1)
+            logdet = torch.sum(logs, [1, 2])
+            return x, logdet
+        else:
+            x1 = (x1 - m) * torch.exp(-logs) * x_mask
+            x = torch.cat([x0, x1], 1)
+            return x
+
+        x1, logabsdet = piecewise_rational_quadratic_transform(
+            x1,
+            unnormalized_widths,
+            unnormalized_heights,
+            unnormalized_derivatives,
+            inverse=reverse,
+            tails="linear",
+            tail_bound=self.tail_bound,
+        )
+
+        x = torch.cat([x0, x1], 1) * x_mask
+        logdet = torch.sum(logabsdet * x_mask, [1, 2])
+        if not reverse:
+            return x, logdet
+        else:
+            return x

tcp_server/src/lib/tts/monotonic_align/__init__.py

@@ -0,0 +1,16 @@
+from numpy import zeros, int32, float32
+from torch import from_numpy
+
+from .core import maximum_path_jit
+
+
+def maximum_path(neg_cent, mask):
+    device = neg_cent.device
+    dtype = neg_cent.dtype
+    neg_cent = neg_cent.data.cpu().numpy().astype(float32)
+    path = zeros(neg_cent.shape, dtype=int32)
+
+    t_t_max = mask.sum(1)[:, 0].data.cpu().numpy().astype(int32)
+    t_s_max = mask.sum(2)[:, 0].data.cpu().numpy().astype(int32)
+    maximum_path_jit(path, neg_cent, t_t_max, t_s_max)
+    return from_numpy(path).to(device=device, dtype=dtype)

tcp_server/src/lib/tts/monotonic_align/core.py

@@ -0,0 +1,46 @@
+import numba
+
+
+@numba.jit(
+    numba.void(
+        numba.int32[:, :, ::1],
+        numba.float32[:, :, ::1],
+        numba.int32[::1],
+        numba.int32[::1],
+    ),
+    nopython=True,
+    nogil=True,
+)
+def maximum_path_jit(paths, values, t_ys, t_xs):
+    b = paths.shape[0]
+    max_neg_val = -1e9
+    for i in range(int(b)):
+        path = paths[i]
+        value = values[i]
+        t_y = t_ys[i]
+        t_x = t_xs[i]
+
+        v_prev = v_cur = 0.0
+        index = t_x - 1
+
+        for y in range(t_y):
+            for x in range(max(0, t_x + y - t_y), min(t_x, y + 1)):
+                if x == y:
+                    v_cur = max_neg_val
+                else:
+                    v_cur = value[y - 1, x]
+                if x == 0:
+                    if y == 0:
+                        v_prev = 0.0
+                    else:
+                        v_prev = max_neg_val
+                else:
+                    v_prev = value[y - 1, x - 1]
+                value[y, x] += max(v_prev, v_cur)
+
+        for y in range(t_y - 1, -1, -1):
+            path[y, index] = 1
+            if index != 0 and (
+                index == y or value[y - 1, index] < value[y - 1, index - 1]
+            ):
+                index = index - 1

tcp_server/src/lib/tts/split_utils.py

@@ -0,0 +1,168 @@
+import re
+
+def split_sentence(text, min_len=10, language_str='EN'):
+    if language_str in ['EN', 'FR', 'ES', 'SP']:
+        sentences = split_sentences_latin(text, min_len=min_len)
+    else:
+        sentences = split_sentences_zh(text, min_len=min_len)
+    return sentences
+
+
+def split_sentences_latin(text, min_len=10):
+    text = re.sub('[。！？；]', '.', text)
+    text = re.sub('[，]', ',', text)
+    text = re.sub('[“”]', '"', text)
+    text = re.sub('[‘’]', "'", text)
+    text = re.sub(r"[\<\>\(\)\[\]\"\«\»]+", "", text)
+    return [item.strip() for item in txtsplit(text, 256, 512) if item.strip()]
+
+
+def split_sentences_zh(text, min_len=10):
+    text = re.sub('[。！？；]', '.', text)
+    text = re.sub('[，]', ',', text)
+    # 将文本中的换行符、空格和制表符替换为空格
+    text = re.sub('[\n\t ]+', ' ', text)
+    # 在标点符号后添加一个空格
+    text = re.sub('([,.!?;])', r'\1 $#!', text)
+    # 分隔句子并去除前后空格
+    # sentences = [s.strip() for s in re.split('(。|！|？|；)', text)]
+    sentences = [s.strip() for s in text.split('$#!')]
+    if len(sentences[-1]) == 0: del sentences[-1]
+
+    new_sentences = []
+    new_sent = []
+    count_len = 0
+    for ind, sent in enumerate(sentences):
+        new_sent.append(sent)
+        count_len += len(sent)
+        if count_len > min_len or ind == len(sentences) - 1:
+            count_len = 0
+            new_sentences.append(' '.join(new_sent))
+            new_sent = []
+    return merge_short_sentences_zh(new_sentences)
+
+
+def merge_short_sentences_en(sens):
+    """Avoid short sentences by merging them with the following sentence.
+
+    Args:
+        List[str]: list of input sentences.
+
+    Returns:
+        List[str]: list of output sentences.
+    """
+    sens_out = []
+    for s in sens:
+        # If the previous sentense is too short, merge them with
+        # the current sentence.
+        if len(sens_out) > 0 and len(sens_out[-1].split(" ")) <= 2:
+            sens_out[-1] = sens_out[-1] + " " + s
+        else:
+            sens_out.append(s)
+    try:
+        if len(sens_out[-1].split(" ")) <= 2:
+            sens_out[-2] = sens_out[-2] + " " + sens_out[-1]
+            sens_out.pop(-1)
+    except:
+        pass
+    return sens_out
+
+
+def merge_short_sentences_zh(sens):
+    # return sens
+    """Avoid short sentences by merging them with the following sentence.
+
+    Args:
+        List[str]: list of input sentences.
+
+    Returns:
+        List[str]: list of output sentences.
+    """
+    sens_out = []
+    for s in sens:
+        # If the previous sentense is too short, merge them with
+        # the current sentence.
+        if len(sens_out) > 0 and len(sens_out[-1]) <= 2:
+            sens_out[-1] = sens_out[-1] + " " + s
+        else:
+            sens_out.append(s)
+    try:
+        if len(sens_out[-1]) <= 2:
+            sens_out[-2] = sens_out[-2] + " " + sens_out[-1]
+            sens_out.pop(-1)
+    except:
+        pass
+    return sens_out
+
+
+
+def txtsplit(text, desired_length=100, max_length=200):
+    """Split text it into chunks of a desired length trying to keep sentences intact."""
+    text = re.sub(r'\n\n+', '\n', text)
+    text = re.sub(r'\s+', ' ', text)
+    text = re.sub(r'[""]', '"', text)
+    text = re.sub(r'([,.?!])', r'\1 ', text)
+    text = re.sub(r'\s+', ' ', text)
+    
+    rv = []
+    in_quote = False
+    current = ""
+    split_pos = []
+    pos = -1
+    end_pos = len(text) - 1
+    def seek(delta):
+        nonlocal pos, in_quote, current
+        is_neg = delta < 0
+        for _ in range(abs(delta)):
+            if is_neg:
+                pos -= 1
+                current = current[:-1]
+            else:
+                pos += 1
+                current += text[pos]
+            if text[pos] == '"':
+                in_quote = not in_quote
+        return text[pos]
+    def peek(delta):
+        p = pos + delta
+        return text[p] if p < end_pos and p >= 0 else ""
+    def commit():
+        nonlocal rv, current, split_pos
+        rv.append(current)
+        current = ""
+        split_pos = []
+    while pos < end_pos:
+        c = seek(1)
+        if len(current) >= max_length:
+            if len(split_pos) > 0 and len(current) > (desired_length / 2):
+                d = pos - split_pos[-1]
+                seek(-d)
+            else:
+                while c not in '!?.\n ' and pos > 0 and len(current) > desired_length:
+                    c = seek(-1)
+            commit()
+        elif not in_quote and (c in '!?\n' or (c in '.,' and peek(1) in '\n ')):
+            while pos < len(text) - 1 and len(current) < max_length and peek(1) in '!?.':
+                c = seek(1)
+            split_pos.append(pos)
+            if len(current) >= desired_length:
+                commit()
+        elif in_quote and peek(1) == '"' and peek(2) in '\n ':
+            seek(2)
+            split_pos.append(pos)
+    rv.append(current)
+    rv = [s.strip() for s in rv]
+    rv = [s for s in rv if len(s) > 0 and not re.match(r'^[\s\.,;:!?]*$', s)]
+    return rv
+
+
+if __name__ == '__main__':
+    zh_text = "好的，我来给你讲一个故事吧。从前有一个小姑娘，她叫做小红。小红非常喜欢在森林里玩耍，她经常会和她的小伙伴们一起去探险。有一天，小红和她的小伙伴们走到了森林深处，突然遇到了一只凶猛的野兽。小红的小伙伴们都吓得不敢动弹，但是小红并没有被吓倒，她勇敢地走向野兽，用她的智慧和勇气成功地制服了野兽，保护了她的小伙伴们。从那以后，小红变得更加勇敢和自信，成为了她小伙伴们心中的英雄。"
+    en_text = "I didn’t know what to do. I said please kill her because it would be better than being kidnapped,” Ben, whose surname CNN is not using for security concerns, said on Wednesday. “It’s a nightmare. I said ‘please kill her, don’t take her there.’"
+    sp_text = "¡Claro! ¿En qué tema te gustaría que te hable en español? Puedo proporcionarte información o conversar contigo sobre una amplia variedad de temas, desde cultura y comida hasta viajes y tecnología. ¿Tienes alguna preferencia en particular?"
+    fr_text = "Bien sûr ! En quelle matière voudriez-vous que je vous parle en français ? Je peux vous fournir des informations ou discuter avec vous sur une grande variété de sujets, que ce soit la culture, la nourriture, les voyages ou la technologie. Avez-vous une préférence particulière ?"
+
+    print(split_sentence(zh_text, language_str='ZH'))
+    print(split_sentence(en_text, language_str='EN'))
+    print(split_sentence(sp_text, language_str='SP'))
+    print(split_sentence(fr_text, language_str='FR'))

tcp_server/src/lib/tts/text/__init__.py

@@ -0,0 +1,29 @@
+from .symbols import *
+
+
+_symbol_to_id = {s: i for i, s in enumerate(symbols)}
+
+
+def cleaned_text_to_sequence(cleaned_text, tones, language, symbol_to_id=None):
+    """Converts a string of text to a sequence of IDs corresponding to the symbols in the text.
+    Args:
+      text: string to convert to a sequence
+    Returns:
+      List of integers corresponding to the symbols in the text
+    """
+    symbol_to_id_map = symbol_to_id if symbol_to_id else _symbol_to_id
+    phones = [symbol_to_id_map[symbol] for symbol in cleaned_text]
+    tone_start = language_tone_start_map[language]
+    tones = [i + tone_start for i in tones]
+    lang_id = language_id_map[language]
+    lang_ids = [lang_id for i in phones]
+    return phones, tones, lang_ids
+
+
+def get_bert(norm_text, word2ph, language, device):
+    from .english_bert import get_bert_feature as en_bert
+    # from .french_bert import get_bert_feature as fr_bert
+
+    lang_bert_func_map = {"EN": en_bert}
+    bert = lang_bert_func_map[language](norm_text, word2ph, device)
+    return bert

tcp_server/src/lib/tts/text/cleaner.py

@@ -0,0 +1,37 @@
+from . import english
+from . import cleaned_text_to_sequence
+import copy
+
+# language_module_map = {"EN": english,
+#                     'FR': french}
+language_module_map = {"EN": english}
+
+
+def clean_text(text, language):
+    language_module = language_module_map[language]
+    norm_text = language_module.text_normalize(text)
+    phones, tones, word2ph = language_module.g2p(norm_text)
+    return norm_text, phones, tones, word2ph
+
+
+def clean_text_bert(text, language, device=None):
+    language_module = language_module_map[language]
+    norm_text = language_module.text_normalize(text)
+    phones, tones, word2ph = language_module.g2p(norm_text)
+    
+    word2ph_bak = copy.deepcopy(word2ph)
+    for i in range(len(word2ph)):
+        word2ph[i] = word2ph[i] * 2
+    word2ph[0] += 1
+    bert = language_module.get_bert_feature(norm_text, word2ph, device=device)
+    
+    return norm_text, phones, tones, word2ph_bak, bert
+
+
+def text_to_sequence(text, language):
+    norm_text, phones, tones, word2ph = clean_text(text, language)
+    return cleaned_text_to_sequence(phones, tones, language)
+
+
+if __name__ == "__main__":
+    pass

