Distil-Whisper: distil-large-v3-fr

Distil-Whisper for English Automatic Speech Recognition (ASR) was proposed in the paper Robust Knowledge Distillation via Large-Scale Pseudo Labelling.

This is the knowledge distilled version of OpenAI's Whisper large-v3 for French ASR.

The result is a distilled model that performs within 2% WER of Whisper large-v3 on out-of-distribution evaluation sets for both short-form and long form transcription. Moreover, it is 5.9x faster than Whisper large-v3 and 1.3 times faster than the tiniest version of whisper while being incomparably more accurate.

Model	Params (M)	Rel. Latency	Short-Form WER	Long-Form WER
whisper-tiny	37.8	4.7	43.7	28.2
whisper-base	72.6	3.7	30.6	18.7
whisper-small	242	2.3	16.2	12.6
whisper-medium	764	1.3	11.7	11.0
whisper-large-v3	1540	1.0	7.8	9.0
distil-large-v3-fr	756	5.9	9.3	11.1

*latencies benchmarked to generate 128 tokens on A100 40GB with a batch size of 1. More details about inference performances in inference speed section.
*WERs are averaged on the test sets. More details in short-form and long-form results sections.

Table of Contents

1. Transformers Usage
   • Short-Form Transcription
   • Sequential Long-Form
   • Chunked Long-Form
   • Speculative Decoding
   • Additional Speed and Memory Improvements
2. Library Integrations
   • Whisper cpp
   • Transformers.js
3. Model Details
   • Architecture
   • Training
4. Results
   • Evaluation methodology
   • Short-Form
   • Long-Form
   • Inference Speed
5. License

Transformers Usage

distil-large-v3-fr is supported in the Hugging Face 🤗 Transformers library from version 4.41 onwards. To run the model, first install the latest version of Transformers. For this example, we'll also install 🤗 Datasets to load a 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 = "eustlb/distil-large-v3-fr"

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("google/fleurs", "fr_fr", split="train", streaming=True)
sample = next(iter(dataset))["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")

For segment-level timestamps, pass the argument return_timestamps=True and return the "chunks" output:

result = pipe(sample, return_timestamps=True)
print(result["chunks"])

For more control over the generation parameters, use the model + processor API directly:

Ad-hoc generation arguments can be passed to model.generate, including num_beams for beam-search, return_timestamps for segment-level timestamps, and prompt_ids for prompting. See the docstrings for more details.

import torch
from transformers import AutoModelForSpeechSeq2Seq, AutoProcessor
from datasets import Audio, 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 = "eustlb/distil-large-v3-fr"

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)

dataset = load_dataset("google/fleurs", "fr_fr", split="train", streaming=True)
dataset = dataset.cast_column("audio", Audio(processor.feature_extractor.sampling_rate))
sample = next(iter(dataset))["audio"] 

input_features = processor(
  sample["array"], sampling_rate=sample["sampling_rate"], return_tensors="pt"
).input_features

input_features = input_features.to(device, dtype=torch_dtype)

gen_kwargs = {
  "max_new_tokens": 128,
  "num_beams": 1,
  "return_timestamps": False,
}

pred_ids = model.generate(input_features, **gen_kwargs)
pred_text = processor.batch_decode(pred_ids, skip_special_tokens=True, decode_with_timestamps=gen_kwargs["return_timestamps"])

print(pred_text)

Sequential Long-Form

distil-large-v3 is compatible with OpenAI's sequential long-form transcription algorithm. This algorithm uses a sliding window for buffered inference of long audio files (> 30-seconds), and returns more accurate transcriptions compared to the chunked long-form algorithm.

The sequential long-form algorithm should be used in either of the following scenarios:

1. Transcription accuracy is the most important factor, and latency is less of a consideration
2. You are transcribing batches of long audio files, in which case the latency of sequential is comparable to chunked, while being up to 0.5% WER more accurate

If you are transcribing single long audio files and latency is the most important factor, you should use the chunked algorithm described below. For a detailed explanation of the different algorithms, refer to Sections 5 of the Distil-Whisper paper.

