Distil-Whisper: distil-small.en

Distil-Whisper was proposed in the paper Robust Knowledge Distillation via Large-Scale Pseudo Labelling. It is a distilled version of the Whisper model that is 6 times faster, 49% smaller, and performs within 1% WER on out-of-distribution evaluation sets.

This is the repository for distil-small.en, a distilled variant of Whisper small.en. It is the smallest Distil-Whisper checkpoint, with just 166M parameters, making it the ideal choice for memory constrained applications (e.g. on-device).

For most other applications, the distil-medium.en or distil-large-v2 checkpoints are recommended, since they are both faster and achieve better WER results:

Model	Params / M	Rel. Latency ↑	Short-Form WER ↓	Long-Form WER ↓
large-v3	1550	1.0	8.4	11.0
large-v2	1550	1.0	9.1	11.7
distil-large-v3	756	6.3	9.7	10.8
distil-large-v2	756	5.8	10.1	11.6
distil-medium.en	394	6.8	11.1	12.4
distil-small.en	166	5.6	12.1	12.8

Note: Distil-Whisper is currently only available for English speech recognition. We are working with the community to distill Whisper on other languages. If you are interested in distilling Whisper in your language, check out the provided training code. We will update the Distil-Whisper repository with multilingual checkpoints when ready!

Why is distil-small.en slower than distil-large-v2?

While distil-medium.en and distil-large-v2 use two decoder layers each, distil-small.en uses four. Using more decoder layers improves the WER performance of the model, at the expense of slower inference speed. We found that four layers was the minimum required to get reasonable WER performance for distil-small.en, where it performs to within 3% WER of Whisper large-v2 while being 5.6x faster. When we tried distilling with just two layers, the model was over 5% worse than large-v2, albeit 7.8x faster. We leave distilling a two layer small.en model as future works.

Usage

Distil-Whisper is supported in Hugging Face 🤗 Transformers from version 4.35 onwards. To run the model, first install the latest version of the Transformers library. For this example, we'll also install 🤗 Datasets to load toy audio dataset from the Hugging Face Hub:

pip install --upgrade pip
pip install --upgrade transformers accelerate datasets[audio]

Short-Form Transcription

The model can be used with the pipeline class to transcribe short-form audio files (< 30-seconds) as follows:

import torch
from transformers import AutoModelForSpeechSeq2Seq, AutoProcessor, pipeline
from datasets import load_dataset


device = "cuda:0" if torch.cuda.is_available() else "cpu"
torch_dtype = torch.float16 if torch.cuda.is_available() else torch.float32

model_id = "distil-whisper/distil-small.en"

model = AutoModelForSpeechSeq2Seq.from_pretrained(
    model_id, torch_dtype=torch_dtype, low_cpu_mem_usage=True, use_safetensors=True
)
model.to(device)

processor = AutoProcessor.from_pretrained(model_id)

pipe = pipeline(
    "automatic-speech-recognition",
    model=model,
    tokenizer=processor.tokenizer,
    feature_extractor=processor.feature_extractor,
    max_new_tokens=128,
    torch_dtype=torch_dtype,
    device=device,
)

dataset = load_dataset("hf-internal-testing/librispeech_asr_dummy", "clean", split="validation")
sample = dataset[0]["audio"]

result = pipe(sample)
print(result["text"])

To transcribe a local audio file, simply pass the path to your audio file when you call the pipeline:

- result = pipe(sample)
+ result = pipe("audio.mp3")

Long-Form Transcription

Distil-Whisper uses a chunked algorithm to transcribe long-form audio files (> 30-seconds). In practice, this chunked long-form algorithm is 9x faster than the sequential algorithm proposed by OpenAI in the Whisper paper (see Table 7 of the Distil-Whisper paper).

To enable chunking, pass the chunk_length_s parameter to the pipeline. For Distil-Whisper, a chunk length of 15-seconds is optimal. To activate batching, pass the argument batch_size:

import torch
from transformers import AutoModelForSpeechSeq2Seq, AutoProcessor, pipeline
from datasets import load_dataset


device = "cuda:0" if torch.cuda.is_available() else "cpu"
torch_dtype = torch.float16 if torch.cuda.is_available() else torch.float32

model_id = "distil-whisper/distil-small.en"

model = AutoModelForSpeechSeq2Seq.from_pretrained(
    model_id, torch_dtype=torch_dtype, low_cpu_mem_usage=True, use_safetensors=True
)
model.to(device)

processor = AutoProcessor.from_pretrained(model_id)

pipe = pipeline(
    "automatic-speech-recognition",
    model=model,
    tokenizer=processor.tokenizer,
    feature_extractor=processor.feature_extractor,
    max_new_tokens=128,
    chunk_length_s=15,
    batch_size=16,
    torch_dtype=torch_dtype,
    device=device,
)

dataset = load_dataset("distil-whisper/librispeech_long", "default", split="validation")
sample = dataset[0]["audio"]

result = pipe(sample)
print(result["text"])

Speculative Decoding

Distil-Whisper can be used as an assistant model to Whisper for speculative decoding. Speculative decoding mathematically ensures the exact same outputs as Whisper are obtained while being 2 times faster. This makes it the perfect drop-in replacement for existing Whisper pipelines, since the same outputs are guaranteed.

