huggingface--transformers/docs/source/en/model_doc/parakeet.md

<!--Copyright 2025 The NVIDIA NeMo Team and The HuggingFace Inc. team. All rights reserved.

Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at

http://www.apache.org/licenses/LICENSE-2.0

Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
specific language governing permissions and limitations under the License.

⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
rendered properly in your Markdown viewer.

-->
*This model was contributed to Hugging Face Transformers on 2025-09-25.*

<div class="flex flex-wrap space-x-1">
<img alt="SDPA" src="https://img.shields.io/badge/SDPA-DE3412?style=flat&logo=pytorch&logoColor=white">
</div>

# Parakeet

## Overview

Parakeet models, [introduced by NVIDIA NeMo](https://developer.nvidia.com/blog/pushing-the-boundaries-of-speech-recognition-with-nemo-parakeet-asr-models/), are models that combine a [Fast Conformer](https://docs.nvidia.com/nemo-framework/user-guide/latest/nemotoolkit/asr/models.html#fast-conformer) encoder with connectionist temporal classification (CTC), recurrent neural network transducer (RNNT) or token and duration transducer (TDT) decoder for automatic speech recognition.

**Model Architecture**

- **Fast Conformer Encoder**: A linearly scalable Conformer architecture that processes mel-spectrogram features and reduces sequence length through subsampling. This is more efficient version of the Conformer Encoder found in [FastSpeech2Conformer](./fastspeech2_conformer.md) (see [`ParakeetEncoder`] for the encoder implementation and details).
- [**ParakeetForCTC**](#parakeetforctc): a Fast Conformer Encoder + a CTC decoder
  - **CTC Decoder**: Simple but effective decoder consisting of:
    - 1D convolution projection from encoder hidden size to vocabulary size (for optimal NeMo compatibility).
    - CTC loss computation for training.
    - Greedy CTC decoding for inference.
- [**ParakeetForRNNT**](#parakeetforrnnt): a Fast Conformer Encoder + an RNN-T (RNN Transducer) decoder
  - **RNN-T Decoder**: Standard neural transducer:
    - LSTM prediction network maintains language context across token predictions.
    - Joint network combines encoder and decoder outputs.
    - Greedy transducer decoding for inference: a blank emission advances the encoder frame by one, a non-blank emission stays on the same frame.
- [**ParakeetForTDT**](#parakeetfortdt): a Fast Conformer Encoder + a TDT (Token Duration Transducer) decoder
  - **TDT Decoder**: Jointly predicts tokens and their durations, enabling efficient decoding:
    - LSTM prediction network maintains language context across token predictions.
    - Joint network combines encoder and decoder outputs.
    - Duration head predicts how many frames to skip, enabling fast inference.

The original implementation can be found in [NVIDIA NeMo](https://github.com/NVIDIA/NeMo).
Model checkpoints are to be found under [the NVIDIA organization](https://huggingface.co/nvidia/models?search=parakeet).

This model was contributed by [Nithin Rao Koluguri](https://huggingface.co/nithinraok), [Eustache Le Bihan](https://huggingface.co/eustlb), [Eric Bezzam](https://huggingface.co/bezzam), [Maksym Lypivskyi](https://huggingface.co/MaksL), and [Hainan Xu](https://huggingface.co/hainanx).

## Usage

### `ParakeetForCTC` usage

<hfoptions id="usage">
<hfoption id="Pipeline">

```python
from transformers import pipeline


pipe = pipeline("automatic-speech-recognition", model="nvidia/parakeet-ctc-1.1b")
out = pipe("https://huggingface.co/datasets/hf-internal-testing/dummy-audio-samples/resolve/main/bcn_weather.mp3")
print(out)
# {'text': 'yesterday it was thirty five degrees in barcelona but today the temperature will go down to minus twenty degrees'}
```

</hfoption>
<hfoption id="AutoModel">

```python
from datasets import Audio, load_dataset

from transformers import AutoModelForCTC, AutoProcessor

model_id = "nvidia/parakeet-ctc-1.1b"
processor = AutoProcessor.from_pretrained(model_id)
model = AutoModelForCTC.from_pretrained(model_id, device_map="auto")

ds = load_dataset("hf-internal-testing/librispeech_asr_dummy", "clean", split="validation")
ds = ds.cast_column("audio", Audio(sampling_rate=processor.feature_extractor.sampling_rate))
speech_samples = [el['array'] for el in ds["audio"][:5]]

inputs = processor(speech_samples, sampling_rate=processor.feature_extractor.sampling_rate)
inputs.to(model.device, dtype=model.dtype)
outputs = model.generate(**inputs)
print(processor.decode(outputs))
```