The pipeline class can be used to transcribe long audio files with the sequential algorithm 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 = "eustlb/distil-large-v3-fr"

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("eustlb/french-long-form-test", split="test", streaming=True)
sample = next(iter(dataset))["audio"]

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

For more control over the generation parameters, use the model + processor API directly:

import torch
from transformers import AutoModelForSpeechSeq2Seq, AutoProcessor
from datasets import Audio, 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 = "eustlb/distil-large-v3-fr"

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)

dataset = load_dataset("eustlb/french-long-form-test", split="test", streaming=True)
dataset = dataset.cast_column("audio", Audio(processor.feature_extractor.sampling_rate))
sample = next(iter(dataset))["audio"]

inputs = processor(
    sample["array"],
    sampling_rate=sample["sampling_rate"],
    return_tensors="pt",
    truncation=False,
    padding="longest",
    return_attention_mask=True,
)
inputs = inputs.to(device, dtype=torch_dtype)

gen_kwargs = {
    "max_new_tokens": 448,
    "num_beams": 1,
    "condition_on_prev_tokens": False,
    "compression_ratio_threshold": 1.35,  # zlib compression ratio threshold (in token space)
    "temperature": (0.0, 0.2, 0.4, 0.6, 0.8, 1.0),
    "logprob_threshold": -1.0,
    "no_speech_threshold": 0.6,
    "return_timestamps": True,
}

pred_ids = model.generate(**inputs, **gen_kwargs)
pred_text = processor.batch_decode(pred_ids, skip_special_tokens=True, decode_with_timestamps=False)

print(pred_text)

Chunked Long-Form

distil-large-v3-fr remains compatible with the Transformers chunked long-form algorithm. This algorithm should be used when a single large audio file is being transcribed and the fastest possible inference is required. In such circumstances, the chunked algorithm is up to 9x faster than OpenAI's sequential long-form implementation (see Table 7 of the Distil-Whisper paper).

To enable chunking, pass the chunk_length_s parameter to the pipeline. For distil-large-v3-fr, a chunk length of 25-seconds is optimal. To activate batching over long audio files, 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 = "eustlb/distil-large-v3-fr"

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=25,
    batch_size=16,
    torch_dtype=torch_dtype,
    device=device,
)

dataset = load_dataset("eustlb/french-long-form-test", split="test", streaming=True)
sample = next(iter(dataset))["audio"] 

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

Speculative Decoding

distil-large-v3 is the first Distil-Whisper model that can be used as an assistant to Whisper large-v3 for speculative decoding. Speculative decoding mathematically ensures that exactly the 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, AutoModelForCausalLM, 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 = "eustlb/distil-large-v3-fr"

assistant_model = AutoModelForCausalLM.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-large-v3"

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("google/fleurs", "fr_fr", split="train", streaming=True)
sample = next(iter(dataset))["audio"] 

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

For more details on speculative decoding, refer to the blog post Speculative Decoding for 2x Faster Whisper Inference.

Additional Speed & Memory Improvements

You can apply additional speed and memory improvements to Distil-Whisper to further reduce the inference speed and VRAM requirements. These optimisations primarily target the attention kernel, swapping it from an eager implementation to a more efficient flash attention version.

Flash Attention 2

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

Then pass attn_implementation="flash_attention_2" 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, attn_implementation="flash_attention_2")

Torch Scale-Product-Attention (SDPA)

If your GPU does not support Flash Attention, we recommend making use of PyTorch scaled dot-product attention (SDPA). This attention implementation is activated by default for PyTorch versions 2.1.1 or greater. To check whether you have a compatible PyTorch version, run the following Python code snippet:

from transformers.utils import is_torch_sdpa_available

print(is_torch_sdpa_available())

If the above returns True, you have a valid version of PyTorch installed and SDPA is activated by default. If it returns False, you need to upgrade your PyTorch version according to the official instructions

Once a valid PyTorch version is installed, SDPA is activated by default. It can also be set explicitly by specifying attn_implementation="sdpa" as follows:

- 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, attn_implementation="sdpa")

For more information about how to use the SDPA refer to the Transformers SDPA documentation.

Torch compile

Coming soon...

4-bit and 8-bit Inference

Coming soon...