In the following code-snippet, we load the assistant Distil-Whisper model standalone to the main Whisper pipeline. We then specify it as the "assistant model" for generation:

from transformers import pipeline, AutoModelForSpeechSeq2Seq, AutoProcessor
import torch
from datasets import load_dataset

device = "cuda:0" if torch.cuda.is_available() else "cpu"
torch_dtype = torch.float16 if torch.cuda.is_available() else torch.float32

assistant_model_id = "distil-whisper/distil-small.en"

assistant_model = AutoModelForSpeechSeq2Seq.from_pretrained(
    assistant_model_id, torch_dtype=torch_dtype, low_cpu_mem_usage=True, use_safetensors=True
)
assistant_model.to(device)

model_id = "openai/whisper-medium.en"

model = AutoModelForSpeechSeq2Seq.from_pretrained(
    model_id, torch_dtype=torch_dtype, low_cpu_mem_usage=True, use_safetensors=True
)
model.to(device)

processor = AutoProcessor.from_pretrained(model_id)

pipe = pipeline(
    "automatic-speech-recognition",
    model=model,
    tokenizer=processor.tokenizer,
    feature_extractor=processor.feature_extractor,
    max_new_tokens=128,
    generate_kwargs={"assistant_model": assistant_model},
    torch_dtype=torch_dtype,
    device=device,
)

dataset = load_dataset("hf-internal-testing/librispeech_asr_dummy", "clean", split="validation")
sample = dataset[0]["audio"]

result = pipe(sample)
print(result["text"])

Additional Speed & Memory Improvements

You can apply additional speed and memory improvements to Distil-Whisper which we cover in the following.

Flash Attention

We recommend using Flash-Attention 2 if your GPU allows for it. To do so, you first need to install Flash Attention:

pip install flash-attn --no-build-isolation

and then all you have to do is to pass use_flash_attention_2=True to from_pretrained:

- model = AutoModelForSpeechSeq2Seq.from_pretrained(model_id, torch_dtype=torch_dtype, low_cpu_mem_usage=True, use_safetensors=True)
+ model = AutoModelForSpeechSeq2Seq.from_pretrained(model_id, torch_dtype=torch_dtype, low_cpu_mem_usage=True, use_safetensors=True, use_flash_attention_2=True)

Torch Scale-Product-Attention (SDPA)

If your GPU does not support Flash Attention, we recommend making use of BetterTransformers. To do so, you first need to install optimum:

pip install --upgrade optimum

And then convert your model to a "BetterTransformer" model before using it:

model = AutoModelForSpeechSeq2Seq.from_pretrained(model_id, torch_dtype=torch_dtype, low_cpu_mem_usage=True, use_safetensors=True)
+ model = model.to_bettertransformer()

Running Distil-Whisper in openai-whisper

To use the model in the original Whisper format, first ensure you have the openai-whisper package installed:

pip install --upgrade openai-whisper

The following code-snippet demonstrates how to transcribe a sample file from the LibriSpeech dataset loaded using 🤗 Datasets:

import torch
from datasets import load_dataset
from huggingface_hub import hf_hub_download
from whisper import load_model, transcribe

distil_small_en = hf_hub_download(repo_id="distil-whisper/distil-small.en", filename="original-model.bin")
model = load_model(distil_small_en)

dataset = load_dataset("hf-internal-testing/librispeech_asr_dummy", "clean", split="validation")
sample = dataset[0]["audio"]["array"]
sample = torch.from_numpy(sample).float()

pred_out = transcribe(model, audio=sample)
print(pred_out["text"])

Note that the model weights will be downloaded and saved to your cache the first time you run the example. Subsequently, you can re-use the same example, and the weights will be loaded directly from your cache without having to download them again.