</hfoption>
</hfoptions>

### `ParakeetForRNNT` usage

<hfoptions id="rnnt-usage">
<hfoption id="Pipeline">

Parakeet RNN-T transcribes without casing or punctuation (like CTC), and the model can also perform token timestamping.

```py
from transformers import pipeline

pipe = pipeline("automatic-speech-recognition", model="nvidia/parakeet-rnnt-0.6b", revision="refs/pr/4")
out = pipe("https://huggingface.co/datasets/hf-internal-testing/dummy-audio-samples/resolve/main/bcn_weather.mp3")
print(out)
# {'text': 'yesterday it was thirty five degrees in barcelona but today the temperature will go down to minus twenty degrees'}
```

</hfoption>
<hfoption id="AutoModel">

```py
from transformers import AutoModelForRNNT, AutoProcessor
from datasets import load_dataset, Audio

model_id = "nvidia/parakeet-rnnt-0.6b"
revision = "refs/pr/4"
processor = AutoProcessor.from_pretrained(model_id, revision=revision)
model = AutoModelForRNNT.from_pretrained(model_id, revision=revision, device_map="auto")

ds = load_dataset("hf-internal-testing/librispeech_asr_dummy", "clean", split="validation")
ds = ds.cast_column("audio", Audio(sampling_rate=processor.feature_extractor.sampling_rate))
speech_samples = [el['array'] for el in ds["audio"][:5]]

inputs = processor(speech_samples, sampling_rate=processor.feature_extractor.sampling_rate)
inputs.to(model.device, dtype=model.dtype)
output = model.generate(**inputs, return_dict_in_generate=True)
print(processor.decode(output.sequences, skip_special_tokens=True))
```

</hfoption>
<hfoption id="Timestamping">

Unlike TDT (which predicts a per-token duration), each RNN-T token is emitted at a single encoder frame, so its start and end span exactly one frame.

```py
from datasets import Audio, load_dataset
from transformers import AutoModelForRNNT, AutoProcessor

model_id = "nvidia/parakeet-rnnt-0.6b"
revision = "refs/pr/4"
processor = AutoProcessor.from_pretrained(model_id, revision=revision)
model = AutoModelForRNNT.from_pretrained(model_id, revision=revision, device_map="auto")

ds = load_dataset("hf-internal-testing/librispeech_asr_dummy", "clean", split="validation")
ds = ds.cast_column("audio", Audio(sampling_rate=processor.feature_extractor.sampling_rate))
speech_samples = [el['array'] for el in ds["audio"][:1]]

inputs = processor(speech_samples, sampling_rate=processor.feature_extractor.sampling_rate)
inputs.to(model.device, dtype=model.dtype)
output = model.generate(**inputs, return_dict_in_generate=True)
decoded_output, decoded_timestamps = processor.decode(
    output.sequences,
    durations=output.durations,
    skip_special_tokens=True,
)
print("Transcription:", decoded_output)
print("\nTimestamped tokens:", decoded_timestamps)

"""
Transcription: ['mister quilter is the apostle of the middle classes and we are glad to welcome his gospel']

Timestamped tokens: [[{'token': 'm', 'start': 0.4, 'end': 0.48}, {'token': 'is', 'start': 0.56, 'end': 0.64}, {'token': 'ter', 'start': 0.56, 'end': 0.64}, {'token': ' qu', 'start': 0.72, 'end': 0.8}, {'token': 'il', 'start': 0.96, 'end': 1.04}, {'token': 'ter', 'start': 1.12, 'end': 1.2}, {'token': ' is', 'start': 1.36, 'end': 1.44}, {'token': ' the', 'start': 1.52, 'end': 1.6}, {'token': ' ap', 'start': 1.68, 'end': 1.76}, {'token': 'o', 'start': 1.76, 'end': 1.84}, {'token': 'st', 'start': 1.84, 'end': 1.92}, {'token': 'le', 'start': 2.0, 'end': 2.08}, {'token': ' of', 'start': 2.16, 'end': 2.24}, {'token': ' the', 'start': 2.24, 'end': 2.32}, {'token': ' m', 'start': 2.4, 'end': 2.48}, {'token': 'id', 'start': 2.48, 'end': 2.56}, {'token': 'd', 'start': 2.56, 'end': 2.64}, {'token': 'le', 'start': 2.56, 'end': 2.64}, {'token': ' cl', 'start': 2.8, 'end': 2.88}, {'token': 'ass', 'start': 2.88, 'end': 2.96}, {'token': 'es', 'start': 3.12, 'end': 3.2}, {'token': ' and', 'start': 3.28, 'end': 3.36}, {'token': ' we', 'start': 3.44, 'end': 3.52}, {'token': ' are', 'start': 3.6, 'end': 3.68}, {'token': ' gl', 'start': 3.84, 'end': 3.92}, {'token': 'ad', 'start': 3.92, 'end': 4.0}, {'token': ' to', 'start': 4.08, 'end': 4.16}, {'token': ' we', 'start': 4.24, 'end': 4.32}, {'token': 'l', 'start': 4.32, 'end': 4.4}, {'token': 'c', 'start': 4.4, 'end': 4.48}, {'token': 'ome', 'start': 4.48, 'end': 4.56}, {'token': ' his', 'start': 4.72, 'end': 4.8}, {'token': ' go', 'start': 4.96, 'end': 5.04}, {'token': 's', 'start': 5.04, 'end': 5.12}, {'token': 'pe', 'start': 5.2, 'end': 5.28}, {'token': 'l', 'start': 5.36, 'end': 5.44}]]
"""
```

