wehub-resource-sync
ba4be087d5
Create PR to main with cherry-pick from release / cherry-pick (push) Failing after 0s
CICD NeMo / pre-flight (push) Failing after 0s
CICD NeMo / configure (push) Has been skipped
Build, validate, and release Neural Modules / pre-flight (push) Failing after 1s
CICD NeMo / code-linting (push) Has been skipped
Build, validate, and release Neural Modules / release (push) Has been skipped
Build, validate, and release Neural Modules / release-summary (push) Has been cancelled
CICD NeMo / cicd-test-container-build (push) Has been cancelled
CICD NeMo / cicd-import-tests (push) Has been cancelled
CICD NeMo / L0_Setup_Test_Data_And_Models (push) Has been cancelled
CICD NeMo / cicd-main-unit-tests (push) Has been cancelled
CICD NeMo / cicd-main-speech (push) Has been cancelled
CICD NeMo / Nemo_CICD_Test (push) Has been cancelled
CICD NeMo / Coverage (e2e) (push) Has been cancelled
CICD NeMo / Coverage (unit-test) (push) Has been cancelled
CodeQL / Analyze (python) (push) Has been cancelled
CICD NeMo / cicd-wait-in-queue (push) Has been cancelled
125 lines
7.8 KiB
YAML
125 lines
7.8 KiB
YAML
# ================================
|
|
# ASR Configuration
|
|
# ================================
|
|
asr:
|
|
model_name: stt_en_fastconformer_hybrid_large_streaming_multi # Pre-trained CTC/hybrid model from NGC/HuggingFace or local .nemo file path
|
|
device: cuda # Device for inference: 'cuda' or 'cpu'
|
|
device_id: 0 # GPU device ID
|
|
compute_dtype: bfloat16 # Compute precision: 'bfloat16' for Ampere+, 'float16' for older GPUs, or 'float32'
|
|
use_amp: false # Enable Automatic Mixed Precision
|
|
|
|
|
|
# ==========================================
|
|
# Inverse Text Normalization Configuration
|
|
# ==========================================
|
|
itn:
|
|
input_case: lower_cased # Input text case handling: 'lower_cased', 'cased'
|
|
whitelist: null # Custom whitelist for ITN processing
|
|
overwrite_cache: false # Whether to overwrite existing cache files
|
|
max_number_of_permutations_per_split: 729 # Maximum permutations allowed per text split during ITN processing
|
|
left_padding_size: 4 # Padding size (#spans) for ITN context
|
|
batch_size: 32 # Batch size for ITN inference
|
|
n_jobs: 16 # Number of parallel jobs for ITN processing
|
|
|
|
|
|
# ================================
|
|
# Neural Machine Translation Configuration
|
|
# ================================
|
|
nmt:
|
|
model_name: "utter-project/EuroLLM-1.7B-Instruct" # vLLM-supported model name
|
|
source_language: "English" # Source language code
|
|
target_language: "Russian" # Target language code
|
|
waitk: -1 # Max allowed lag (in words) between ASR transcript and translation; -1 disables it and uses only the longest common prefix between current and previous translations.
|
|
device: cuda # Device for translation: 'cuda'. 'cpu' is not supported.
