• Stars
    star
    2,964
  • Rank 14,712 (Top 0.3 %)
  • Language
    Python
  • License
    MIT License
  • Created 7 months ago
  • Updated 3 months ago

Reviews

There are no reviews yet. Be the first to send feedback to the community and the maintainers!

Repository Details

Distilled variant of Whisper for speech recognition. 6x faster, 50% smaller, within 1% word error rate.

Distil-Whisper

[Paper] [Models] [Colab] [Training Code]

Distil-Whisper is a distilled version of Whisper that is 6 times faster, 49% smaller, and performs within 1% word error rate (WER) on out-of-distribution evaluation sets:

Model Params / M Rel. Latency Short-Form WER Long-Form WER
whisper-large-v2 1550 1.0 9.1 11.7
distil-large-v2 756 5.8 10.1 11.6
distil-medium.en 394 6.8 11.1 12.4

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 soon update the repository with multilingual checkpoints when ready!

1. Usage

Distil-Whisper is supported in Hugging Face ๐Ÿค— 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 ๐Ÿค— 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

Short-form transcription is the process of transcribing audio samples that are less than 30-seconds long, which is the maximum receptive field of the Whisper models. This means the entire audio clip can be processed in one go without the need for chunking.

First, we load Distil-Whisper via the convenient AutoModelForSpeechSeq2Seq and AutoProcessor classes.

We load the model in float16 precision and make sure that loading time takes as little time as possible by passing low_cpu_mem_usage=True. In addition, we want to make sure that the model is loaded in safetensors format by passing use_safetensors=True:

import torch
from transformers import AutoModelForSpeechSeq2Seq, AutoProcessor, pipeline

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-large-v2"

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)

The model and processor can then be passed to the pipeline. Note that if you would like to have more control over the generation process, you can directly make use of model.generate(...) as shown here.

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,
)

Next, we load an example short-form audio from the LibriSpeech corpus:

from datasets import load_dataset

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

Finally, we can call the pipeline to transcribe the 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("audio.mp3")
print(result["text"])

For more information on how to customize the automatic speech recognition pipeline, please refer to the ASR pipeline docs. We also provide an end-to-end Google Colab that benchmarks Whisper against Distil-Whisper.

Long-Form Transcription

Distil-Whisper uses a chunked algorithm to transcribe long-form audio files longer than 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).

We can load the model and processor as before:

import torch
from transformers import AutoModelForSpeechSeq2Seq, AutoProcessor, pipeline

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-large-v2"

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)

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:

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,
)

The argument max_new_tokens controls the maximum number of generated tokens per-chunk. In the typical speech setting, we have no more than 3 words spoken per-second. Therefore, for a 15-second input, we have at most 45 words (approx 60 tokens). We set the maximum number of generated tokens per-chunk to 128 to truncate any possible hallucinations that occur at the end of the segment. These tokens get removed at the chunk borders using the long-form chunking transcription algorithm, so it is more efficient to truncate them directly during generation to avoid redundant generation steps in the decoder.

Now, let's load a long-form audio sample. Here, we use an example of concatenated samples from the LibriSpeech corpus:

from datasets import load_dataset

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

Finally, we can call the pipeline to transcribe the audio:

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

For more information on how to customize the automatic speech recognition pipeline, please refer to the ASR pipeline docs.

Speculative Decoding

Distil-Whisper can be used as an assistant model to Whisper for speculative decoding. As a refresher, we recommend reading Joao's amazing blog post or taking a look at the original paper.

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.

For speculative decoding, we need to load both the teacher: openai/whisper-large-v2. As well as the assistant (a.k.a student) distil-whisper/distil-large-v2.

Let's start by loading the teacher model and processor. We do this in much the same way we loaded the Distil-Whisper model in the previous examples:

from transformers import AutoModelForSpeechSeq2Seq, AutoProcessor
import torch

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

model_id = "openai/whisper-large-v2"

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)

Now let's load the assistant. Since Distil-Whisper shares exactly same encoder as the teacher model, we only need to load the 2-layer decoder as a "Decoder-only" model:

from transformers import AutoModelForCausalLM
assistant_model_id = "distil-whisper/distil-large-v2"

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

The assistant model shares the same processor as the teacher, so there's no need to load a student processor.

We can now pass the assistant model to the pipeline to be used for speculative decoding. We pass it as a generate_kwarg with the key "assistant_model" so that speculative decoding is enabled:

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,
)

As before, we can pass any sample to the pipeline to be transcribed:

from datasets import load_dataset

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

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

Note: speculative decoding should be on average 2x faster than using "only" Whisper large-v2 at a mere 8% increase in VRAM memory usage while mathematically ensuring the same results. This makes it the perfect replacement for Whisper large-v2 in existing speech recognition pipelines.

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

You can then pass use_flash_attention_2=True to from_pretrained to enable Flash Attention 2:

- 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()

Exporting to Other Libraries

Distil-Whisper has support in the following libraries with the original "sequential" long-form transcription algorithm. Click the links in the table to see the relevant code-snippets for each:

Library distil-medium.en distil-large-v2
OpenAI Whisper link link
Whisper cpp link link
Transformers js link link
Candle (Rust) link link

Updates will be posted here with the integration of the "chunked" long-form transcription algorithm into the respective libraries.