</hfoption>
</hfoptions>

### `ParakeetForTDT` usage

<hfoptions id="tdt-usage">
<hfoption id="Pipeline">

Parakeet TDT transcripts include casing, and the model can also perform token timestamping.

```py
from transformers import pipeline

pipe = pipeline("automatic-speech-recognition", model="nvidia/parakeet-tdt-0.6b-v3")
out = pipe("https://huggingface.co/datasets/hf-internal-testing/dummy-audio-samples/resolve/main/bcn_weather.mp3")
print(out)
# {'text': 'Yesterday it was 35 degrees in Barcelona, but today the temperature will go down to minus 20 degrees.'}
```

</hfoption>
<hfoption id="AutoModel">

```py
from transformers import AutoModelForTDT, AutoProcessor
from datasets import load_dataset, Audio

model_id = "nvidia/parakeet-tdt-0.6b-v3"
processor = AutoProcessor.from_pretrained(model_id)
model = AutoModelForTDT.from_pretrained(model_id, device_map="auto")

ds = load_dataset("hf-internal-testing/librispeech_asr_dummy", "clean", split="validation")
ds = ds.cast_column("audio", Audio(sampling_rate=processor.feature_extractor.sampling_rate))
speech_samples = [el['array'] for el in ds["audio"][:5]]

inputs = processor(speech_samples, sampling_rate=processor.feature_extractor.sampling_rate)
inputs.to(model.device, dtype=model.dtype)
output = model.generate(**inputs, return_dict_in_generate=True)
print(processor.decode(output.sequences, skip_special_tokens=True))
```

</hfoption>
<hfoption id="Timestamping">

```py
from datasets import Audio, load_dataset
from transformers import AutoModelForTDT, AutoProcessor

model_id = "nvidia/parakeet-tdt-0.6b-v3"
processor = AutoProcessor.from_pretrained(model_id)
model = AutoModelForTDT.from_pretrained(model_id, device_map="auto")

ds = load_dataset("hf-internal-testing/librispeech_asr_dummy", "clean", split="validation")
ds = ds.cast_column("audio", Audio(sampling_rate=processor.feature_extractor.sampling_rate))
speech_samples = [el['array'] for el in ds["audio"][:1]]

inputs = processor(speech_samples, sampling_rate=processor.feature_extractor.sampling_rate)
inputs.to(model.device, dtype=model.dtype)
output = model.generate(**inputs, return_dict_in_generate=True)
decoded_output, decoded_timestamps = processor.decode(
    output.sequences,
    durations=output.durations,
    skip_special_tokens=True,
)
print("Transcription:", decoded_output)
print("\nTimestamped tokens:", decoded_timestamps)

"""
Transcription: ['mister Quilter is the apostle of the middle classes, and we are glad to welcome his gospel.']

Timestamped tokens: [[{'token': 'm', 'start': 0.24, 'end': 0.48}, {'token': 'ister', 'start': 0.48, 'end': 0.64}, {'token': 'Qu', 'start': 0.64, 'end': 0.88}, {'token': 'il', 'start': 0.88, 'end': 1.12}, {'token': 'ter', 'start': 1.12, 'end': 1.36}, {'token': 'is', 'start': 1.36, 'end': 1.44}, {'token': 'the', 'start': 1.44, 'end': 1.6}, {'token': 'ap', 'start': 1.6, 'end': 1.76}, {'token': 'ost', 'start': 1.76, 'end': 1.92}, {'token': 'le', 'start': 2.0, 'end': 2.16}, {'token': 'of', 'start': 2.16, 'end': 2.24}, {'token': 'the', 'start': 2.24, 'end': 2.4}, {'token': 'mid', 'start': 2.4, 'end': 2.48}, {'token': 'd', 'start': 2.48, 'end': 2.56}, {'token': 'le', 'start': 2.56, 'end': 2.64}, {'token': 'clas', 'start': 2.72, 'end': 2.88}, {'token': 's', 'start': 2.88, 'end': 3.04}, {'token': 'es', 'start': 3.04, 'end': 3.12}, {'token': ',', 'start': 3.12, 'end': 3.12}, {'token': 'and', 'start': 3.2800000000000002, 'end': 3.44}, {'token': 'we', 'start': 3.44, 'end': 3.6}, {'token': 'are', 'start': 3.6, 'end': 3.7600000000000002}, {'token': 'gl', 'start': 3.7600000000000002, 'end': 3.92}, {'token': 'ad', 'start': 3.92, 'end': 4.08}, {'token': 'to', 'start': 4.08, 'end': 4.24}, {'token': 'wel', 'start': 4.24, 'end': 4.4}, {'token': 'c', 'start': 4.4, 'end': 4.48}, {'token': 'ome', 'start': 4.48, 'end': 4.72}, {'token': 'his', 'start': 4.72, 'end': 4.96}, {'token': 'gos', 'start': 4.96, 'end': 5.12}, {'token': 'pel', 'start': 5.36, 'end': 5.6000000000000005}, {'token': '.', 'start': 5.6000000000000005, 'end': 5.6000000000000005}]]
"""
```