tcp_server/src/lib/tts/text/cleaner_multiling.py

@@ -0,0 +1,110 @@
+"""Set of default text cleaners"""
+# TODO: pick the cleaner for languages dynamically
+
+import re
+
+# Regular expression matching whitespace:
+_whitespace_re = re.compile(r"\s+")
+
+rep_map = {
+    "：": ",",
+    "；": ",",
+    "，": ",",
+    "。": ".",
+    "！": "!",
+    "？": "?",
+    "\n": ".",
+    "·": ",",
+    "、": ",",
+    "...": ".",
+    "…": ".",
+    "$": ".",
+    "“": "'",
+    "”": "'",
+    "‘": "'",
+    "’": "'",
+    "（": "'",
+    "）": "'",
+    "(": "'",
+    ")": "'",
+    "《": "'",
+    "》": "'",
+    "【": "'",
+    "】": "'",
+    "[": "'",
+    "]": "'",
+    "—": "",
+    "～": "-",
+    "~": "-",
+    "「": "'",
+    "」": "'",
+}
+
+def replace_punctuation(text):
+    pattern = re.compile("|".join(re.escape(p) for p in rep_map.keys()))
+    replaced_text = pattern.sub(lambda x: rep_map[x.group()], text)
+    return replaced_text
+
+def lowercase(text):
+    return text.lower()
+
+
+def collapse_whitespace(text):
+    return re.sub(_whitespace_re, " ", text).strip()
+
+def remove_punctuation_at_begin(text):
+    return re.sub(r'^[,.!?]+', '', text)
+
+def remove_aux_symbols(text):
+    text = re.sub(r"[\<\>\(\)\[\]\"\«\»\']+", "", text)
+    return text
+
+
+def replace_symbols(text, lang="en"):
+    """Replace symbols based on the lenguage tag.
+
+    Args:
+      text:
+       Input text.
+      lang:
+        Lenguage identifier. ex: "en", "fr", "pt", "ca".
+
+    Returns:
+      The modified text
+      example:
+        input args:
+            text: "si l'avi cau, diguem-ho"
+            lang: "ca"
+        Output:
+            text: "si lavi cau, diguemho"
+    """
+    text = text.replace(";", ",")
+    text = text.replace("-", " ") if lang != "ca" else text.replace("-", "")
+    text = text.replace(":", ",")
+    if lang == "en":
+        text = text.replace("&", " and ")
+    elif lang == "fr":
+        text = text.replace("&", " et ")
+    elif lang == "pt":
+        text = text.replace("&", " e ")
+    elif lang == "ca":
+        text = text.replace("&", " i ")
+        text = text.replace("'", "")
+    elif lang== "es":
+        text=text.replace("&","y")
+        text = text.replace("'", "")
+    return text
+
+def unicleaners(text, cased=False, lang='en'):
+    """Basic pipeline for Portuguese text. There is no need to expand abbreviation and
+    numbers, phonemizer already does that"""
+    if not cased:
+        text = lowercase(text)
+    text = replace_punctuation(text)
+    text = replace_symbols(text, lang=lang)
+    text = remove_aux_symbols(text)
+    text = remove_punctuation_at_begin(text)
+    text = collapse_whitespace(text)
+    text = re.sub(r'([^\.,!\?\-…])$', r'\1.', text)
+    return text
+

tcp_server/src/lib/tts/text/english.py

@@ -0,0 +1,298 @@
+import pickle
+import os
+import re
+from g2p_en import G2p
+from transformers import AutoTokenizer
+
+
+from lib.constants import TTS_BERT_BASE_MODEL_DIR_PATH
+
+from . import symbols
+
+from .english_utils.abbreviations import expand_abbreviations
+from .english_utils.time_norm import expand_time_english
+from .english_utils.number_norm import normalize_numbers
+
+
+current_file_path = os.path.dirname(__file__)
+CMU_DICT_PATH = os.path.join(current_file_path, "cmudict.rep")
+CACHE_PATH = os.path.join(current_file_path, "cmudict_cache.pickle")
+_g2p = G2p()
+
+arpa = {
+    "AH0",
+    "S",
+    "AH1",
+    "EY2",
+    "AE2",
+    "EH0",
+    "OW2",
+    "UH0",
+    "NG",
+    "B",
+    "G",
+    "AY0",
+    "M",
+    "AA0",
+    "F",
+    "AO0",
+    "ER2",
+    "UH1",
+    "IY1",
+    "AH2",
+    "DH",
+    "IY0",
+    "EY1",
+    "IH0",
+    "K",
+    "N",
+    "W",
+    "IY2",
+    "T",
+    "AA1",
+    "ER1",
+    "EH2",
+    "OY0",
+    "UH2",
+    "UW1",
+    "Z",
+    "AW2",
+    "AW1",
+    "V",
+    "UW2",
+    "AA2",
+    "ER",
+    "AW0",
+    "UW0",
+    "R",
+    "OW1",
+    "EH1",
+    "ZH",
+    "AE0",
+    "IH2",
+    "IH",
+    "Y",
+    "JH",
+    "P",
+    "AY1",
+    "EY0",
+    "OY2",
+    "TH",
+    "HH",
+    "D",
+    "ER0",
+    "CH",
+    "AO1",
+    "AE1",
+    "AO2",
+    "OY1",
+    "AY2",
+    "IH1",
+    "OW0",
+    "L",
+    "SH",
+}
+
+
+def distribute_phone(n_phone, n_word):
+    phones_per_word = [0] * n_word
+    for task in range(n_phone):
+        min_tasks = min(phones_per_word)
+        min_index = phones_per_word.index(min_tasks)
+        phones_per_word[min_index] += 1
+    return phones_per_word
+
+def post_replace_ph(ph):
+    rep_map = {
+        "：": ",",
+        "；": ",",
+        "，": ",",
+        "。": ".",
+        "！": "!",
+        "？": "?",
+        "\n": ".",
+        "·": ",",
+        "、": ",",
+        "...": "…",
+        "v": "V",
+    }
+    if ph in rep_map.keys():
+        ph = rep_map[ph]
+    if ph in symbols:
+        return ph
+    if ph not in symbols:
+        ph = "UNK"
+    return ph
+
+
+def read_dict():
+    g2p_dict = {}
+    start_line = 49
+    with open(CMU_DICT_PATH) as f:
+        line = f.readline()
+        line_index = 1
+        while line:
+            if line_index >= start_line:
+                line = line.strip()
+                word_split = line.split("  ")
+                word = word_split[0]
+
+                syllable_split = word_split[1].split(" - ")
+                g2p_dict[word] = []
+                for syllable in syllable_split:
+                    phone_split = syllable.split(" ")
+                    g2p_dict[word].append(phone_split)
+
+            line_index = line_index + 1
+            line = f.readline()
+
+    return g2p_dict
+
+
+def cache_dict(g2p_dict, file_path):
+    with open(file_path, "wb") as pickle_file:
+        pickle.dump(g2p_dict, pickle_file)
+
+
+def get_dict():
+    if os.path.exists(CACHE_PATH):
+        with open(CACHE_PATH, "rb") as pickle_file:
+            g2p_dict = pickle.load(pickle_file)
+    else:
+        g2p_dict = read_dict()
+        cache_dict(g2p_dict, CACHE_PATH)
+
+    return g2p_dict
+
+
+eng_dict = get_dict()
+
+
+def refine_ph(phn):
+    tone = 0
+    if re.search(r"\d$", phn):
+        tone = int(phn[-1]) + 1
+        phn = phn[:-1]
+    return phn.lower(), tone
+
+
+def refine_syllables(syllables):
+    tones = []
+    phonemes = []
+    for phn_list in syllables:
+        for i in range(len(phn_list)):
+            phn = phn_list[i]
+            phn, tone = refine_ph(phn)
+            phonemes.append(phn)
+            tones.append(tone)
+    return phonemes, tones
+
+
+def text_normalize(text):
+    text = text.lower()
+    text = expand_time_english(text)
+    text = normalize_numbers(text)
+    text = expand_abbreviations(text)
+    return text
+
+load_model_params = {
+    "pretrained_model_name_or_path": TTS_BERT_BASE_MODEL_DIR_PATH,
+    "local_files_only": True
+}
+tokenizer = AutoTokenizer.from_pretrained(**load_model_params)
+
+def g2p_old(text):
+    tokenized = tokenizer.tokenize(text)
+    # import pdb; pdb.set_trace()
+    phones = []
+    tones = []
+    words = re.split(r"([,;.\-\?\!\s+])", text)
+    for w in words:
+        if w.upper() in eng_dict:
+            phns, tns = refine_syllables(eng_dict[w.upper()])
+            phones += phns
+            tones += tns
+        else:
+            phone_list = list(filter(lambda p: p != " ", _g2p(w)))
+            for ph in phone_list:
+                if ph in arpa:
+                    ph, tn = refine_ph(ph)
+                    phones.append(ph)
+                    tones.append(tn)
+                else:
+                    phones.append(ph)
+                    tones.append(0)
+    # todo: implement word2ph
+    word2ph = [1 for i in phones]
+
+    phones = [post_replace_ph(i) for i in phones]
+    return phones, tones, word2ph
+
+def g2p(text, pad_start_end=True, tokenized=None):
+    if tokenized is None:
+        tokenized = tokenizer.tokenize(text)
+    # import pdb; pdb.set_trace()
+    phs = []
+    ph_groups = []
+    for t in tokenized:
+        if not t.startswith("#"):
+            ph_groups.append([t])
+        else:
+            ph_groups[-1].append(t.replace("#", ""))
+    
+    phones = []
+    tones = []
+    word2ph = []
+    for group in ph_groups:
+        w = "".join(group)
+        phone_len = 0
+        word_len = len(group)
+        if w.upper() in eng_dict:
+            phns, tns = refine_syllables(eng_dict[w.upper()])
+            phones += phns
+            tones += tns
+            phone_len += len(phns)
+        else:
+            phone_list = list(filter(lambda p: p != " ", _g2p(w)))
+            for ph in phone_list:
+                if ph in arpa:
+                    ph, tn = refine_ph(ph)
+                    phones.append(ph)
+                    tones.append(tn)
+                else:
+                    phones.append(ph)
+                    tones.append(0)
+                phone_len += 1
+        aaa = distribute_phone(phone_len, word_len)
+        word2ph += aaa
+    phones = [post_replace_ph(i) for i in phones]
+
+    if pad_start_end:
+        phones = ["_"] + phones + ["_"]
+        tones = [0] + tones + [0]
+        word2ph = [1] + word2ph + [1]
+    return phones, tones, word2ph
+
+def get_bert_feature(text, word2ph, device=None):
+    from text import english_bert
+
+    return english_bert.get_bert_feature(text, word2ph, device=device)
+
+if __name__ == "__main__":
+    # print(get_dict())
+    # print(eng_word_to_phoneme("hello"))
+    from text.english_bert import get_bert_feature
+    text = "In this paper, we propose 1 DSPGAN, a N-F-T GAN-based universal vocoder."
+    text = text_normalize(text)
+    phones, tones, word2ph = g2p(text)
+    import pdb; pdb.set_trace()
+    bert = get_bert_feature(text, word2ph)
+    
+    print(phones, tones, word2ph, bert.shape)
+
+    # all_phones = set()
+    # for k, syllables in eng_dict.items():
+    #     for group in syllables:
+    #         for ph in group:
+    #             all_phones.add(ph)
+    # print(all_phones)

tcp_server/src/lib/tts/text/english_bert.py

@@ -0,0 +1,44 @@
+import torch
+from transformers import AutoTokenizer, AutoModelForMaskedLM
+import sys
+
+from lib.constants import TTS_BERT_BASE_MODEL_DIR_PATH
+
+load_model_params = {
+    "pretrained_model_name_or_path": TTS_BERT_BASE_MODEL_DIR_PATH,
+    "local_files_only": True
+}
+tokenizer = AutoTokenizer.from_pretrained(**load_model_params)
+model = None
+
+def get_bert_feature(text, word2ph, device=None):
+    global model
+    if (
+        sys.platform == "darwin"
+        and torch.backends.mps.is_available()
+        and device == "cpu"
+    ):
+        device = "mps"
+    if not device:
+        device = "cuda"
+    if model is None:
+        model = AutoModelForMaskedLM.from_pretrained(**load_model_params).to(
+            device
+        )
+    with torch.no_grad():
+        inputs = tokenizer(text, return_tensors="pt")
+        for i in inputs:
+            inputs[i] = inputs[i].to(device)
+        res = model(**inputs, output_hidden_states=True)
+        res = torch.cat(res["hidden_states"][-3:-2], -1)[0].cpu()
+        
+    assert inputs["input_ids"].shape[-1] == len(word2ph)
+    word2phone = word2ph
+    phone_level_feature = []
+    for i in range(len(word2phone)):
+        repeat_feature = res[i].repeat(word2phone[i], 1)
+        phone_level_feature.append(repeat_feature)
+
+    phone_level_feature = torch.cat(phone_level_feature, dim=0)
+
+    return phone_level_feature.T

tcp_server/src/lib/tts/text/english_utils/abbreviations.py

@@ -0,0 +1,35 @@
+import re
+
+# List of (regular expression, replacement) pairs for abbreviations in english:
+abbreviations_en = [
+    (re.compile("\\b%s\\." % x[0], re.IGNORECASE), x[1])
+    for x in [
+        ("mrs", "misess"),
+        ("mr", "mister"),
+        ("dr", "doctor"),
+        ("st", "saint"),
+        ("co", "company"),
+        ("jr", "junior"),
+        ("maj", "major"),
+        ("gen", "general"),
+        ("drs", "doctors"),
+        ("rev", "reverend"),
+        ("lt", "lieutenant"),
+        ("hon", "honorable"),
+        ("sgt", "sergeant"),
+        ("capt", "captain"),
+        ("esq", "esquire"),
+        ("ltd", "limited"),
+        ("col", "colonel"),
+        ("ft", "fort"),
+    ]
+]
+
+def expand_abbreviations(text, lang="en"):
+    if lang == "en":
+        _abbreviations = abbreviations_en
+    else:
+        raise NotImplementedError()
+    for regex, replacement in _abbreviations:
+        text = re.sub(regex, replacement, text)
+    return text