To transcribe a local audio file, simply pass the path to the audio file as the audio argument to transcribe:

pred_out = transcribe(model, audio="audio.mp3")

Whisper.cpp

Distil-Whisper can be run from the Whisper.cpp repository with the original sequential long-form transcription algorithm. In a provisional benchmark on Mac M1, distil-small.en is over 4x faster than large-v2, while performing to within 1.4% WER over long-form audio.

Steps for getting started:

1. Clone the Whisper.cpp repository:

git clone https://github.com/ggerganov/whisper.cpp.git
cd whisper.cpp

2. Download the ggml weights for distil-small.en from the Hugging Face Hub:

python -c "from huggingface_hub import hf_hub_download; hf_hub_download(repo_id='distil-whisper/distil-small.en', filename='ggml-distil-small.en.bin', local_dir='./models')"

Note that if you do not have the huggingface_hub package installed, you can also download the weights with wget:

wget https://huggingface.co/distil-whisper/distil-small.en/resolve/main/ggml-distil-small.en.bin -P ./models

3. Run inference using the provided sample audio:

make -j && ./main -m models/ggml-distil-small.en.bin -f samples/jfk.wav

Transformers.js

Distil-Whisper can even run completely in your web browser with Transformers.js:

1. Install Transformers.js from NPM:

npm i @xenova/transformers

2. Import the library and perform inference with the pipeline API.

import { pipeline } from '@xenova/transformers';

const transcriber = await pipeline('automatic-speech-recognition', 'distil-whisper/distil-small.en');

const url = 'https://huggingface.co/datasets/Xenova/transformers.js-docs/resolve/main/jfk.wav';
const output = await transcriber(url);
// { text: " And so my fellow Americans, ask not what your country can do for you. Ask what you can do for your country." }

Check out the online Distil-Whisper Web demo to try it out yourself. As you'll see, it runs locally in your browser: no server required!

See the docs for more information.

Candle

Coming soon!

8bit & 4bit Quantization

Coming soon!

Model Details

Distil-Whisper inherits the encoder-decoder architecture from Whisper. The encoder maps a sequence of speech vector inputs to a sequence of hidden-state vectors. The decoder auto-regressively predicts text tokens, conditional on all previous tokens and the encoder hidden-states. Consequently, the encoder is only run forward once, whereas the decoder is run as many times as the number of tokens generated. In practice, this means the decoder accounts for over 90% of total inference time. Thus, to optimise for latency, the focus is on minimising the inference time of the decoder.

To distill the Whisper model, we reduce the number of decoder layers while keeping the encoder fixed. The encoder (shown in green) is entirely copied from the teacher to the student and frozen during training. The student's decoder consists of a subset of the teacher decoder layers, which are intialised from maximally spaced layers. The model is then trained on a weighted sum of the KL divergence and pseudo-label loss terms.

Evaluation

The following code-snippets demonstrates how to evaluate the Distil-Whisper model on the LibriSpeech validation.clean dataset with streaming mode, meaning no audio data has to be downloaded to your local device.

First, we need to install the required packages, including 🤗 Datasets to stream and load the audio data, and 🤗 Evaluate to perform the WER calculation:

pip install --upgrade pip
pip install --upgrade transformers datasets[audio] evaluate jiwer

Evaluation can then be run end-to-end with the following example:

from transformers import AutoModelForSpeechSeq2Seq, AutoProcessor
from transformers.models.whisper.english_normalizer import EnglishTextNormalizer
from datasets import load_dataset
from evaluate import load
import torch
from tqdm import tqdm

# define our torch configuration
device = "cuda:0" if torch.cuda.is_available() else "cpu"
torch_dtype = torch.float16 if torch.cuda.is_available() else torch.float32

model_id = "distil-whisper/distil-small.en"

# load the model + processor
model =  AutoModelForSpeechSeq2Seq.from_pretrained(model_id, torch_dtype=torch_dtype, use_safetensors=True, low_cpu_mem_usage=True)
model = model.to(device)
processor = AutoProcessor.from_pretrained(model_id)

# load the dataset with streaming mode
dataset = load_dataset("librispeech_asr", "clean", split="validation", streaming=True)

# define the evaluation metric
wer_metric = load("wer")
normalizer = EnglishTextNormalizer(processor.tokenizer.english_spelling_normalizer)

def inference(batch):
    # 1. Pre-process the audio data to log-mel spectrogram inputs
    audio = [sample["array"] for sample in batch["audio"]]
    input_features = processor(audio, sampling_rate=batch["audio"][0]["sampling_rate"], return_tensors="pt").input_features
    input_features = input_features.to(device, dtype=torch_dtype)
    
    # 2. Auto-regressively generate the predicted token ids
    pred_ids = model.generate(input_features, max_new_tokens=128)
    