</hfoption>
</hfoptions>

### Making The Model Go Brrr

Parakeet supports full-graph compilation with CUDA graphs! This optimization is most effective when you know the maximum audio length you want to transcribe. The key idea is using static input shapes to avoid recompilation. For example, if you know your audio will be under 30 seconds, you can use the processor to pad all inputs to 30 seconds, preparing consistent input features and attention masks. See the example below!

```python
import torch
from datasets import Audio, load_dataset

from transformers import AutoModelForCTC, AutoProcessor


processor = AutoProcessor.from_pretrained("nvidia/parakeet-ctc-1.1b")
model = AutoModelForCTC.from_pretrained("nvidia/parakeet-ctc-1.1b", device_map="auto")

ds = load_dataset("hf-internal-testing/librispeech_asr_dummy", "clean", split="validation")
ds = ds.cast_column("audio", Audio(sampling_rate=processor.feature_extractor.sampling_rate))
speech_samples = [el['array'] for el in ds["audio"][:5]]

# Compile the generate method with fullgraph and CUDA graphs
model.generate = torch.compile(model.generate, fullgraph=True, mode="reduce-overhead")

# let's define processor kwargs to pad to 30 seconds
processor_kwargs = {
    "padding": "max_length",
    "max_length": 30 * processor.feature_extractor.sampling_rate,
}

# Define a timing context using CUDA events
class TimerContext:
    def __init__(self, name="Execution"):
        self.name = name
        self.start_event = None
        self.end_event = None

    def __enter__(self):
        # Use CUDA events for more accurate GPU timing
        self.start_event = torch.cuda.Event(enable_timing=True)
        self.end_event = torch.cuda.Event(enable_timing=True)
        self.start_event.record()
        return self

    def __exit__(self, *args):
        self.end_event.record()
        torch.cuda.synchronize()
        elapsed_time = self.start_event.elapsed_time(self.end_event) / 1000.0
        print(f"{self.name} time: {elapsed_time:.4f} seconds")


inputs = processor(speech_samples[0], **processor_kwargs)
inputs.to(model.device, dtype=model.dtype)
print("\n" + "="*50)
print("First generation - compiling...")
# Generate with the compiled model
with TimerContext("First generation"):
    outputs = model.generate(**inputs)
print(processor.decode(outputs))

inputs = processor(speech_samples[1], **processor_kwargs)
inputs.to(model.device, dtype=model.dtype)
print("\n" + "="*50)
print("Second generation - recording CUDA graphs...")
with TimerContext("Second generation"):
    outputs = model.generate(**inputs)
print(processor.decode(outputs))

inputs = processor(speech_samples[2], **processor_kwargs)
inputs.to(model.device, dtype=model.dtype)
print("\n" + "="*50)
print("Third generation - fast !!!")
with TimerContext("Third generation"):
    outputs = model.generate(**inputs)
print(processor.decode(outputs))

inputs = processor(speech_samples[3], **processor_kwargs)
inputs.to(model.device, dtype=model.dtype)
print("\n" + "="*50)
print("Fourth generation - still fast !!!")
with TimerContext("Fourth generation"):
    outputs = model.generate(**inputs)
print(processor.decode(outputs))
```