tcp_server/src/lib/tts/text/english_utils/number_norm.py

@@ -0,0 +1,97 @@
+""" from https://github.com/keithito/tacotron """
+
+import re
+from typing import Dict
+
+import inflect
+
+_inflect = inflect.engine()
+_comma_number_re = re.compile(r"([0-9][0-9\,]+[0-9])")
+_decimal_number_re = re.compile(r"([0-9]+\.[0-9]+)")
+_currency_re = re.compile(r"(£|\$|¥)([0-9\,\.]*[0-9]+)")
+_ordinal_re = re.compile(r"[0-9]+(st|nd|rd|th)")
+_number_re = re.compile(r"-?[0-9]+")
+
+
+def _remove_commas(m):
+    return m.group(1).replace(",", "")
+
+
+def _expand_decimal_point(m):
+    return m.group(1).replace(".", " point ")
+
+
+def __expand_currency(value: str, inflection: Dict[float, str]) -> str:
+    parts = value.replace(",", "").split(".")
+    if len(parts) > 2:
+        return f"{value} {inflection[2]}"  # Unexpected format
+    text = []
+    integer = int(parts[0]) if parts[0] else 0
+    if integer > 0:
+        integer_unit = inflection.get(integer, inflection[2])
+        text.append(f"{integer} {integer_unit}")
+    fraction = int(parts[1]) if len(parts) > 1 and parts[1] else 0
+    if fraction > 0:
+        fraction_unit = inflection.get(fraction / 100, inflection[0.02])
+        text.append(f"{fraction} {fraction_unit}")
+    if len(text) == 0:
+        return f"zero {inflection[2]}"
+    return " ".join(text)
+
+
+def _expand_currency(m: "re.Match") -> str:
+    currencies = {
+        "$": {
+            0.01: "cent",
+            0.02: "cents",
+            1: "dollar",
+            2: "dollars",
+        },
+        "€": {
+            0.01: "cent",
+            0.02: "cents",
+            1: "euro",
+            2: "euros",
+        },
+        "£": {
+            0.01: "penny",
+            0.02: "pence",
+            1: "pound sterling",
+            2: "pounds sterling",
+        },
+        "¥": {
+            # TODO rin
+            0.02: "sen",
+            2: "yen",
+        },
+    }
+    unit = m.group(1)
+    currency = currencies[unit]
+    value = m.group(2)
+    return __expand_currency(value, currency)
+
+
+def _expand_ordinal(m):
+    return _inflect.number_to_words(m.group(0))
+
+
+def _expand_number(m):
+    num = int(m.group(0))
+    if 1000 < num < 3000:
+        if num == 2000:
+            return "two thousand"
+        if 2000 < num < 2010:
+            return "two thousand " + _inflect.number_to_words(num % 100)
+        if num % 100 == 0:
+            return _inflect.number_to_words(num // 100) + " hundred"
+        return _inflect.number_to_words(num, andword="", zero="oh", group=2).replace(", ", " ")
+    return _inflect.number_to_words(num, andword="")
+
+
+def normalize_numbers(text):
+    text = re.sub(_comma_number_re, _remove_commas, text)
+    text = re.sub(_currency_re, _expand_currency, text)
+    text = re.sub(_decimal_number_re, _expand_decimal_point, text)
+    text = re.sub(_ordinal_re, _expand_ordinal, text)
+    text = re.sub(_number_re, _expand_number, text)
+    return text

tcp_server/src/lib/tts/text/english_utils/time_norm.py

@@ -0,0 +1,47 @@
+import re
+
+import inflect
+
+_inflect = inflect.engine()
+
+_time_re = re.compile(
+    r"""\b
+                          ((0?[0-9])|(1[0-1])|(1[2-9])|(2[0-3]))  # hours
+                          :
+                          ([0-5][0-9])                            # minutes
+                          \s*(a\\.m\\.|am|pm|p\\.m\\.|a\\.m|p\\.m)? # am/pm
+                          \b""",
+    re.IGNORECASE | re.X,
+)
+
+
+def _expand_num(n: int) -> str:
+    return _inflect.number_to_words(n)
+
+
+def _expand_time_english(match: "re.Match") -> str:
+    hour = int(match.group(1))
+    past_noon = hour >= 12
+    time = []
+    if hour > 12:
+        hour -= 12
+    elif hour == 0:
+        hour = 12
+        past_noon = True
+    time.append(_expand_num(hour))
+
+    minute = int(match.group(6))
+    if minute > 0:
+        if minute < 10:
+            time.append("oh")
+        time.append(_expand_num(minute))
+    am_pm = match.group(7)
+    if am_pm is None:
+        time.append("p m" if past_noon else "a m")
+    else:
+        time.extend(list(am_pm.replace(".", "")))
+    return " ".join(time)
+
+
+def expand_time_english(text: str) -> str:
+    return re.sub(_time_re, _expand_time_english, text)

tcp_server/src/lib/tts/text/fr_phonemizer/base.py

@@ -0,0 +1,140 @@
+import abc
+from typing import List, Tuple
+
+from .punctuation import Punctuation
+
+
+class BasePhonemizer(abc.ABC):
+    """Base phonemizer class
+
+    Phonemization follows the following steps:
+        1. Preprocessing:
+            - remove empty lines
+            - remove punctuation
+            - keep track of punctuation marks
+
+        2. Phonemization:
+            - convert text to phonemes
+
+        3. Postprocessing:
+            - join phonemes
+            - restore punctuation marks
+
+    Args:
+        language (str):
+            Language used by the phonemizer.
+
+        punctuations (List[str]):
+            List of punctuation marks to be preserved.
+
+        keep_puncs (bool):
+            Whether to preserve punctuation marks or not.
+    """
+
+    def __init__(self, language, punctuations=Punctuation.default_puncs(), keep_puncs=False):
+        # ensure the backend is installed on the system
+        if not self.is_available():
+            raise RuntimeError("{} not installed on your system".format(self.name()))  # pragma: nocover
+
+        # ensure the backend support the requested language
+        self._language = self._init_language(language)
+
+        # setup punctuation processing
+        self._keep_puncs = keep_puncs
+        self._punctuator = Punctuation(punctuations)
+
+    def _init_language(self, language):
+        """Language initialization
+
+        This method may be overloaded in child classes (see Segments backend)
+
+        """
+        if not self.is_supported_language(language):
+            raise RuntimeError(f'language "{language}" is not supported by the ' f"{self.name()} backend")
+        return language
+
+    @property
+    def language(self):
+        """The language code configured to be used for phonemization"""
+        return self._language
+
+    @staticmethod
+    @abc.abstractmethod
+    def name():
+        """The name of the backend"""
+        ...
+
+    @classmethod
+    @abc.abstractmethod
+    def is_available(cls):
+        """Returns True if the backend is installed, False otherwise"""
+        ...
+
+    @classmethod
+    @abc.abstractmethod
+    def version(cls):
+        """Return the backend version as a tuple (major, minor, patch)"""
+        ...
+
+    @staticmethod
+    @abc.abstractmethod
+    def supported_languages():
+        """Return a dict of language codes -> name supported by the backend"""
+        ...
+
+    def is_supported_language(self, language):
+        """Returns True if `language` is supported by the backend"""
+        return language in self.supported_languages()
+
+    @abc.abstractmethod
+    def _phonemize(self, text, separator):
+        """The main phonemization method"""
+
+    def _phonemize_preprocess(self, text) -> Tuple[List[str], List]:
+        """Preprocess the text before phonemization
+
+        1. remove spaces
+        2. remove punctuation
+
+        Override this if you need a different behaviour
+        """
+        text = text.strip()
+        if self._keep_puncs:
+            # a tuple (text, punctuation marks)
+            return self._punctuator.strip_to_restore(text)
+        return [self._punctuator.strip(text)], []
+
+    def _phonemize_postprocess(self, phonemized, punctuations) -> str:
+        """Postprocess the raw phonemized output
+
+        Override this if you need a different behaviour
+        """
+        if self._keep_puncs:
+            return self._punctuator.restore(phonemized, punctuations)[0]
+        return phonemized[0]
+
+    def phonemize(self, text: str, separator="|", language: str = None) -> str:  # pylint: disable=unused-argument
+        """Returns the `text` phonemized for the given language
+
+        Args:
+            text (str):
+                Text to be phonemized.
+
+            separator (str):
+                string separator used between phonemes. Default to '_'.
+
+        Returns:
+            (str): Phonemized text
+        """
+        text, punctuations = self._phonemize_preprocess(text)
+        phonemized = []
+        for t in text:
+            p = self._phonemize(t, separator)
+            phonemized.append(p)
+        phonemized = self._phonemize_postprocess(phonemized, punctuations)
+        return phonemized
+
+    def print_logs(self, level: int = 0):
+        indent = "\t" * level
+        print(f"{indent}| > phoneme language: {self.language}")
+        print(f"{indent}| > phoneme backend: {self.name()}")

tcp_server/src/lib/tts/text/fr_phonemizer/cleaner.py

@@ -0,0 +1,122 @@
+"""Set of default text cleaners"""
+# TODO: pick the cleaner for languages dynamically
+
+import re
+from .french_abbreviations import abbreviations_fr
+
+# Regular expression matching whitespace:
+_whitespace_re = re.compile(r"\s+")
+
+
+rep_map = {
+    "：": ",",
+    "；": ",",
+    "，": ",",
+    "。": ".",
+    "！": "!",
+    "？": "?",
+    "\n": ".",
+    "·": ",",
+    "、": ",",
+    "...": ".",
+    "…": ".",
+    "$": ".",
+    "“": "",
+    "”": "",
+    "‘": "",
+    "’": "",
+    "（": "",
+    "）": "",
+    "(": "",
+    ")": "",
+    "《": "",
+    "》": "",
+    "【": "",
+    "】": "",
+    "[": "",
+    "]": "",
+    "—": "",
+    "～": "-",
+    "~": "-",
+    "「": "",
+    "」": "",
+    "¿" : "",
+    "¡" : ""
+}
+
+
+def replace_punctuation(text):
+    pattern = re.compile("|".join(re.escape(p) for p in rep_map.keys()))
+    replaced_text = pattern.sub(lambda x: rep_map[x.group()], text)
+    return replaced_text
+
+def expand_abbreviations(text, lang="fr"):
+    if lang == "fr":
+        _abbreviations = abbreviations_fr
+    for regex, replacement in _abbreviations:
+        text = re.sub(regex, replacement, text)
+    return text
+
+
+def lowercase(text):
+    return text.lower()
+
+
+def collapse_whitespace(text):
+    return re.sub(_whitespace_re, " ", text).strip()
+
+def remove_punctuation_at_begin(text):
+    return re.sub(r'^[,.!?]+', '', text)
+
+def remove_aux_symbols(text):
+    text = re.sub(r"[\<\>\(\)\[\]\"\«\»]+", "", text)
+    return text
+
+
+def replace_symbols(text, lang="en"):
+    """Replace symbols based on the lenguage tag.
+
+    Args:
+      text:
+       Input text.
+      lang:
+        Lenguage identifier. ex: "en", "fr", "pt", "ca".
+
+    Returns:
+      The modified text
+      example:
+        input args:
+            text: "si l'avi cau, diguem-ho"
+            lang: "ca"
+        Output:
+            text: "si lavi cau, diguemho"
+    """
+    text = text.replace(";", ",")
+    text = text.replace("-", " ") if lang != "ca" else text.replace("-", "")
+    text = text.replace(":", ",")
+    if lang == "en":
+        text = text.replace("&", " and ")
+    elif lang == "fr":
+        text = text.replace("&", " et ")
+    elif lang == "pt":
+        text = text.replace("&", " e ")
+    elif lang == "ca":
+        text = text.replace("&", " i ")
+        text = text.replace("'", "")
+    elif lang== "es":
+        text=text.replace("&","y")
+        text = text.replace("'", "")
+    return text
+
+def french_cleaners(text):
+    """Pipeline for French text. There is no need to expand numbers, phonemizer already does that"""
+    text = expand_abbreviations(text, lang="fr")
+    # text = lowercase(text) # as we use the cased bert
+    text = replace_punctuation(text)
+    text = replace_symbols(text, lang="fr")
+    text = remove_aux_symbols(text)
+    text = remove_punctuation_at_begin(text)
+    text = collapse_whitespace(text)
+    text = re.sub(r'([^\.,!\?\-…])$', r'\1.', text)
+    return text
+

tcp_server/src/lib/tts/text/fr_phonemizer/en_symbols.json

@@ -0,0 +1,79 @@
+{
+  "symbols": [
+    "_",
+    ",",
+    ".",
+    "!",
+    "?",
+    "-",
+    "~",
+    "\u2026",
+    "N",
+    "Q",
+    "a",
+    "b",
+    "d",
+    "e",
+    "f",
+    "g",
+    "h",
+    "i",
+    "j",
+    "k",
+    "l",
+    "m",
+    "n",
+    "o",
+    "p",
+    "s",
+    "t",
+    "u",
+    "v",
+    "w",
+    "x",
+    "y",
+    "z",
+    "\u0251",
+    "\u00e6",
+    "\u0283",
+    "\u0291",
+    "\u00e7",
+    "\u026f",
+    "\u026a",
+    "\u0254",
+    "\u025b",
+    "\u0279",
+    "\u00f0",
+    "\u0259",
+    "\u026b",
+    "\u0265",
+    "\u0278",
+    "\u028a",
+    "\u027e",
+    "\u0292",
+    "\u03b8",
+    "\u03b2",
+    "\u014b",
+    "\u0266",
+    "\u207c",
+    "\u02b0",
+    "`",
+    "^",
+    "#",
+    "*",
+    "=",
+    "\u02c8",
+    "\u02cc",
+    "\u2192",
+    "\u2193",
+    "\u2191",
+    " ",
+    "ɣ",
+    "ɡ",
+    "r",
+    "ɲ",
+    "ʝ",
+    "ʎ",
+    "ː"
+  ]
+}

tcp_server/src/lib/tts/text/fr_phonemizer/fr_symbols.json

@@ -0,0 +1,89 @@
+{
+  "symbols": [
+    "_",
+    ",",
+    ".",
+    "!",
+    "?",
+    "-",
+    "~",
+    "\u2026",
+    "N",
+    "Q",
+    "a",
+    "b",
+    "d",
+    "e",
+    "f",
+    "g",
+    "h",
+    "i",
+    "j",
+    "k",
+    "l",
+    "m",
+    "n",
+    "o",
+    "p",
+    "s",
+    "t",
+    "u",
+    "v",
+    "w",
+    "x",
+    "y",
+    "z",
+    "\u0251",
+    "\u00e6",
+    "\u0283",
+    "\u0291",
+    "\u00e7",
+    "\u026f",
+    "\u026a",
+    "\u0254",
+    "\u025b",
+    "\u0279",
+    "\u00f0",
+    "\u0259",
+    "\u026b",
+    "\u0265",
+    "\u0278",
+    "\u028a",
+    "\u027e",
+    "\u0292",
+    "\u03b8",
+    "\u03b2",
+    "\u014b",
+    "\u0266",
+    "\u207c",
+    "\u02b0",
+    "`",
+    "^",
+    "#",
+    "*",
+    "=",
+    "\u02c8",
+    "\u02cc",
+    "\u2192",
+    "\u2193",
+    "\u2191",
+    " ",
+    "\u0263",
+    "\u0261",
+    "r",
+    "\u0272",
+    "\u029d",
+    "\u028e",
+    "\u02d0",
+
+    "\u0303",
+    "\u0153",
+    "\u00f8",
+    "\u0281",
+    "\u0252",
+    "\u028c",
+    "\u2014",
+    "\u025c",
+    "\u0250"
+  ]
+}