|
|
device_id: 1 # GPU device ID for translation
|
|
batch_size: 16 # Batch size for translation, if -1, the batch size is equal to the ASR batch size
|
|
llm_params: # See https://docs.vllm.ai/en/v0.8.1/api/offline_inference/llm.html for more details
|
|
dtype: "auto" # Compute precision
|
|
seed: 42 # The seed to initialize the random number generator for sampling
|
|
sampling_params: # See https://docs.vllm.ai/en/v0.6.4/dev/sampling_params.html for more details
|
|
max_tokens: 100 # Maximum number of tokens to generate with LLM
|
|
temperature: 0.0 # LLM sampling temperature, default for translation is 0 (greedy)
|
|
top_p: 0.9 # The cumulative probability threshold for nucleus sampling
|
|
seed: 42 # The seed to initialize the random number generator for sampling
|
|
|
|
|
|
# ========================
|
|
# Confidence estimation
|
|
# ========================
|
|
confidence:
|
|
exclude_blank: true # Exclude blank tokens when calculating confidence
|
|
aggregation: mean # Aggregation method for confidence across time steps
|
|
method_cfg:
|
|
name: entropy # Confidence estimation method: 'max_prob' or 'entropy'
|
|
entropy_type: tsallis
|
|
alpha: 0.5
|
|
entropy_norm: exp
|
|
|
|
|
|
# ========================
|
|
# Endpointing settings
|
|
# ========================
|
|
endpointing:
|
|
stop_history_eou: 800 # Time window (ms) for evaluating EoU
|
|
residue_tokens_at_end: 2 # Number of residual tokens used for EoU
|
|
|
|
|
|
# ========================
|
|
# Streaming configuration
|
|
# ========================
|
|
streaming:
|
|
sample_rate: 16000 # Audio sample rate in Hz
|
|
batch_size: 256 # Number of audio frames per batch
|
|
word_boundary_tolerance: 4 # Tolerance for word boundaries
|
|
att_context_size: [70,13] # Attention context size: [70,13],[70,6],[70,1],[70,0]
|
|
use_cache: true # Whether to use cache for streaming
|
|
use_feat_cache: true # Whether to cache mel-spec features, set false to re-calculate all mel-spec features in audio buffer
|
|
chunk_size_in_secs: null # Amount of audio to load for each streaming step, e.g., 0.08s for FastConformer. Set to `null` for using default size equal to 1+lookahead frames.
|
|
request_type: frame # Type of request: frame or feature_buffer
|
|
num_slots: 1024 # Number of slots in the context manager: must be >= batch_size
|
|
|
|
|
|
# ========================
|
|
# Pipeline settings
|
|
# ========================
|
|
matmul_precision: high # Matrix multiplication precision: highest, high, medium
|
|
log_level: 20 # Logging level: 0 (NOTSET), 10 (DEBUG), 20 (INFO), 30 (WARNING), 40 (ERROR), 50 (CRITICAL)
|
|
pipeline_type: cache_aware # Pipeline type: buffered, cache_aware
|
|
asr_decoding_type: ctc # Decoding method: ctc or rnnt
|
|
|
|
# ========================
|
|
# Runtime arguments defined at runtime via command line
|
|
# ========================
|
|
audio_file: null # Path to audio file, directory, or manifest JSON
|
|
output_filename: null # Path to output transcription JSON file
|
|
output_dir: null # Directory to save time-aligned output
|
|
enable_itn: false # Whether to apply inverse text normalization
|
|
enable_nmt: false # Whether to apply neural machine translation
|
|
asr_output_granularity: segment # Output granularity: word or segment
|
|
cache_dir: null # Directory to store cache (e.g., .far files)
|
|
lang: null # Language code for ASR model
|
|
return_tail_result: false # Whether to return the tail labels left in the right padded side of the buffer
|
|
calculate_wer: true # Whether to calculate WER
|
|
calculate_bleu: true # Whether to calculate BLEU score
|
|
warmup_steps: 0 # Number of warmup steps for RTFx and LAAL calculation
|
|
run_steps: 1 # Number of run steps for RTFx and LAAL calculation
|
|
|
|
|
|
# ========================
|
|
# Metrics
|
|
# ========================
|
|
metrics:
|
|
asr:
|
|
gt_text_attr_name: text # Attribute name for ground truth text
|
|
clean_groundtruth_text: false # Whether to clean ground truth text
|
|
langid: en # Language code for text normalization; only "en" is supported
|
|
use_cer: false # Whether to use character error rate
|
|
ignore_capitalization: true # Whether to ignore capitalization
|
|
ignore_punctuation: true # Whether to ignore punctuation
|
|
strip_punc_space: false # Whether to strip punctuation and space
|
|
nmt:
|
|
gt_text_attr_name: answer # Attribute name for ground truth text
|
|
ignore_capitalization: false # Whether to ignore capitalization
|
|
ignore_punctuation: false # Whether to ignore punctuation
|
|
strip_punc_space: false # Whether to strip punctuation and space
|