### CTC Training

```python
from datasets import Audio, load_dataset
from transformers import AutoModelForCTC, AutoProcessor

model_id = "nvidia/parakeet-ctc-1.1b"
NUM_SAMPLES = 5

processor = AutoProcessor.from_pretrained(model_id)
model = AutoModelForCTC.from_pretrained(model_id, device_map="auto")
model.train()

ds = load_dataset("hf-internal-testing/librispeech_asr_dummy", "clean", split="validation")
ds = ds.cast_column("audio", Audio(sampling_rate=processor.feature_extractor.sampling_rate))
speech_samples = [el['array'] for el in ds["audio"][:NUM_SAMPLES]]
text_samples = ds["text"][:NUM_SAMPLES]

# passing `text` to the processor will prepare inputs' `labels` key
inputs = processor(audio=speech_samples, text=text_samples, sampling_rate=processor.feature_extractor.sampling_rate)
inputs.to(model.device, dtype=model.dtype)

outputs = model(**inputs)
print("Loss:", outputs.loss.item())
outputs.loss.backward()
```

### RNN-T Training

> [!NOTE]
> Computing the RNN-T loss requires [torchaudio](https://pytorch.org/audio) (`pip install torchaudio`).

```py
from datasets import Audio, load_dataset
import torch
from transformers import AutoModelForRNNT, AutoProcessor

model_id = "nvidia/parakeet-rnnt-0.6b"
revision = "refs/pr/4"
NUM_SAMPLES = 4

processor = AutoProcessor.from_pretrained(model_id, revision=revision)
model = AutoModelForRNNT.from_pretrained(model_id, revision=revision,  device_map="auto")
model.train()

ds = load_dataset("hf-internal-testing/librispeech_asr_dummy", "clean", split="validation")
ds = ds.cast_column("audio", Audio(sampling_rate=processor.feature_extractor.sampling_rate))
speech_samples = [el['array'] for el in ds["audio"][:NUM_SAMPLES]]
text_samples = ds["text"][:NUM_SAMPLES]

# passing `text` to the processor will prepare inputs' `labels` key
inputs = processor(audio=speech_samples, text=text_samples, sampling_rate=processor.feature_extractor.sampling_rate)
inputs.to(model.device, dtype=model.dtype)

outputs = model(**inputs)
print("Loss:", outputs.loss.item())
outputs.loss.backward()
```


### TDT Training

```py
from datasets import Audio, load_dataset
import torch
from transformers import AutoModelForTDT, AutoProcessor

model_id = "nvidia/parakeet-tdt-0.6b-v3"
NUM_SAMPLES = 4

processor = AutoProcessor.from_pretrained(model_id)
model = AutoModelForTDT.from_pretrained(model_id, device_map="auto")
model.train()

ds = load_dataset("hf-internal-testing/librispeech_asr_dummy", "clean", split="validation")
ds = ds.cast_column("audio", Audio(sampling_rate=processor.feature_extractor.sampling_rate))
speech_samples = [el['array'] for el in ds["audio"][:NUM_SAMPLES]]
text_samples = ds["text"][:NUM_SAMPLES]

# passing `text` to the processor will prepare inputs' `labels` key
inputs = processor(audio=speech_samples, text=text_samples, sampling_rate=processor.feature_extractor.sampling_rate)
inputs.to(model.device, dtype=model.dtype)

outputs = model(**inputs)
print("Loss:", outputs.loss.item())
outputs.loss.backward()
```


## ParakeetTokenizer

[[autodoc]] ParakeetTokenizer

## ParakeetFeatureExtractor

[[autodoc]] ParakeetFeatureExtractor
    - __call__

## ParakeetProcessor

[[autodoc]] ParakeetProcessor
    - __call__
    - decode

## ParakeetEncoderConfig

[[autodoc]] ParakeetEncoderConfig

## ParakeetCTCConfig

[[autodoc]] ParakeetCTCConfig

## ParakeetRNNTConfig

[[autodoc]] ParakeetRNNTConfig

## ParakeetTDTConfig

[[autodoc]] ParakeetTDTConfig

## ParakeetEncoder

[[autodoc]] ParakeetEncoder

## ParakeetForCTC

[[autodoc]] ParakeetForCTC

## ParakeetForRNNT

[[autodoc]] ParakeetForRNNT

## ParakeetForTDT

[[autodoc]] ParakeetForTDT