tcp_server/src/lib/tts/text/fr_phonemizer/fr_to_ipa.py

@@ -0,0 +1,30 @@
+from .cleaner import french_cleaners
+from .gruut_wrapper import Gruut
+
+
+def remove_consecutive_t(input_str):
+    result = []
+    count = 0
+
+    for char in input_str:
+        if char == 't':
+            count += 1
+        else:
+            if count < 3:  
+                result.extend(['t'] * count)
+            count = 0
+            result.append(char)
+
+    if count < 3:
+        result.extend(['t'] * count)
+
+    return ''.join(result)
+
+def fr2ipa(text):
+    e = Gruut(language="fr-fr", keep_puncs=True, keep_stress=True, use_espeak_phonemes=True)
+    # text = french_cleaners(text)
+    phonemes = e.phonemize(text, separator="")
+    # print(phonemes)
+    phonemes = remove_consecutive_t(phonemes)
+    # print(phonemes)
+    return phonemes

tcp_server/src/lib/tts/text/fr_phonemizer/french_abbreviations.py

@@ -0,0 +1,48 @@
+import re
+
+# List of (regular expression, replacement) pairs for abbreviations in french:
+abbreviations_fr = [
+    (re.compile("\\b%s\\." % x[0], re.IGNORECASE), x[1])
+    for x in [
+        ("M", "monsieur"),
+        ("Mlle", "mademoiselle"),
+        ("Mlles", "mesdemoiselles"),
+        ("Mme", "Madame"),
+        ("Mmes", "Mesdames"),
+        ("N.B", "nota bene"),
+        ("M", "monsieur"),
+        ("p.c.q", "parce que"),
+        ("Pr", "professeur"),
+        ("qqch", "quelque chose"),
+        ("rdv", "rendez-vous"),
+        ("max", "maximum"),
+        ("min", "minimum"),
+        ("no", "numéro"),
+        ("adr", "adresse"),
+        ("dr", "docteur"),
+        ("st", "saint"),
+        ("co", "companie"),
+        ("jr", "junior"),
+        ("sgt", "sergent"),
+        ("capt", "capitain"),
+        ("col", "colonel"),
+        ("av", "avenue"),
+        ("av. J.-C", "avant Jésus-Christ"),
+        ("apr. J.-C", "après Jésus-Christ"),
+        ("art", "article"),
+        ("boul", "boulevard"),
+        ("c.-à-d", "c’est-à-dire"),
+        ("etc", "et cetera"),
+        ("ex", "exemple"),
+        ("excl", "exclusivement"),
+        ("boul", "boulevard"),
+    ]
+] + [
+    (re.compile("\\b%s" % x[0]), x[1])
+    for x in [
+        ("Mlle", "mademoiselle"),
+        ("Mlles", "mesdemoiselles"),
+        ("Mme", "Madame"),
+        ("Mmes", "Mesdames"),
+    ]
+]

tcp_server/src/lib/tts/text/fr_phonemizer/french_symbols.txt

@@ -0,0 +1 @@
+_,.!?-~…NQabdefghijklmnopstuvwxyzɑæʃʑçɯɪɔɛɹðəɫɥɸʊɾʒθβŋɦ⁼ʰ`^#*=ˈˌ→↓↑ ɣɡrɲʝʎː̃œøʁɒʌ—ɜɐ

tcp_server/src/lib/tts/text/fr_phonemizer/gruut_wrapper.py

@@ -0,0 +1,258 @@
+import importlib
+from typing import List
+
+import gruut
+from gruut_ipa import IPA # pip install gruut_ipa
+
+from .base import BasePhonemizer
+from .punctuation import Punctuation
+
+# Table for str.translate to fix gruut/TTS phoneme mismatch
+GRUUT_TRANS_TABLE = str.maketrans("g", "ɡ")
+
+
+class Gruut(BasePhonemizer):
+    """Gruut wrapper for G2P
+
+    Args:
+        language (str):
+            Valid language code for the used backend.
+
+        punctuations (str):
+            Characters to be treated as punctuation. Defaults to `Punctuation.default_puncs()`.
+
+        keep_puncs (bool):
+            If true, keep the punctuations after phonemization. Defaults to True.
+
+        use_espeak_phonemes (bool):
+            If true, use espeak lexicons instead of default Gruut lexicons. Defaults to False.
+
+        keep_stress (bool):
+            If true, keep the stress characters after phonemization. Defaults to False.
+
+    Example:
+
+        >>> from TTS.tts.utils.text.phonemizers.gruut_wrapper import Gruut
+        >>> phonemizer = Gruut('en-us')
+        >>> phonemizer.phonemize("Be a voice, not an! echo?", separator="|")
+        'b|i| ə| v|ɔ|ɪ|s, n|ɑ|t| ə|n! ɛ|k|o|ʊ?'
+    """
+
+    def __init__(
+        self,
+        language: str,
+        punctuations=Punctuation.default_puncs(),
+        keep_puncs=True,
+        use_espeak_phonemes=False,
+        keep_stress=False,
+    ):
+        super().__init__(language, punctuations=punctuations, keep_puncs=keep_puncs)
+        self.use_espeak_phonemes = use_espeak_phonemes
+        self.keep_stress = keep_stress
+
+    @staticmethod
+    def name():
+        return "gruut"
+
+    def phonemize_gruut(self, text: str, separator: str = "|", tie=False) -> str:  # pylint: disable=unused-argument
+        """Convert input text to phonemes.
+
+        Gruut phonemizes the given `str` by seperating each phoneme character with `separator`, even for characters
+        that constitude a single sound.
+
+        It doesn't affect 🐸TTS since it individually converts each character to token IDs.
+
+        Examples::
+            "hello how are you today?" -> `h|ɛ|l|o|ʊ| h|a|ʊ| ɑ|ɹ| j|u| t|ə|d|e|ɪ`
+
+        Args:
+            text (str):
+                Text to be converted to phonemes.
+
+            tie (bool, optional) : When True use a '͡' character between
+                consecutive characters of a single phoneme. Else separate phoneme
+                with '_'. This option requires espeak>=1.49. Default to False.
+        """
+        ph_list = []
+        for sentence in gruut.sentences(text, lang=self.language, espeak=self.use_espeak_phonemes):
+            for word in sentence:
+                if word.is_break:
+                    # Use actual character for break phoneme (e.g., comma)
+                    if ph_list:
+                        # Join with previous word
+                        ph_list[-1].append(word.text)
+                    else:
+                        # First word is punctuation
+                        ph_list.append([word.text])
+                elif word.phonemes:
+                    # Add phonemes for word
+                    word_phonemes = []
+
+                    for word_phoneme in word.phonemes:
+                        if not self.keep_stress:
+                            # Remove primary/secondary stress
+                            word_phoneme = IPA.without_stress(word_phoneme)
+
+                        word_phoneme = word_phoneme.translate(GRUUT_TRANS_TABLE)
+
+                        if word_phoneme:
+                            # Flatten phonemes
+                            word_phonemes.extend(word_phoneme)
+
+                    if word_phonemes:
+                        ph_list.append(word_phonemes)
+
+        ph_words = [separator.join(word_phonemes) for word_phonemes in ph_list]
+        ph = f"{separator} ".join(ph_words)
+        return ph
+
+    def _phonemize(self, text, separator):
+        return self.phonemize_gruut(text, separator, tie=False)
+
+    def is_supported_language(self, language):
+        """Returns True if `language` is supported by the backend"""
+        return gruut.is_language_supported(language)
+
+    @staticmethod
+    def supported_languages() -> List:
+        """Get a dictionary of supported languages.
+
+        Returns:
+            List: List of language codes.
+        """
+        return list(gruut.get_supported_languages())
+
+    def version(self):
+        """Get the version of the used backend.
+
+        Returns:
+            str: Version of the used backend.
+        """
+        return gruut.__version__
+
+    @classmethod
+    def is_available(cls):
+        """Return true if ESpeak is available else false"""
+        return importlib.util.find_spec("gruut") is not None
+
+
+if __name__ == "__main__":
+    from cleaner import french_cleaners
+    import json
+
+    e = Gruut(language="fr-fr", keep_puncs=True, keep_stress=True, use_espeak_phonemes=True)
+    symbols = [  # en + sp
+        "_",
+        ",",
+        ".",
+        "!",
+        "?",
+        "-",
+        "~",
+        "\u2026",
+        "N",
+        "Q",
+        "a",
+        "b",
+        "d",
+        "e",
+        "f",
+        "g",
+        "h",
+        "i",
+        "j",
+        "k",
+        "l",
+        "m",
+        "n",
+        "o",
+        "p",
+        "s",
+        "t",
+        "u",
+        "v",
+        "w",
+        "x",
+        "y",
+        "z",
+        "\u0251",
+        "\u00e6",
+        "\u0283",
+        "\u0291",
+        "\u00e7",
+        "\u026f",
+        "\u026a",
+        "\u0254",
+        "\u025b",
+        "\u0279",
+        "\u00f0",
+        "\u0259",
+        "\u026b",
+        "\u0265",
+        "\u0278",
+        "\u028a",
+        "\u027e",
+        "\u0292",
+        "\u03b8",
+        "\u03b2",
+        "\u014b",
+        "\u0266",
+        "\u207c",
+        "\u02b0",
+        "`",
+        "^",
+        "#",
+        "*",
+        "=",
+        "\u02c8",
+        "\u02cc",
+        "\u2192",
+        "\u2193",
+        "\u2191",
+        " ",
+        "ɣ",
+        "ɡ", 
+        "r", 
+        "ɲ", 
+        "ʝ", 
+        "ʎ",
+        "ː"
+    ]
+    with open('/home/xumin/workspace/VITS-Training-Multiling/230715_fr/metadata.txt', 'r') as f:
+        lines = f.readlines()
+    
+
+    used_sym = []
+    not_existed_sym = []
+    phonemes = []
+
+    for line in lines:
+        text = line.split('|')[-1].strip()
+        text = french_cleaners(text)
+        ipa =  e.phonemize(text, separator="")
+        phonemes.append(ipa)
+        for s in ipa:
+            if s not in symbols:
+                if s not in not_existed_sym:
+                    print(f'not_existed char: {s}')
+                    not_existed_sym.append(s)
+            else:
+                if s not in used_sym:
+                    # print(f'used char: {s}')
+                    used_sym.append(s)
+    
+    print(used_sym)
+    print(not_existed_sym)
+
+
+    with open('./text/fr_phonemizer/french_symbols.txt', 'w') as g:
+        g.writelines(symbols + not_existed_sym)
+        
+    with open('./text/fr_phonemizer/example_ipa.txt', 'w') as g:
+        g.writelines(phonemes)
+
+    data = {'symbols': symbols + not_existed_sym}
+
+    with open('./text/fr_phonemizer/fr_symbols.json', 'w') as f:
+        json.dump(data, f, indent=4)
+

tcp_server/src/lib/tts/text/fr_phonemizer/punctuation.py

@@ -0,0 +1,172 @@
+import collections
+import re
+from enum import Enum
+
+import six
+
+_DEF_PUNCS = ';:,.!?¡¿—…"«»“”'
+
+_PUNC_IDX = collections.namedtuple("_punc_index", ["punc", "position"])
+
+
+class PuncPosition(Enum):
+    """Enum for the punctuations positions"""
+
+    BEGIN = 0
+    END = 1
+    MIDDLE = 2
+    ALONE = 3
+
+
+class Punctuation:
+    """Handle punctuations in text.
+
+    Just strip punctuations from text or strip and restore them later.
+
+    Args:
+        puncs (str): The punctuations to be processed. Defaults to `_DEF_PUNCS`.
+
+    Example:
+        >>> punc = Punctuation()
+        >>> punc.strip("This is. example !")
+        'This is example'
+
+        >>> text_striped, punc_map = punc.strip_to_restore("This is. example !")
+        >>> ' '.join(text_striped)
+        'This is example'
+
+        >>> text_restored = punc.restore(text_striped, punc_map)
+        >>> text_restored[0]
+        'This is. example !'
+    """
+
+    def __init__(self, puncs: str = _DEF_PUNCS):
+        self.puncs = puncs
+
+    @staticmethod
+    def default_puncs():
+        """Return default set of punctuations."""
+        return _DEF_PUNCS
+
+    @property
+    def puncs(self):
+        return self._puncs
+
+    @puncs.setter
+    def puncs(self, value):
+        if not isinstance(value, six.string_types):
+            raise ValueError("[!] Punctuations must be of type str.")
+        self._puncs = "".join(list(dict.fromkeys(list(value))))  # remove duplicates without changing the oreder
+        self.puncs_regular_exp = re.compile(rf"(\s*[{re.escape(self._puncs)}]+\s*)+")
+
+    def strip(self, text):
+        """Remove all the punctuations by replacing with `space`.
+
+        Args:
+            text (str): The text to be processed.
+
+        Example::
+
+            "This is. example !" -> "This is example "
+        """
+        return re.sub(self.puncs_regular_exp, " ", text).rstrip().lstrip()
+
+    def strip_to_restore(self, text):
+        """Remove punctuations from text to restore them later.
+
+        Args:
+            text (str): The text to be processed.
+
+        Examples ::
+
+            "This is. example !" -> [["This is", "example"], [".", "!"]]
+
+        """
+        text, puncs = self._strip_to_restore(text)
+        return text, puncs
+
+    def _strip_to_restore(self, text):
+        """Auxiliary method for Punctuation.preserve()"""
+        matches = list(re.finditer(self.puncs_regular_exp, text))
+        if not matches:
+            return [text], []
+        # the text is only punctuations
+        if len(matches) == 1 and matches[0].group() == text:
+            return [], [_PUNC_IDX(text, PuncPosition.ALONE)]
+        # build a punctuation map to be used later to restore punctuations
+        puncs = []
+        for match in matches:
+            position = PuncPosition.MIDDLE
+            if match == matches[0] and text.startswith(match.group()):
+                position = PuncPosition.BEGIN
+            elif match == matches[-1] and text.endswith(match.group()):
+                position = PuncPosition.END
+            puncs.append(_PUNC_IDX(match.group(), position))
+        # convert str text to a List[str], each item is separated by a punctuation
+        splitted_text = []
+        for idx, punc in enumerate(puncs):
+            split = text.split(punc.punc)
+            prefix, suffix = split[0], punc.punc.join(split[1:])
+            splitted_text.append(prefix)
+            # if the text does not end with a punctuation, add it to the last item
+            if idx == len(puncs) - 1 and len(suffix) > 0:
+                splitted_text.append(suffix)
+            text = suffix
+        return splitted_text, puncs
+
+    @classmethod
+    def restore(cls, text, puncs):
+        """Restore punctuation in a text.
+
+        Args:
+            text (str): The text to be processed.
+            puncs (List[str]): The list of punctuations map to be used for restoring.
+
+        Examples ::
+
+            ['This is', 'example'], ['.', '!'] -> "This is. example!"
+
+        """
+        return cls._restore(text, puncs, 0)
+
+    @classmethod
+    def _restore(cls, text, puncs, num):  # pylint: disable=too-many-return-statements
+        """Auxiliary method for Punctuation.restore()"""
+        if not puncs:
+            return text
+
+        # nothing have been phonemized, returns the puncs alone
+        if not text:
+            return ["".join(m.punc for m in puncs)]
+
+        current = puncs[0]
+
+        if current.position == PuncPosition.BEGIN:
+            return cls._restore([current.punc + text[0]] + text[1:], puncs[1:], num)
+
+        if current.position == PuncPosition.END:
+            return [text[0] + current.punc] + cls._restore(text[1:], puncs[1:], num + 1)
+
+        if current.position == PuncPosition.ALONE:
+            return [current.mark] + cls._restore(text, puncs[1:], num + 1)
+
+        # POSITION == MIDDLE
+        if len(text) == 1:  # pragma: nocover
+            # a corner case where the final part of an intermediate
+            # mark (I) has not been phonemized
+            return cls._restore([text[0] + current.punc], puncs[1:], num)
+
+        return cls._restore([text[0] + current.punc + text[1]] + text[2:], puncs[1:], num)
+
+
+# if __name__ == "__main__":
+#     punc = Punctuation()
+#     text = "This is. This is, example!"
+
+#     print(punc.strip(text))
+
+#     split_text, puncs = punc.strip_to_restore(text)
+#     print(split_text, " ---- ", puncs)
+
+#     restored_text = punc.restore(split_text, puncs)
+#     print(restored_text)