For the ๐Ÿค— Transformers code-examples, refer to the sections Short-Form and Long-Form Transcription.

2. Why use Distil-Whisper? โ‰๏ธ

Distil-Whisper is designed to be a drop-in replacement for Whisper on English speech recognition. Here are 5 reasons for making the switch to Distil-Whisper:

  1. Faster inference: 6 times faster inference speed, while performing to within 1% WER of Whisper on out-of-distribution audio:

  1. Robustness to noise: demonstrated by strong WER performance at low signal-to-noise ratios:

  1. Robustness to hallucinations: quantified by 1.3 times fewer repeated 5-gram word duplicates (5-Dup.) and 2.1% lower insertion error rate (IER) than Whisper:

  1. Designed for speculative decoding: Distil-Whisper can be used as an assistant model to Whisper, giving 2 times faster inference speed while mathematically ensuring the same outputs as the Whisper model.
  2. Permissive license: Distil-Whisper is MIT licensed, meaning it can be used for commercial applications.

3. Approach โœ๏ธ

To distill Whisper, we copy the entire encoder module and freeze it during training. We copy only two decoder layers, which are initialised from the first and last decoder layers from Whisper. All other decoder layers from Whisper are discarded:

Distil-Whisper is trained on a knowledge distillation objective. Specifically, it is trained to minimise the KL divergence between the distilled model and the Whisper model, as well as the cross-entropy loss on pseudo-labelled audio data.

We train Distil-Whisper on a total of 22k hours of pseudo-labelled audio data, spanning 10 domains with over 18k speakers:

This diverse audio dataset is paramount to ensuring robustness of Distil-Whisper to different datasets and domains.

In addition, we use a WER filter to discard pseudo-labels where Whisper mis-transcribes or hallucinates. This greatly improves WER performance of the downstream distilled model.

For full details on the distillation set-up and evaluation results, refer to the Distil-Whisper paper.

4. Training Code

Training code to reproduce Distil-Whisper can be found in the directory training. This code has been adapted be general enough to distill Whisper for multilingual speech recognition, facilitating anyone in the community to distill Whisper on their choice of language.

5. Acknowledgements

6. 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}
}

And also the Whisper paper:

@misc{radford2022robust,
      title={Robust Speech Recognition via Large-Scale Weak Supervision}, 
      author={Alec Radford and Jong Wook Kim and Tao Xu and Greg Brockman and Christine McLeavey and Ilya Sutskever},
      year={2022},
      eprint={2212.04356},
      archivePrefix={arXiv},
      primaryClass={eess.AS}
}

More Repositories

1

transformers

๐Ÿค— Transformers: State-of-the-art Machine Learning for Pytorch, TensorFlow, and JAX.
Python
125,891
star
2

pytorch-image-models

PyTorch image models, scripts, pretrained weights -- ResNet, ResNeXT, EfficientNet, NFNet, Vision Transformer (ViT), MobileNet-V3/V2, RegNet, DPN, CSPNet, Swin Transformer, MaxViT, CoAtNet, ConvNeXt, and more
Python
28,073
star
3

diffusers

๐Ÿค— Diffusers: State-of-the-art diffusion models for image and audio generation in PyTorch and FLAX.
Python
22,776
star
4

datasets

๐Ÿค— The largest hub of ready-to-use datasets for ML models with fast, easy-to-use and efficient data manipulation tools
Python
17,530
star
5

peft

๐Ÿค— PEFT: State-of-the-art Parameter-Efficient Fine-Tuning.
Python
14,007
star
6

candle

Minimalist ML framework for Rust
Rust
12,686
star
7

tokenizers

๐Ÿ’ฅ Fast State-of-the-Art Tokenizers optimized for Research and Production
Rust
8,286
star
8

trl

Train transformer language models with reinforcement learning.
Python
8,181
star
9

text-generation-inference

Large Language Model Text Generation Inference
Python
7,240
star
10

accelerate

๐Ÿš€ A simple way to launch, train, and use PyTorch models on almost any device and distributed configuration, automatic mixed precision (including fp8), and easy-to-configure FSDP and DeepSpeed support
Python
7,008
star
11

chat-ui

Open source codebase powering the HuggingChat app
TypeScript
6,369
star
12

deep-rl-class

This repo contains the syllabus of the Hugging Face Deep Reinforcement Learning Course.
MDX
3,541
star
13

alignment-handbook

Robust recipes to align language models with human and AI preferences
Python
3,485
star
14

autotrain-advanced

๐Ÿค— AutoTrain Advanced
Python
3,283
star
15

diffusion-models-class

Materials for the Hugging Face Diffusion Models Course
Jupyter Notebook
3,126
star
16