Library Integrations

Whisper.cpp

distil-large-v3-fr can be run with the Whisper.cpp package with the original sequential long-form transcription algorithm. In a provisional benchmark on Mac M1, distil-large-v3 is over 5x faster than Whisper large-v3, while performing to within 0.8% 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. Install the Hugging Face Hub Python package:

pip install --upgrade huggingface_hub

And download the GGML weights for distil-large-v3 using the following Python snippet:

from huggingface_hub import hf_hub_download

hf_hub_download(repo_id='eustlb/distil-large-v3-fr-ggml', filename='ggml-distil-large-v3-fr.bin', local_dir='./models')

Note that if you do not have a Python environment set-up, you can also download the weights directly with wget:

wget https://huggingface.co/eustlb/distil-large-v3-fr-ggml/resolve/main/ggml-distil-large-v3-fr.bin -P ./models

3. Run inference

wget https://huggingface.co/spaces/eustlb/whisper-vs-distil-whisper-fr/resolve/main/assets/example_1.wav
make -j && ./main -m models/ggml-distil-large-v3-fr.bin -f example_1.wav

Transformers.js

Distil-Whisper can be 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', 'eustlb/distil-large-v3-fr');

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." }

Refer to the Transformers.js docs for further information.

Model Details

Architecture

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 structure is copied from Whisper large-v3, with the only difference being a reduction from 32 to 2 decoder layers. These layers are initialized from distil-large-v3 to leverage language transfer from English to French (more details here).

Training

Data

distil-large-v3-fr is trained on 4,515 hours of audio data from three open-source, permissively licensed speech datasets on the Hugging Face Hub:

Dataset	Size / h	Speakers	Domain	Licence
Common Voice 17	1,014	unknown	Narrated Wikipedia	CC0-1.0
MultiLingual LibriSpeech	1,077	142	Audiobook	CC-BY-4.0
YODAS fr000 split	2,424	unknown	YouTube	CC-BY-3.0
Total	4,515	142+

The audio data is then pseudo-labelled using the Whisper large-v3 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.

We chose for this training a WER threshold of 20%, resulting in an effective training set of 2110 hours (750 for Common Voice 17, 1040 for MultiLingual LibriSpeech and 320 for YODAS fr000 split).

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 procedure

The model was trained for 18,000 optimisation steps (or 14 epochs) with batch size 256. We saved the best model, based on the global wer score on validation splits, reached after 14,000 optimization steps (or 11 epochs). See the Distil-Whisper paper for more details (training objective, etc).

Results

The distilled model performs to within 1% WER of Whisper large-v3 on out-of-distribution (Voxpopuli, Fleurs) short-form audio and within 2.5% WER on out-of-distribuion sequential long-form decoding.

Evaluation methodology

The model has been tested for both in-distribution (Common Voice 17 and Multilingual Librispeech) and out-of-distribution (Fleurs, Voxpopuli, custom long-form test set) short-form and long-form transcription performances. Models have been evaluated with SDPA, float32 and batch size 32.

Short-form evaluations are conducted on the four given datasets by first applying a filter to exclude samples longer than 30 seconds.

Long-form evaluation is conducted on a custom out-of-distribution long-form test set using OpenAI's sequential long-form transcription algorithm (see Sequential Long-Form section) with long-form generation parameters that can be found here.

Short-Form

Model	Common Voice 17	Multilingual Librispeech	voxpopuli	fleurs	RTFx
whisper-tiny	57.141	38.049	32.346	47.4	265.226
whisper-base	42.58	25.235	26.701	27.773	237.195
whisper-small	22.56	13.576	14.486	14.165	196.932
whisper-medium	15.51	9.541	11.836	9.992	93.428
whisper-large-v3	11.038	4.762	9.83	5.624	62.845
distil-large-v3-fr	12.675	5.865	10.832	7.989	106.291

*the above datasets correspond to test splits

*RTFx = 1 / RTF, where RTF is the Real Time Factor. To be interpreted as audio processed (in seconds) per second of processing.

Long-Form

Model Name	RTFx	long-form test set
whisper-tiny	121.389	28.158
whisper-base	109.366	18.665
whisper-small	83.049	12.557
whisper-medium	47.807	11.023
whisper-large-v3	38.294	9.008
distil-large-v3-fr	101.326	11.13

Inference speed

Reported latencies were benchmarked on a 40GB nvidia A100, generating 128 tokens with SDPA, bfloat16, 3 warmup steps, 5 measures, one beam. The benchmarking script can be found here. The time measured is the time do one forward pass of the encoder and 128 autoregressive forward passes of the decoder.

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-large-v3-fr 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, in particular Jong Wook Kim for the original codebase and training discussions
• Hugging Face 🤗 Transformers for the model integration
• Georgi Gerganov for the Whisper cpp integration
• Joshua Lochner for the Transformers.js integration
• Vaibhav Srivastav for Distil-Whisper distribution
• Raghav Sonavane for an early iteration of Distil-Whisper on the LibriSpeech datasets