tcp_server/src/lib/tts/text/french.py

@@ -0,0 +1,95 @@
+from transformers import AutoTokenizer
+
+from lib.constants import TTS_BERT_FRENCH_MODEL_DIR_PATH
+
+from .fr_phonemizer import cleaner as fr_cleaner
+from .fr_phonemizer import fr_to_ipa
+
+
+def distribute_phone(n_phone, n_word):
+    phones_per_word = [0] * n_word
+    for task in range(n_phone):
+        min_tasks = min(phones_per_word)
+        min_index = phones_per_word.index(min_tasks)
+        phones_per_word[min_index] += 1
+    return phones_per_word
+
+def text_normalize(text):
+    text = fr_cleaner.french_cleaners(text)
+    return text
+
+load_model_params = {
+    "pretrained_model_name_or_path": 'dbmdz/bert-base-french-europeana-cased',
+    "local_files_only": True
+}
+tokenizer = AutoTokenizer.from_pretrained(**load_model_params)
+
+def g2p(text, pad_start_end=True, tokenized=None):
+    if tokenized is None:
+        tokenized = tokenizer.tokenize(text)
+    # import pdb; pdb.set_trace()
+    phs = []
+    ph_groups = []
+    for t in tokenized:
+        if not t.startswith("#"):
+            ph_groups.append([t])
+        else:
+            ph_groups[-1].append(t.replace("#", ""))
+    
+    phones = []
+    tones = []
+    word2ph = []
+    # print(ph_groups)
+    for group in ph_groups:
+        w = "".join(group)
+        phone_len = 0
+        word_len = len(group)
+        if w == '[UNK]':
+            phone_list = ['UNK']
+        else:
+            phone_list = list(filter(lambda p: p != " ", fr_to_ipa.fr2ipa(w)))
+        
+        for ph in phone_list:
+            phones.append(ph)
+            tones.append(0)
+            phone_len += 1
+        aaa = distribute_phone(phone_len, word_len)
+        word2ph += aaa
+        # print(phone_list, aaa)
+        # print('=' * 10)
+
+    if pad_start_end:
+        phones = ["_"] + phones + ["_"]
+        tones = [0] + tones + [0]
+        word2ph = [1] + word2ph + [1]
+    return phones, tones, word2ph
+
+def get_bert_feature(text, word2ph, device=None):
+    from text import french_bert
+    return french_bert.get_bert_feature(text, word2ph, device=device)
+
+if __name__ == "__main__":
+    ori_text = 'Ce service gratuit est“”"" 【disponible》 en chinois 【simplifié] et autres 123'
+    # ori_text = "Ils essayaient vainement de faire comprendre à ma mère qu'avec les cent mille francs que m'avait laissé mon père,"
+    # print(ori_text)
+    text = text_normalize(ori_text)
+    print(text)
+    phoneme = fr_to_ipa.fr2ipa(text)
+    print(phoneme)
+
+    
+    from TTS.tts.utils.text.phonemizers.multi_phonemizer import MultiPhonemizer
+    from text.cleaner_multiling import unicleaners
+
+    def text_normalize(text):
+        text = unicleaners(text, cased=True, lang='fr')
+        return text
+
+    # print(ori_text)
+    text = text_normalize(ori_text)
+    print(text)
+    phonemizer = MultiPhonemizer({"fr-fr": "espeak"})
+    # phonemizer.lang_to_phonemizer['fr'].keep_stress = True
+    # phonemizer.lang_to_phonemizer['fr'].use_espeak_phonemes = True
+    phoneme = phonemizer.phonemize(text, separator="", language='fr-fr')
+    print(phoneme)

tcp_server/src/lib/tts/text/french_bert.py

@@ -0,0 +1,44 @@
+import torch
+from transformers import AutoTokenizer, AutoModelForMaskedLM
+import sys
+
+from lib.constants import TTS_BERT_FRENCH_MODEL_DIR_PATH
+
+load_model_params = {
+    "pretrained_model_name_or_path": TTS_BERT_FRENCH_MODEL_DIR_PATH,
+    "local_files_only": True
+}
+tokenizer = AutoTokenizer.from_pretrained(**load_model_params)
+model = None
+
+def get_bert_feature(text, word2ph, device=None):
+    global model
+    if (
+        sys.platform == "darwin"
+        and torch.backends.mps.is_available()
+        and device == "cpu"
+    ):
+        device = "mps"
+    if not device:
+        device = "cuda"
+    if model is None:
+        model = AutoModelForMaskedLM.from_pretrained(**load_model_params).to(
+            device
+        )
+    with torch.no_grad():
+        inputs = tokenizer(text, return_tensors="pt")
+        for i in inputs:
+            inputs[i] = inputs[i].to(device)
+        res = model(**inputs, output_hidden_states=True)
+        res = torch.cat(res["hidden_states"][-3:-2], -1)[0].cpu()
+        
+    assert inputs["input_ids"].shape[-1] == len(word2ph)
+    word2phone = word2ph
+    phone_level_feature = []
+    for i in range(len(word2phone)):
+        repeat_feature = res[i].repeat(word2phone[i], 1)
+        phone_level_feature.append(repeat_feature)
+
+    phone_level_feature = torch.cat(phone_level_feature, dim=0)
+
+    return phone_level_feature.T

tcp_server/src/lib/tts/text/opencpop-strict.txt

@@ -0,0 +1,429 @@
+a	AA a
+ai	AA ai
+an	AA an
+ang	AA ang
+ao	AA ao
+ba	b a
+bai	b ai
+ban	b an
+bang	b ang
+bao	b ao
+bei	b ei
+ben	b en
+beng	b eng
+bi	b i
+bian	b ian
+biao	b iao
+bie	b ie
+bin	b in
+bing	b ing
+bo	b o
+bu	b u
+ca	c a
+cai	c ai
+can	c an
+cang	c ang
+cao	c ao
+ce	c e
+cei	c ei
+cen	c en
+ceng	c eng
+cha	ch a
+chai	ch ai
+chan	ch an
+chang	ch ang
+chao	ch ao
+che	ch e
+chen	ch en
+cheng	ch eng
+chi	ch ir
+chong	ch ong
+chou	ch ou
+chu	ch u
+chua	ch ua
+chuai	ch uai
+chuan	ch uan
+chuang	ch uang
+chui	ch ui
+chun	ch un
+chuo	ch uo
+ci	c i0
+cong	c ong
+cou	c ou
+cu	c u
+cuan	c uan
+cui	c ui
+cun	c un
+cuo	c uo
+da	d a
+dai	d ai
+dan	d an
+dang	d ang
+dao	d ao
+de	d e
+dei	d ei
+den	d en
+deng	d eng
+di	d i
+dia	d ia
+dian	d ian
+diao	d iao
+die	d ie
+ding	d ing
+diu	d iu
+dong	d ong
+dou	d ou
+du	d u
+duan	d uan
+dui	d ui
+dun	d un
+duo	d uo
+e	EE e
+ei	EE ei
+en	EE en
+eng	EE eng
+er	EE er
+fa	f a
+fan	f an
+fang	f ang
+fei	f ei
+fen	f en
+feng	f eng
+fo	f o
+fou	f ou
+fu	f u
+ga	g a
+gai	g ai
+gan	g an
+gang	g ang
+gao	g ao
+ge	g e
+gei	g ei
+gen	g en
+geng	g eng
+gong	g ong
+gou	g ou
+gu	g u
+gua	g ua
+guai	g uai
+guan	g uan
+guang	g uang
+gui	g ui
+gun	g un
+guo	g uo
+ha	h a
+hai	h ai
+han	h an
+hang	h ang
+hao	h ao
+he	h e
+hei	h ei
+hen	h en
+heng	h eng
+hong	h ong
+hou	h ou
+hu	h u
+hua	h ua
+huai	h uai
+huan	h uan
+huang	h uang
+hui	h ui
+hun	h un
+huo	h uo
+ji	j i
+jia	j ia
+jian	j ian
+jiang	j iang
+jiao	j iao
+jie	j ie
+jin	j in
+jing	j ing
+jiong	j iong
+jiu	j iu
+ju	j v
+jv	j v
+juan	j van
+jvan	j van
+jue	j ve
+jve	j ve
+jun	j vn
+jvn	j vn
+ka	k a
+kai	k ai
+kan	k an
+kang	k ang
+kao	k ao
+ke	k e
+kei	k ei
+ken	k en
+keng	k eng
+kong	k ong
+kou	k ou
+ku	k u
+kua	k ua
+kuai	k uai
+kuan	k uan
+kuang	k uang
+kui	k ui
+kun	k un
+kuo	k uo
+la	l a
+lai	l ai
+lan	l an
+lang	l ang
+lao	l ao
+le	l e
+lei	l ei
+leng	l eng
+li	l i
+lia	l ia
+lian	l ian
+liang	l iang
+liao	l iao
+lie	l ie
+lin	l in
+ling	l ing
+liu	l iu
+lo	l o
+long	l ong
+lou	l ou
+lu	l u
+luan	l uan
+lun	l un
+luo	l uo
+lv	l v
+lve	l ve
+ma	m a
+mai	m ai
+man	m an
+mang	m ang
+mao	m ao
+me	m e
+mei	m ei
+men	m en
+meng	m eng
+mi	m i
+mian	m ian
+miao	m iao
+mie	m ie
+min	m in
+ming	m ing
+miu	m iu
+mo	m o
+mou	m ou
+mu	m u
+na	n a
+nai	n ai
+nan	n an
+nang	n ang
+nao	n ao
+ne	n e
+nei	n ei
+nen	n en
+neng	n eng
+ni	n i
+nian	n ian
+niang	n iang
+niao	n iao
+nie	n ie
+nin	n in
+ning	n ing
+niu	n iu
+nong	n ong
+nou	n ou
+nu	n u
+nuan	n uan
+nun	n un
+nuo	n uo
+nv	n v
+nve	n ve
+o	OO o
+ou	OO ou
+pa	p a
+pai	p ai
+pan	p an
+pang	p ang
+pao	p ao
+pei	p ei
+pen	p en
+peng	p eng
+pi	p i
+pian	p ian
+piao	p iao
+pie	p ie
+pin	p in
+ping	p ing
+po	p o
+pou	p ou
+pu	p u
+qi	q i
+qia	q ia
+qian	q ian
+qiang	q iang
+qiao	q iao
+qie	q ie
+qin	q in
+qing	q ing
+qiong	q iong
+qiu	q iu
+qu	q v
+qv	q v
+quan	q van
+qvan	q van
+que	q ve
+qve	q ve
+qun	q vn
+qvn	q vn
+ran	r an
+rang	r ang
+rao	r ao
+re	r e
+ren	r en
+reng	r eng
+ri	r ir
+rong	r ong
+rou	r ou
+ru	r u
+rua	r ua
+ruan	r uan
+rui	r ui
+run	r un
+ruo	r uo
+sa	s a
+sai	s ai
+san	s an
+sang	s ang
+sao	s ao
+se	s e
+sen	s en
+seng	s eng
+sha	sh a
+shai	sh ai
+shan	sh an
+shang	sh ang
+shao	sh ao
+she	sh e
+shei	sh ei
+shen	sh en
+sheng	sh eng
+shi	sh ir
+shou	sh ou
+shu	sh u
+shua	sh ua
+shuai	sh uai
+shuan	sh uan
+shuang	sh uang
+shui	sh ui
+shun	sh un
+shuo	sh uo
+si	s i0
+song	s ong
+sou	s ou
+su	s u
+suan	s uan
+sui	s ui
+sun	s un
+suo	s uo
+ta	t a
+tai	t ai
+tan	t an
+tang	t ang
+tao	t ao
+te	t e
+tei	t ei
+teng	t eng
+ti	t i
+tian	t ian
+tiao	t iao
+tie	t ie
+ting	t ing
+tong	t ong
+tou	t ou
+tu	t u
+tuan	t uan
+tui	t ui
+tun	t un
+tuo	t uo
+wa	w a
+wai	w ai
+wan	w an
+wang	w ang
+wei	w ei
+wen	w en
+weng	w eng
+wo	w o
+wu	w u
+xi	x i
+xia	x ia
+xian	x ian
+xiang	x iang
+xiao	x iao
+xie	x ie
+xin	x in
+xing	x ing
+xiong	x iong
+xiu	x iu
+xu	x v
+xv	x v
+xuan	x van
+xvan	x van
+xue	x ve
+xve	x ve
+xun	x vn
+xvn	x vn
+ya	y a
+yan	y En
+yang	y ang
+yao	y ao
+ye	y E
+yi	y i
+yin	y in
+ying	y ing
+yo	y o
+yong	y ong
+you	y ou
+yu	y v
+yv	y v
+yuan	y van
+yvan	y van
+yue	y ve
+yve	y ve
+yun	y vn
+yvn	y vn
+za	z a
+zai	z ai
+zan	z an
+zang	z ang
+zao	z ao
+ze	z e
+zei	z ei
+zen	z en
+zeng	z eng
+zha	zh a
+zhai	zh ai
+zhan	zh an
+zhang	zh ang
+zhao	zh ao
+zhe	zh e
+zhei	zh ei
+zhen	zh en
+zheng	zh eng
+zhi	zh ir
+zhong	zh ong
+zhou	zh ou
+zhu	zh u
+zhua	zh ua
+zhuai	zh uai
+zhuan	zh uan
+zhuang	zh uang
+zhui	zh ui
+zhun	zh un
+zhuo	zh uo
+zi	z i0
+zong	z ong
+zou	z ou
+zu	z u
+zuan	z uan
+zui	z ui
+zun	z un
+zuo	z uo