    # 3. Decode the token ids to the final transcription
    batch["transcription"] = processor.batch_decode(pred_ids, skip_special_tokens=True)
    batch["reference"] = batch["text"]
    return batch

dataset = dataset.map(function=inference, batched=True, batch_size=16)

all_transcriptions = []
all_references = []

# iterate over the dataset and run inference
for i, result in tqdm(enumerate(dataset), desc="Evaluating..."):
    all_transcriptions.append(result["transcription"])
    all_references.append(result["reference"])

# normalize predictions and references
all_transcriptions = [normalizer(transcription) for transcription in all_transcriptions]
all_references = [normalizer(reference) for reference in all_references]

# compute the WER metric
wer = 100 * wer_metric.compute(predictions=all_transcriptions, references=all_references)
print(wer)

Print Output:

3.4326070294536297

Intended Use

Distil-Whisper is intended to be a drop-in replacement for Whisper on English speech recognition. In particular, it achieves comparable WER results over out-of-distribution test data, while being 6x faster over both short and long-form audio.

Data

Distil-Whisper is trained on 22,000 hours of audio data from 9 open-source, permissively licensed speech datasets on the Hugging Face Hub:

Dataset	Size / h	Speakers	Domain	Licence
People's Speech	12,000	unknown	Internet Archive	CC-BY-SA-4.0
Common Voice 13	3,000	unknown	Narrated Wikipedia	CC0-1.0
GigaSpeech	2,500	unknown	Audiobook, podcast, YouTube	apache-2.0
Fisher	1,960	11,900	Telephone conversations	LDC
LibriSpeech	960	2,480	Audiobooks	CC-BY-4.0
VoxPopuli	540	1,310	European Parliament	CC0
TED-LIUM	450	2,030	TED talks	CC-BY-NC-ND 3.0
SwitchBoard	260	540	Telephone conversations	LDC
AMI	100	unknown	Meetings	CC-BY-4.0
Total	21,770	18,260+

The combined dataset spans 10 distinct domains and over 50k speakers. The diversity of this dataset is crucial to ensuring the distilled model is robust to audio distributions and noise.

The audio data is then pseudo-labelled using the Whisper large-v2 model: we use Whisper to generate predictions for all the audio in our training set and use these as the target labels during training. Using pseudo-labels ensures that the transcriptions are consistently formatted across datasets and provides sequence-level distillation signal during training.

WER Filter

The Whisper pseudo-label predictions are subject to mis-transcriptions and hallucinations. To ensure we only train on accurate pseudo-labels, we employ a simple WER heuristic during training. First, we normalise the Whisper pseudo-labels and the ground truth labels provided by each dataset. We then compute the WER between these labels. If the WER exceeds a specified threshold, we discard the training example. Otherwise, we keep it for training.

Section 9.2 of the Distil-Whisper paper demonstrates the effectiveness of this filter for improving downstream performance of the distilled model. We also partially attribute Distil-Whisper's robustness to hallucinations to this filter.

Training

The model was trained for 50,000 optimisation steps (or 12 epochs) with batch size 2056. The Tensorboard training logs can be found under: https://huggingface.co/distil-whisper/distil-small.en/tensorboard?params=scalars#frame

Results

The distilled model performs to within 1% WER of Whisper on out-of-distribution (OOD) short-form audio, and outperforms Whisper by 0.1% on OOD long-form audio. This performance gain is attributed to lower hallucinations.

For a detailed per-dataset breakdown of the evaluation results, refer to Tables 16 and 17 of the Distil-Whisper paper

Distil-Whisper is also evaluated on the ESB benchmark datasets as part of the OpenASR leaderboard, where it performs to within 0.2% WER of Whisper.

Reproducing Distil-Whisper

Training and evaluation code to reproduce Distil-Whisper is available under the Distil-Whisper repository: https://github.com/huggingface/distil-whisper/tree/main/training

License

Distil-Whisper inherits the MIT license from OpenAI's Whisper model.

Citation

If you use this model, please consider citing the Distil-Whisper paper:

@misc{gandhi2023distilwhisper,
      title={Distil-Whisper: Robust Knowledge Distillation via Large-Scale Pseudo Labelling}, 
      author={Sanchit Gandhi and Patrick von Platen and Alexander M. Rush},
      year={2023},
      eprint={2311.00430},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}

Acknowledgements

• OpenAI for the Whisper model and original codebase
• Hugging Face 🤗 Transformers for the model integration
• Google's TPU Research Cloud (TRC) programme for Cloud TPU v4s
• @rsonavane for releasing an early iteration of Distil-Whisper on the LibriSpeech dataset