tcp_server/src/lib/tts/text/symbols.py

@@ -0,0 +1,290 @@
+# punctuation = ["!", "?", "…", ",", ".", "'", "-"]
+punctuation = ["!", "?", "…", ",", ".", "'", "-", "¿", "¡"]
+pu_symbols = punctuation + ["SP", "UNK"]
+pad = "_"
+
+# chinese
+zh_symbols = [
+    "E",
+    "En",
+    "a",
+    "ai",
+    "an",
+    "ang",
+    "ao",
+    "b",
+    "c",
+    "ch",
+    "d",
+    "e",
+    "ei",
+    "en",
+    "eng",
+    "er",
+    "f",
+    "g",
+    "h",
+    "i",
+    "i0",
+    "ia",
+    "ian",
+    "iang",
+    "iao",
+    "ie",
+    "in",
+    "ing",
+    "iong",
+    "ir",
+    "iu",
+    "j",
+    "k",
+    "l",
+    "m",
+    "n",
+    "o",
+    "ong",
+    "ou",
+    "p",
+    "q",
+    "r",
+    "s",
+    "sh",
+    "t",
+    "u",
+    "ua",
+    "uai",
+    "uan",
+    "uang",
+    "ui",
+    "un",
+    "uo",
+    "v",
+    "van",
+    "ve",
+    "vn",
+    "w",
+    "x",
+    "y",
+    "z",
+    "zh",
+    "AA",
+    "EE",
+    "OO",
+]
+num_zh_tones = 6
+
+# japanese
+ja_symbols = [
+    "N",
+    "a",
+    "a:",
+    "b",
+    "by",
+    "ch",
+    "d",
+    "dy",
+    "e",
+    "e:",
+    "f",
+    "g",
+    "gy",
+    "h",
+    "hy",
+    "i",
+    "i:",
+    "j",
+    "k",
+    "ky",
+    "m",
+    "my",
+    "n",
+    "ny",
+    "o",
+    "o:",
+    "p",
+    "py",
+    "q",
+    "r",
+    "ry",
+    "s",
+    "sh",
+    "t",
+    "ts",
+    "ty",
+    "u",
+    "u:",
+    "w",
+    "y",
+    "z",
+    "zy",
+]
+num_ja_tones = 1
+
+# English
+en_symbols = [
+    "aa",
+    "ae",
+    "ah",
+    "ao",
+    "aw",
+    "ay",
+    "b",
+    "ch",
+    "d",
+    "dh",
+    "eh",
+    "er",
+    "ey",
+    "f",
+    "g",
+    "hh",
+    "ih",
+    "iy",
+    "jh",
+    "k",
+    "l",
+    "m",
+    "n",
+    "ng",
+    "ow",
+    "oy",
+    "p",
+    "r",
+    "s",
+    "sh",
+    "t",
+    "th",
+    "uh",
+    "uw",
+    "V",
+    "w",
+    "y",
+    "z",
+    "zh",
+]
+num_en_tones = 4
+
+# Korean
+kr_symbols = ['ᄌ', 'ᅥ', 'ᆫ', 'ᅦ', 'ᄋ', 'ᅵ', 'ᄅ', 'ᅴ', 'ᄀ', 'ᅡ', 'ᄎ', 'ᅪ', 'ᄑ', 'ᅩ', 'ᄐ', 'ᄃ', 'ᅢ', 'ᅮ', 'ᆼ', 'ᅳ', 'ᄒ', 'ᄆ', 'ᆯ', 'ᆷ', 'ᄂ', 'ᄇ', 'ᄉ', 'ᆮ', 'ᄁ', 'ᅬ', 'ᅣ', 'ᄄ', 'ᆨ', 'ᄍ', 'ᅧ', 'ᄏ', 'ᆸ', 'ᅭ', '(', 'ᄊ', ')', 'ᅲ', 'ᅨ', 'ᄈ', 'ᅱ', 'ᅯ', 'ᅫ', 'ᅰ', 'ᅤ', '~', '\\', '[', ']', '/', '^', ':', 'ㄸ', '*']
+num_kr_tones = 1
+
+# Spanish
+es_symbols = [
+        "N",
+        "Q",
+        "a",
+        "b",
+        "d",
+        "e",
+        "f",
+        "g",
+        "h",
+        "i",
+        "j",
+        "k",
+        "l",
+        "m",
+        "n",
+        "o",
+        "p",
+        "s",
+        "t",
+        "u",
+        "v",
+        "w",
+        "x",
+        "y",
+        "z",
+        "ɑ",
+        "æ",
+        "ʃ",
+        "ʑ",
+        "ç",
+        "ɯ",
+        "ɪ",
+        "ɔ",
+        "ɛ",
+        "ɹ",
+        "ð",
+        "ə",
+        "ɫ",
+        "ɥ",
+        "ɸ",
+        "ʊ",
+        "ɾ",
+        "ʒ",
+        "θ",
+        "β",
+        "ŋ",
+        "ɦ",
+        "ɡ",
+        "r",
+        "ɲ",
+        "ʝ",
+        "ɣ",
+        "ʎ",
+        "ˈ",
+        "ˌ",
+        "ː"
+    ]
+num_es_tones = 1
+
+# French 
+fr_symbols = [
+    "\u0303",
+    "œ",
+    "ø",
+    "ʁ",
+    "ɒ",
+    "ʌ",
+    "ɜ",
+    "ɐ"
+]
+num_fr_tones = 1
+
+# German 
+de_symbols = [
+    "ʏ",
+    "̩"
+  ]
+num_de_tones = 1
+
+# Russian 
+ru_symbols = [
+    "ɭ",
+    "ʲ",
+    "ɕ",
+    "\"",
+    "ɵ",
+    "^",
+    "ɬ"
+]
+num_ru_tones = 1
+
+# combine all symbols
+normal_symbols = sorted(set(zh_symbols + ja_symbols + en_symbols + kr_symbols + es_symbols + fr_symbols + de_symbols + ru_symbols))
+symbols = [pad] + normal_symbols + pu_symbols
+sil_phonemes_ids = [symbols.index(i) for i in pu_symbols]
+
+# combine all tones
+num_tones = num_zh_tones + num_ja_tones + num_en_tones + num_kr_tones + num_es_tones + num_fr_tones + num_de_tones + num_ru_tones
+
+# language maps
+language_id_map = {"ZH": 0, "JP": 1, "EN": 2, "ZH_MIX_EN": 3, 'KR': 4, 'ES': 5, 'SP': 5 ,'FR': 6}
+num_languages = len(language_id_map.keys())
+
+language_tone_start_map = {
+    "ZH": 0,
+    "ZH_MIX_EN": 0,
+    "JP": num_zh_tones,
+    "EN": num_zh_tones + num_ja_tones,
+    'KR': num_zh_tones + num_ja_tones + num_en_tones,
+    "ES": num_zh_tones + num_ja_tones + num_en_tones + num_kr_tones,
+    "SP": num_zh_tones + num_ja_tones + num_en_tones + num_kr_tones,
+    "FR": num_zh_tones + num_ja_tones + num_en_tones + num_kr_tones + num_es_tones,
+}
+
+if __name__ == "__main__":
+    a = set(zh_symbols)
+    b = set(en_symbols)
+    print(sorted(a & b))

tcp_server/src/lib/tts/transforms.py

@@ -0,0 +1,209 @@
+import torch
+from torch.nn import functional as F
+
+import numpy as np
+
+
+DEFAULT_MIN_BIN_WIDTH = 1e-3
+DEFAULT_MIN_BIN_HEIGHT = 1e-3
+DEFAULT_MIN_DERIVATIVE = 1e-3
+
+
+def piecewise_rational_quadratic_transform(
+    inputs,
+    unnormalized_widths,
+    unnormalized_heights,
+    unnormalized_derivatives,
+    inverse=False,
+    tails=None,
+    tail_bound=1.0,
+    min_bin_width=DEFAULT_MIN_BIN_WIDTH,
+    min_bin_height=DEFAULT_MIN_BIN_HEIGHT,
+    min_derivative=DEFAULT_MIN_DERIVATIVE,
+):
+    if tails is None:
+        spline_fn = rational_quadratic_spline
+        spline_kwargs = {}
+    else:
+        spline_fn = unconstrained_rational_quadratic_spline
+        spline_kwargs = {"tails": tails, "tail_bound": tail_bound}
+
+    outputs, logabsdet = spline_fn(
+        inputs=inputs,
+        unnormalized_widths=unnormalized_widths,
+        unnormalized_heights=unnormalized_heights,
+        unnormalized_derivatives=unnormalized_derivatives,
+        inverse=inverse,
+        min_bin_width=min_bin_width,
+        min_bin_height=min_bin_height,
+        min_derivative=min_derivative,
+        **spline_kwargs
+    )
+    return outputs, logabsdet
+
+
+def searchsorted(bin_locations, inputs, eps=1e-6):
+    bin_locations[..., -1] += eps
+    return torch.sum(inputs[..., None] >= bin_locations, dim=-1) - 1
+
+
+def unconstrained_rational_quadratic_spline(
+    inputs,
+    unnormalized_widths,
+    unnormalized_heights,
+    unnormalized_derivatives,
+    inverse=False,
+    tails="linear",
+    tail_bound=1.0,
+    min_bin_width=DEFAULT_MIN_BIN_WIDTH,
+    min_bin_height=DEFAULT_MIN_BIN_HEIGHT,
+    min_derivative=DEFAULT_MIN_DERIVATIVE,
+):
+    inside_interval_mask = (inputs >= -tail_bound) & (inputs <= tail_bound)
+    outside_interval_mask = ~inside_interval_mask
+
+    outputs = torch.zeros_like(inputs)
+    logabsdet = torch.zeros_like(inputs)
+
+    if tails == "linear":
+        unnormalized_derivatives = F.pad(unnormalized_derivatives, pad=(1, 1))
+        constant = np.log(np.exp(1 - min_derivative) - 1)
+        unnormalized_derivatives[..., 0] = constant
+        unnormalized_derivatives[..., -1] = constant
+
+        outputs[outside_interval_mask] = inputs[outside_interval_mask]
+        logabsdet[outside_interval_mask] = 0
+    else:
+        raise RuntimeError("{} tails are not implemented.".format(tails))
+
+    (
+        outputs[inside_interval_mask],
+        logabsdet[inside_interval_mask],
+    ) = rational_quadratic_spline(
+        inputs=inputs[inside_interval_mask],
+        unnormalized_widths=unnormalized_widths[inside_interval_mask, :],
+        unnormalized_heights=unnormalized_heights[inside_interval_mask, :],
+        unnormalized_derivatives=unnormalized_derivatives[inside_interval_mask, :],
+        inverse=inverse,
+        left=-tail_bound,
+        right=tail_bound,
+        bottom=-tail_bound,
+        top=tail_bound,
+        min_bin_width=min_bin_width,
+        min_bin_height=min_bin_height,
+        min_derivative=min_derivative,
+    )
+
+    return outputs, logabsdet
+
+
+def rational_quadratic_spline(
+    inputs,
+    unnormalized_widths,
+    unnormalized_heights,
+    unnormalized_derivatives,
+    inverse=False,
+    left=0.0,
+    right=1.0,
+    bottom=0.0,
+    top=1.0,
+    min_bin_width=DEFAULT_MIN_BIN_WIDTH,
+    min_bin_height=DEFAULT_MIN_BIN_HEIGHT,
+    min_derivative=DEFAULT_MIN_DERIVATIVE,
+):
+    if torch.min(inputs) < left or torch.max(inputs) > right:
+        raise ValueError("Input to a transform is not within its domain")
+
+    num_bins = unnormalized_widths.shape[-1]
+
+    if min_bin_width * num_bins > 1.0:
+        raise ValueError("Minimal bin width too large for the number of bins")
+    if min_bin_height * num_bins > 1.0:
+        raise ValueError("Minimal bin height too large for the number of bins")
+
+    widths = F.softmax(unnormalized_widths, dim=-1)
+    widths = min_bin_width + (1 - min_bin_width * num_bins) * widths
+    cumwidths = torch.cumsum(widths, dim=-1)
+    cumwidths = F.pad(cumwidths, pad=(1, 0), mode="constant", value=0.0)
+    cumwidths = (right - left) * cumwidths + left
+    cumwidths[..., 0] = left
+    cumwidths[..., -1] = right
+    widths = cumwidths[..., 1:] - cumwidths[..., :-1]
+
+    derivatives = min_derivative + F.softplus(unnormalized_derivatives)
+
+    heights = F.softmax(unnormalized_heights, dim=-1)
+    heights = min_bin_height + (1 - min_bin_height * num_bins) * heights
+    cumheights = torch.cumsum(heights, dim=-1)
+    cumheights = F.pad(cumheights, pad=(1, 0), mode="constant", value=0.0)
+    cumheights = (top - bottom) * cumheights + bottom
+    cumheights[..., 0] = bottom
+    cumheights[..., -1] = top
+    heights = cumheights[..., 1:] - cumheights[..., :-1]
+
+    if inverse:
+        bin_idx = searchsorted(cumheights, inputs)[..., None]
+    else:
+        bin_idx = searchsorted(cumwidths, inputs)[..., None]
+
+    input_cumwidths = cumwidths.gather(-1, bin_idx)[..., 0]
+    input_bin_widths = widths.gather(-1, bin_idx)[..., 0]
+
+    input_cumheights = cumheights.gather(-1, bin_idx)[..., 0]
+    delta = heights / widths
+    input_delta = delta.gather(-1, bin_idx)[..., 0]
+
+    input_derivatives = derivatives.gather(-1, bin_idx)[..., 0]
+    input_derivatives_plus_one = derivatives[..., 1:].gather(-1, bin_idx)[..., 0]
+
+    input_heights = heights.gather(-1, bin_idx)[..., 0]
+
+    if inverse:
+        a = (inputs - input_cumheights) * (
+            input_derivatives + input_derivatives_plus_one - 2 * input_delta
+        ) + input_heights * (input_delta - input_derivatives)
+        b = input_heights * input_derivatives - (inputs - input_cumheights) * (
+            input_derivatives + input_derivatives_plus_one - 2 * input_delta
+        )
+        c = -input_delta * (inputs - input_cumheights)
+
+        discriminant = b.pow(2) - 4 * a * c
+        assert (discriminant >= 0).all()
+
+        root = (2 * c) / (-b - torch.sqrt(discriminant))
+        outputs = root * input_bin_widths + input_cumwidths
+
+        theta_one_minus_theta = root * (1 - root)
+        denominator = input_delta + (
+            (input_derivatives + input_derivatives_plus_one - 2 * input_delta)
+            * theta_one_minus_theta
+        )
+        derivative_numerator = input_delta.pow(2) * (
+            input_derivatives_plus_one * root.pow(2)
+            + 2 * input_delta * theta_one_minus_theta
+            + input_derivatives * (1 - root).pow(2)
+        )
+        logabsdet = torch.log(derivative_numerator) - 2 * torch.log(denominator)
+
+        return outputs, -logabsdet
+    else:
+        theta = (inputs - input_cumwidths) / input_bin_widths
+        theta_one_minus_theta = theta * (1 - theta)
+
+        numerator = input_heights * (
+            input_delta * theta.pow(2) + input_derivatives * theta_one_minus_theta
+        )
+        denominator = input_delta + (
+            (input_derivatives + input_derivatives_plus_one - 2 * input_delta)
+            * theta_one_minus_theta
+        )
+        outputs = input_cumheights + numerator / denominator
+
+        derivative_numerator = input_delta.pow(2) * (
+            input_derivatives_plus_one * theta.pow(2)
+            + 2 * input_delta * theta_one_minus_theta
+            + input_derivatives * (1 - theta).pow(2)
+        )
+        logabsdet = torch.log(derivative_numerator) - 2 * torch.log(denominator)
+
+        return outputs, logabsdet

tcp_server/src/lib/tts/utils.py

@@ -0,0 +1,404 @@
+import os
+import glob
+import argparse
+import logging
+import json
+import subprocess
+import torch
+from lib.tts.text import cleaned_text_to_sequence, get_bert
+from lib.tts.text.cleaner import clean_text
+from lib.tts import commons
+
+MATPLOTLIB_FLAG = False
+
+logger = logging.getLogger(__name__)
+
+
+
+def get_text_for_tts_infer(text, language_str, hps, device, symbol_to_id=None):
+    norm_text, phone, tone, word2ph = clean_text(text, language_str)
+    phone, tone, language = cleaned_text_to_sequence(phone, tone, language_str, symbol_to_id)
+
+    if hps.data.add_blank:
+        phone = commons.intersperse(phone, 0)
+        tone = commons.intersperse(tone, 0)
+        language = commons.intersperse(language, 0)
+        for i in range(len(word2ph)):
+            word2ph[i] = word2ph[i] * 2
+        word2ph[0] += 1
+
+    if getattr(hps.data, "disable_bert", False):
+        bert = torch.zeros(1024, len(phone))
+        ja_bert = torch.zeros(768, len(phone))
+    else:
+        bert = get_bert(norm_text, word2ph, language_str, device)
+        del word2ph
+        assert bert.shape[-1] == len(phone), phone
+
+        if language_str == "ZH":
+            bert = bert
+            ja_bert = torch.zeros(768, len(phone))
+        elif language_str in ["JP", "EN", "ZH_MIX_EN", 'KR', 'SP', 'ES', 'FR', 'DE', 'RU']:
+            ja_bert = bert
+            bert = torch.zeros(1024, len(phone))
+        else:
+            raise NotImplementedError()
+
+    assert bert.shape[-1] == len(
+        phone
+    ), f"Bert seq len {bert.shape[-1]} != {len(phone)}"
+
+    phone = torch.LongTensor(phone)
+    tone = torch.LongTensor(tone)
+    language = torch.LongTensor(language)
+    return bert, ja_bert, phone, tone, language
+
+def load_checkpoint(checkpoint_path, model, optimizer=None, skip_optimizer=False):
+    assert os.path.isfile(checkpoint_path)
+    checkpoint_dict = torch.load(checkpoint_path, map_location="cpu")
+    iteration = checkpoint_dict.get("iteration", 0)
+    learning_rate = checkpoint_dict.get("learning_rate", 0.)
+    if (
+        optimizer is not None
+        and not skip_optimizer
+        and checkpoint_dict["optimizer"] is not None
+    ):
+        optimizer.load_state_dict(checkpoint_dict["optimizer"])
+    elif optimizer is None and not skip_optimizer:
+        # else:      Disable this line if Infer and resume checkpoint,then enable the line upper
+        new_opt_dict = optimizer.state_dict()
+        new_opt_dict_params = new_opt_dict["param_groups"][0]["params"]
+        new_opt_dict["param_groups"] = checkpoint_dict["optimizer"]["param_groups"]
+        new_opt_dict["param_groups"][0]["params"] = new_opt_dict_params
+        optimizer.load_state_dict(new_opt_dict)
+
+    saved_state_dict = checkpoint_dict["model"]
+    if hasattr(model, "module"):
+        state_dict = model.module.state_dict()
+    else:
+        state_dict = model.state_dict()
+
+    new_state_dict = {}
+    for k, v in state_dict.items():
+        try:
+            # assert "emb_g" not in k
+            new_state_dict[k] = saved_state_dict[k]
+            assert saved_state_dict[k].shape == v.shape, (
+                saved_state_dict[k].shape,
+                v.shape,
+            )
+        except Exception as e:
+            print(e)
+            # For upgrading from the old version
+            if "ja_bert_proj" in k:
+                v = torch.zeros_like(v)
+                logger.warn(
+                    f"Seems you are using the old version of the model, the {k} is automatically set to zero for backward compatibility"
+                )
+            else:
+                logger.error(f"{k} is not in the checkpoint")
+
+            new_state_dict[k] = v
+
+    if hasattr(model, "module"):
+        model.module.load_state_dict(new_state_dict, strict=False)
+    else:
+        model.load_state_dict(new_state_dict, strict=False)
+
+    logger.info(
+        "Loaded checkpoint '{}' (iteration {})".format(checkpoint_path, iteration)
+    )
+
+    return model, optimizer, learning_rate, iteration
+
+
+def save_checkpoint(model, optimizer, learning_rate, iteration, checkpoint_path):
+    logger.info(
+        "Saving model and optimizer state at iteration {} to {}".format(
+            iteration, checkpoint_path
+        )
+    )
+    if hasattr(model, "module"):
+        state_dict = model.module.state_dict()
+    else:
+        state_dict = model.state_dict()
+    torch.save(
+        {
+            "model": state_dict,
+            "iteration": iteration,
+            "optimizer": optimizer.state_dict(),
+            "learning_rate": learning_rate,
+        },
+        checkpoint_path,
+    )
+
+
+def summarize(
+    writer,
+    global_step,
+    scalars={},
+    histograms={},
+    images={},
+    audios={},
+    audio_sampling_rate=22050,
+):
+    for k, v in scalars.items():
+        writer.add_scalar(k, v, global_step)
+    for k, v in histograms.items():
+        writer.add_histogram(k, v, global_step)
+    for k, v in images.items():
+        writer.add_image(k, v, global_step, dataformats="HWC")
+    for k, v in audios.items():
+        writer.add_audio(k, v, global_step, audio_sampling_rate)
+
+
+def latest_checkpoint_path(dir_path, regex="G_*.pth"):
+    f_list = glob.glob(os.path.join(dir_path, regex))
+    f_list.sort(key=lambda f: int("".join(filter(str.isdigit, f))))
+    x = f_list[-1]
+    return x
+
+
+def plot_spectrogram_to_numpy(spectrogram):
+    global MATPLOTLIB_FLAG
+    if not MATPLOTLIB_FLAG:
+        import matplotlib
+
+        matplotlib.use("Agg")
+        MATPLOTLIB_FLAG = True
+        mpl_logger = logging.getLogger("matplotlib")
+        mpl_logger.setLevel(logging.WARNING)
+    import matplotlib.pylab as plt
+    import numpy as np
+
+    fig, ax = plt.subplots(figsize=(10, 2))
+    im = ax.imshow(spectrogram, aspect="auto", origin="lower", interpolation="none")
+    plt.colorbar(im, ax=ax)
+    plt.xlabel("Frames")
+    plt.ylabel("Channels")
+    plt.tight_layout()
+
+    fig.canvas.draw()
+    data = np.fromstring(fig.canvas.tostring_rgb(), dtype=np.uint8, sep="")
+    data = data.reshape(fig.canvas.get_width_height()[::-1] + (3,))
+    plt.close()
+    return data
+
+
+def plot_alignment_to_numpy(alignment, info=None):
+    global MATPLOTLIB_FLAG
+    if not MATPLOTLIB_FLAG:
+        import matplotlib
+
+        matplotlib.use("Agg")
+        MATPLOTLIB_FLAG = True
+        mpl_logger = logging.getLogger("matplotlib")
+        mpl_logger.setLevel(logging.WARNING)
+    import matplotlib.pylab as plt
+    import numpy as np
+
+    fig, ax = plt.subplots(figsize=(6, 4))
+    im = ax.imshow(
+        alignment.transpose(), aspect="auto", origin="lower", interpolation="none"
+    )
+    fig.colorbar(im, ax=ax)
+    xlabel = "Decoder timestep"
+    if info is not None:
+        xlabel += "\n\n" + info
+    plt.xlabel(xlabel)
+    plt.ylabel("Encoder timestep")
+    plt.tight_layout()
+
+    fig.canvas.draw()
+    data = np.fromstring(fig.canvas.tostring_rgb(), dtype=np.uint8, sep="")
+    data = data.reshape(fig.canvas.get_width_height()[::-1] + (3,))
+    plt.close()
+    return data
+
+def load_filepaths_and_text(filename, split="|"):
+    with open(filename, encoding="utf-8") as f:
+        filepaths_and_text = [line.strip().split(split) for line in f]
+    return filepaths_and_text
+
+
+def get_hparams(init=True):
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "-c",
+        "--config",
+        type=str,
+        default="./configs/base.json",
+        help="JSON file for configuration",
+    )
+    parser.add_argument('--local_rank', type=int, default=0)
+    parser.add_argument('--world-size', type=int, default=1)
+    parser.add_argument('--port', type=int, default=10000)
+    parser.add_argument("-m", "--model", type=str, required=True, help="Model name")
+    parser.add_argument('--pretrain_G', type=str, default=None,
+                            help='pretrain model')
+    parser.add_argument('--pretrain_D', type=str, default=None,
+                            help='pretrain model D')
+    parser.add_argument('--pretrain_dur', type=str, default=None,
+                            help='pretrain model duration')
+
+    args = parser.parse_args()
+    model_dir = os.path.join("./logs", args.model)
+
+    os.makedirs(model_dir, exist_ok=True)
+
+    config_path = args.config
+    config_save_path = os.path.join(model_dir, "config.json")
+    if init:
+        with open(config_path, "r") as f:
+            data = f.read()
+        with open(config_save_path, "w") as f:
+            f.write(data)
+    else:
+        with open(config_save_path, "r") as f:
+            data = f.read()
+    config = json.loads(data)
+
+    hparams = HParams(**config)
+    hparams.model_dir = model_dir
+    hparams.pretrain_G = args.pretrain_G
+    hparams.pretrain_D = args.pretrain_D
+    hparams.pretrain_dur = args.pretrain_dur
+    hparams.port = args.port
+    return hparams
+
+
+def clean_checkpoints(path_to_models="logs/44k/", n_ckpts_to_keep=2, sort_by_time=True):
+    """Freeing up space by deleting saved ckpts
+
+    Arguments:
+    path_to_models    --  Path to the model directory
+    n_ckpts_to_keep   --  Number of ckpts to keep, excluding G_0.pth and D_0.pth
+    sort_by_time      --  True -> chronologically delete ckpts
+                          False -> lexicographically delete ckpts
+    """
+    import re
+
+    ckpts_files = [
+        f
+        for f in os.listdir(path_to_models)
+        if os.path.isfile(os.path.join(path_to_models, f))
+    ]
+
+    def name_key(_f):
+        return int(re.compile("._(\\d+)\\.pth").match(_f).group(1))
+
+    def time_key(_f):
+        return os.path.getmtime(os.path.join(path_to_models, _f))
+
+    sort_key = time_key if sort_by_time else name_key
+
+    def x_sorted(_x):
+        return sorted(
+            [f for f in ckpts_files if f.startswith(_x) and not f.endswith("_0.pth")],
+            key=sort_key,
+        )
+
+    to_del = [
+        os.path.join(path_to_models, fn)
+        for fn in (x_sorted("G")[:-n_ckpts_to_keep] + x_sorted("D")[:-n_ckpts_to_keep])
+    ]
+
+    def del_info(fn):
+        return logger.info(f".. Free up space by deleting ckpt {fn}")
+
+    def del_routine(x):
+        return [os.remove(x), del_info(x)]
+
+    [del_routine(fn) for fn in to_del]
+
+
+def get_hparams_from_dir(model_dir):
+    config_save_path = os.path.join(model_dir, "config.json")
+    with open(config_save_path, "r", encoding="utf-8") as f:
+        data = f.read()
+    config = json.loads(data)
+
+    hparams = HParams(**config)
+    hparams.model_dir = model_dir
+    return hparams
+
+
+def get_hparams_from_file(config_path):
+    with open(config_path, "r", encoding="utf-8") as f:
+        data = f.read()
+    config = json.loads(data)
+
+    hparams = HParams(**config)
+    return hparams
+
+
+def check_git_hash(model_dir):
+    source_dir = os.path.dirname(os.path.realpath(__file__))
+    if not os.path.exists(os.path.join(source_dir, ".git")):
+        logger.warn(
+            "{} is not a git repository, therefore hash value comparison will be ignored.".format(
+                source_dir
+            )
+        )
+        return
+
+    cur_hash = subprocess.getoutput("git rev-parse HEAD")
+
+    path = os.path.join(model_dir, "githash")
+    if os.path.exists(path):
+        saved_hash = open(path).read()
+        if saved_hash != cur_hash:
+            logger.warn(
+                "git hash values are different. {}(saved) != {}(current)".format(
+                    saved_hash[:8], cur_hash[:8]
+                )
+            )
+    else:
+        open(path, "w").write(cur_hash)
+
+
+def get_logger(model_dir, filename="train.log"):
+    global logger
+    logger = logging.getLogger(os.path.basename(model_dir))
+    logger.setLevel(logging.DEBUG)
+
+    formatter = logging.Formatter("%(asctime)s\t%(name)s\t%(levelname)s\t%(message)s")
+    if not os.path.exists(model_dir):
+        os.makedirs(model_dir, exist_ok=True)
+    h = logging.FileHandler(os.path.join(model_dir, filename))
+    h.setLevel(logging.DEBUG)
+    h.setFormatter(formatter)
+    logger.addHandler(h)
+    return logger
+
+
+class HParams:
+    def __init__(self, **kwargs):
+        for k, v in kwargs.items():
+            if type(v) == dict:
+                v = HParams(**v)
+            self[k] = v
+
+    def keys(self):
+        return self.__dict__.keys()
+
+    def items(self):
+        return self.__dict__.items()
+
+    def values(self):
+        return self.__dict__.values()
+
+    def __len__(self):
+        return len(self.__dict__)
+
+    def __getitem__(self, key):
+        return getattr(self, key)
+
+    def __setitem__(self, key, value):
+        return setattr(self, key, value)
+
+    def __contains__(self, key):
+        return key in self.__dict__
+
+    def __repr__(self):
+        return self.__dict__.__repr__()

tcp_server/src/lib/utils.py

@@ -0,0 +1,36 @@
+import time
+import sys
+import json
+from .constants import SETTINGS_PATH
+
+
+class ThrottledCallback:
+    def __init__(self, callback, min_interval):
+        self.callback = callback
+        self.min_interval = min_interval
+        self.last_call = 0
+
+    def __call__(self, *args, **kwargs):
+        current_time = time.time()
+        if current_time - self.last_call > self.min_interval:
+            self.callback(*args, **kwargs)
+            self.last_call = current_time
+
+
+def is_macos():
+    return sys.platform == 'darwin'
+
+
+def is_windows():
+    return sys.platform == 'win32'
+
+
+def is_linux():
+    return sys.platform == 'linux'
+
+
+def get_settings(key):
+    with open(SETTINGS_PATH) as f:
+        settings = json.load(f)
+
+    return settings[key]

tcp_server/src/lib/wake_word/api.py

@@ -0,0 +1,91 @@
+import os
+import time
+import numpy as np
+from openwakeword.model import Model as WakeWordModel
+
+from ..constants import WAKE_WORD_MODEL_FOLDER_PATH
+
+class WakeWord:
+    def __init__(self, asr, model_path, device='cpu', detection_threshold=0.5):
+        tic = time.perf_counter()
+        self.log('Loading model...')
+
+        self.log(f'Device: {device}')
+
+        self.asr = asr
+        self.model_path = model_path
+        self.device = device
+        self.detection_threshold = detection_threshold
+        self.chunk_size = 1280
+        self.audio = None
+        self.is_listening = False
+        self.is_enabled = False
+
+        if not os.path.exists(model_path):
+            self.log(f'Wake word model not found at {model_path}')
+            return
+
+        # @see https://github.com/dscripka/openWakeWord/blob/main/openwakeword/model.py#L38
+        # @see https://github.com/dscripka/openWakeWord/blob/main/openwakeword/utils.py#L38
+        self.model = WakeWordModel(
+            device=self.device,
+            wakeword_models=[self.model_path],
+            melspec_model_path=os.path.join(WAKE_WORD_MODEL_FOLDER_PATH, 'melspectrogram.onnx'),
+            embedding_model_path=os.path.join(WAKE_WORD_MODEL_FOLDER_PATH, 'embedding.onnx'),
+            ncpu=1,
+            inference_framework='onnx'
+        )
+
+        self.log('Model loaded')
+        toc = time.perf_counter()
+
+        self.log(f'Time taken to load model: {toc - tic:0.4f} seconds')
+
+        self.is_enabled = True
+
+    def reset_model_state(self):
+        """
+        Reset the wake word model's prediction buffer to avoid false triggers
+        """
+        for mdl in self.model.prediction_buffer.keys():
+            self.model.prediction_buffer[mdl] = []
+
+    def start_listening(self):
+        if self.is_enabled:
+            self.asr.is_recording = False
+            self.is_listening = True
+            self.audio = None
+
+            self.reset_model_state()
+
+            try:
+                self.log('Listening...')
+
+                while self.is_listening:
+                    # Get audio
+                    # Reuse the shared mic audio stream with ASR
+                    self.audio = np.frombuffer(self.asr.mic_stream.read(self.chunk_size), dtype=np.int16)
+
+                    # Feed to openWakeWord model
+                    prediction = self.model.predict(self.audio)
+
+                    for mdl in self.model.prediction_buffer.keys():
+                        scores = list(self.model.prediction_buffer[mdl])
+
+                        if scores[-1] > self.detection_threshold:
+                            self.log(f'Wakeword Detected! ({mdl})')
+                            self.stop_listening()
+                            self.asr.transcribed_callback('')
+                            self.asr.start_recording()
+            except Exception as e:
+                self.stop_listening()
+                self.log('Error:', e)
+
+    def stop_listening(self):
+        if self.is_enabled:
+            self.is_listening = False
+            self.log('Stopped listening')
+
+    @staticmethod
+    def log(*args, **kwargs):
+        print('[Wake word]', *args, **kwargs)

tcp_server/src/main.py

@@ -0,0 +1,172 @@
+import argparse
+import ctypes
+import glob
+import os
+import sys
+import threading
+from os.path import join
+from dotenv import load_dotenv
+
+DEFAULT_LEON_PROFILE = "just-me"
+DEFAULT_TCP_SERVER_HOST = "127.0.0.1"
+DEFAULT_TCP_SERVER_PORT = 5_367
+
+
+def resolve_leon_home() -> str:
+    configured_leon_home = os.getenv("LEON_HOME", "").strip()
+
+    if configured_leon_home:
+        return os.path.abspath(configured_leon_home)
+
+    return os.path.join(os.path.expanduser("~"), ".leon")
+
+
+def resolve_leon_profile() -> str:
+    return os.getenv("LEON_PROFILE", "").strip() or DEFAULT_LEON_PROFILE
+
+
+def resolve_leon_profile_path() -> str:
+    configured_profile_path = os.getenv("LEON_PROFILE_PATH", "").strip()
+
+    if configured_profile_path:
+        return os.path.abspath(configured_profile_path)
+
+    return os.path.join(resolve_leon_home(), "profiles", resolve_leon_profile())
+
+
+def _resolve_torch_root(pytorch_path: str) -> str | None:
+    normalized_path = os.path.abspath(pytorch_path)
+    if os.path.basename(normalized_path) == "torch" and os.path.isfile(
+        os.path.join(normalized_path, "__init__.py")
+    ):
+        return normalized_path
+
+    torch_candidate = os.path.join(normalized_path, "torch")
+    if os.path.isfile(os.path.join(torch_candidate, "__init__.py")):
+        return torch_candidate
+
+    torch_nested_candidate = os.path.join(normalized_path, "torch", "torch")
+    if os.path.isfile(os.path.join(torch_nested_candidate, "__init__.py")):
+        return torch_nested_candidate
+
+    return None
+
+
+def _add_pytorch_path(pytorch_path: str | None) -> str | None:
+    if not pytorch_path:
+        return None
+
+    torch_root = _resolve_torch_root(pytorch_path)
+    if torch_root:
+        sys.path.insert(0, os.path.dirname(torch_root))
+        return torch_root
+
+    sys.path.insert(0, os.path.abspath(pytorch_path))
+    return None
+
+
+def _set_library_paths(paths: list[str]) -> None:
+    if not paths:
+        return
+
+    existing_path = ""
+
+    if sys.platform.startswith("win"):
+        add_dll_directory = getattr(os, "add_dll_directory", None)
+        for path in paths:
+            if os.path.isdir(path) and add_dll_directory:
+                add_dll_directory(path)
+        existing_path = os.environ.get("PATH", "")
+        os.environ["PATH"] = (
+            os.pathsep.join([*paths, existing_path])
+            if existing_path
+            else os.pathsep.join(paths)
+        )
+        return
+
+    if sys.platform == "darwin":
+        existing_path = os.environ.get("DYLD_LIBRARY_PATH", "")
+        os.environ["DYLD_LIBRARY_PATH"] = (
+            os.pathsep.join([*paths, existing_path])
+            if existing_path
+            else os.pathsep.join(paths)
+        )
+        return
+
+    existing_path = os.environ.get("LD_LIBRARY_PATH", "")
+    os.environ["LD_LIBRARY_PATH"] = (
+        os.pathsep.join([*paths, existing_path])
+        if existing_path
+        else os.pathsep.join(paths)
+    )
+
+
+def _configure_external_libraries(
+    pytorch_path: str | None, nvidia_path: str | None
+) -> None:
+    lib_paths = []
+    torch_root = _add_pytorch_path(pytorch_path)
+
+    if torch_root:
+        torch_lib_path = os.path.join(torch_root, "lib")
+        if os.path.isdir(torch_lib_path):
+            lib_paths.append(torch_lib_path)
+
+    if nvidia_path:
+        nvidia_root = os.path.abspath(nvidia_path)
+        nvjitlink_pattern = os.path.join(
+            nvidia_root, "nvjitlink", "lib", "libnvJitLink.so.*"
+        )
+        for library in [
+            "cublas",
+            "cudnn",
+            "cusparse",
+            "cusparse_full",
+            "nccl",
+            "nvshmem",
+            "nvjitlink",
+        ]:
+            candidate = os.path.join(nvidia_root, library, "lib")
+            if os.path.isdir(candidate):
+                lib_paths.append(candidate)
+
+        if sys.platform.startswith("linux"):
+            nvjitlink_candidates = sorted(glob.glob(nvjitlink_pattern), reverse=True)
+            if nvjitlink_candidates:
+                ctypes.CDLL(nvjitlink_candidates[0], mode=ctypes.RTLD_GLOBAL)
+
+    _set_library_paths(lib_paths)
+
+
+def _parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser(description="Leon TCP server")
+    parser.add_argument(
+        "lang", nargs="?", default="en", help="Language code (e.g. en, fr)"
+    )
+    parser.add_argument("--pytorch-path", dest="pytorch_path", type=str, default=None)
+    parser.add_argument("--nvidia-path", dest="nvidia_path", type=str, default=None)
+    return parser.parse_args()
+
+
+args = _parse_args()
+os.environ["LEON_PY_TCP_SERVER_LANG"] = args.lang
+_configure_external_libraries(args.pytorch_path, args.nvidia_path)
+
+dotenv_path = join(resolve_leon_profile_path(), ".env")
+load_dotenv(dotenv_path)
+
+from lib.tcp_server import TCPServer
+
+tcp_server_host = os.environ.get("LEON_PY_TCP_SERVER_HOST", DEFAULT_TCP_SERVER_HOST)
+tcp_server_port = os.environ.get("LEON_PY_TCP_SERVER_PORT", DEFAULT_TCP_SERVER_PORT)
+
+tcp_server = TCPServer(tcp_server_host, tcp_server_port)
+
+# Use thread as ASR starts recording audio and it blocks the main thread
+asr_thread = threading.Thread(target=tcp_server.init_asr)
+asr_thread.start()
+
+tcp_server.init_tts()
+
+tcp_server_thread = threading.Thread(target=tcp_server.init)
+tcp_server_thread.start()

tcp_server/src/pyproject.toml

@@ -0,0 +1,23 @@
+[project]
+name = "leon-tcp-server"
+# Keep this metadata version for uv project compatibility only.
+# Leon runtime versioning still comes from `version.py`.
+version = "1.0.0"
+requires-python = "==3.11.9"
+dependencies = [
+  "python-dotenv==0.19.2",
+  "transformers==4.27.4",
+  "g2p-en==2.1.0",
+  "gruut[de,es,fr]==2.2.3",
+  "inflect==7.0.0",
+  "tqdm==4.66.4",
+  "soundfile==0.12.1",
+  "numba==0.59.1",
+  "faster-whisper==1.1.1",
+  "numpy==1.26.4",
+  "openwakeword==0.6.0",
+  "soundcard==0.4.5",
+]
+
+[tool.uv]
+package = false

tcp_server/src/version.py

@@ -0,0 +1 @@
+__version__ = '2.0.0'