notebooks

Notebooks using the Hugging Face libraries ๐Ÿค—
Jupyter Notebook
3,114
star
17

neuralcoref

โœจFast Coreference Resolution in spaCy with Neural Networks
C
2,806
star
18

knockknock

๐ŸšชโœŠKnock Knock: Get notified when your training ends with only two additional lines of code
Python
2,682
star
19

swift-coreml-diffusers

Swift app demonstrating Core ML Stable Diffusion
Swift
2,377
star
20

safetensors

Simple, safe way to store and distribute tensors
Python
2,347
star
21

optimum

๐Ÿš€ Accelerate training and inference of ๐Ÿค— Transformers and ๐Ÿค— Diffusers with easy to use hardware optimization tools
Python
2,086
star
22

awesome-papers

Papers & presentation materials from Hugging Face's internal science day
1,996
star
23

blog

Public repo for HF blog posts
Jupyter Notebook
1,962
star
24

setfit

Efficient few-shot learning with Sentence Transformers
Jupyter Notebook
1,912
star
25

text-embeddings-inference

A blazing fast inference solution for text embeddings models
Rust
1,845
star
26

course

The Hugging Face course on Transformers
MDX
1,832
star
27

evaluate

๐Ÿค— Evaluate: A library for easily evaluating machine learning models and datasets.
Python
1,825
star
28

transfer-learning-conv-ai

๐Ÿฆ„ State-of-the-Art Conversational AI with Transfer Learning
Python
1,654
star
29

swift-coreml-transformers

Swift Core ML 3 implementations of GPT-2, DistilGPT-2, BERT, and DistilBERT for Question answering. Other Transformers coming soon!
Swift
1,543
star
30

pytorch-openai-transformer-lm

๐ŸฅA PyTorch implementation of OpenAI's finetuned transformer language model with a script to import the weights pre-trained by OpenAI
Python
1,464
star
31

cookbook

Open-source AI cookbook
Jupyter Notebook
1,357
star
32

huggingface_hub

All the open source things related to the Hugging Face Hub.
Python
1,311
star
33

Mongoku

๐Ÿ”ฅThe Web-scale GUI for MongoDB
TypeScript
1,289
star
34

huggingface.js

Utilities to use the Hugging Face Hub API
TypeScript
1,193
star
35

hmtl

๐ŸŒŠHMTL: Hierarchical Multi-Task Learning - A State-of-the-Art neural network model for several NLP tasks based on PyTorch and AllenNLP
Python
1,185
star
36

gsplat.js

JavaScript Gaussian Splatting library.
TypeScript
1,114
star
37

llm-vscode

LLM powered development for VSCode
TypeScript
1,060
star
38

datatrove

Freeing data processing from scripting madness by providing a set of platform-agnostic customizable pipeline processing blocks.
Python
1,033
star
39

pytorch-pretrained-BigGAN

๐Ÿฆ‹A PyTorch implementation of BigGAN with pretrained weights and conversion scripts.
Python
986
star
40

torchMoji

๐Ÿ˜‡A pyTorch implementation of the DeepMoji model: state-of-the-art deep learning model for analyzing sentiment, emotion, sarcasm etc
Python
880
star
41

nanotron

Minimalistic large language model 3D-parallelism training
Python
810
star
42

naacl_transfer_learning_tutorial

Repository of code for the tutorial on Transfer Learning in NLP held at NAACL 2019 in Minneapolis, MN, USA
Python
718
star
43

awesome-huggingface

๐Ÿค— A list of wonderful open-source projects & applications integrated with Hugging Face libraries.
698
star
44

optimum-nvidia

Python
680
star
45

dataset-viewer

Lightweight web API for visualizing and exploring any dataset - computer vision, speech, text, and tabular - stored on the Hugging Face Hub
Python
614
star
46

transformers-bloom-inference

Fast Inference Solutions for BLOOM
Python
546
star
47

exporters

Export Hugging Face models to Core ML and TensorFlow Lite
Python
540
star
48

pytorch_block_sparse

Fast Block Sparse Matrices for Pytorch
C++
523
star
49

llm.nvim

LLM powered development for Neovim
Lua
507
star
50

swift-transformers

Swift Package to implement a transformers-like API in Swift
Swift
482
star
51

node-question-answering

Fast and production-ready question answering in Node.js
TypeScript
459
star
52

large_language_model_training_playbook

An open collection of implementation tips, tricks and resources for training large language models
Python
431
star
53

llm-ls

LSP server leveraging LLMs for code completion (and more?)
Rust
416
star
54

llm_training_handbook

An open collection of methodologies to help with successful training of large language models.
Python
385
star
55

swift-chat

Mac app to demonstrate swift-transformers
Swift
375
star
56

tflite-android-transformers

DistilBERT / GPT-2 for on-device inference thanks to TensorFlow Lite with Android demo apps
Java
368
star
57

community-events

Place where folks can contribute to ๐Ÿค— community events
Jupyter Notebook
368
star
58

nn_pruning

Prune a model while finetuning or training.
Jupyter Notebook
360
star
59

text-clustering

Easily embed, cluster and semantically label text datasets
Python
335
star
60

speechbox

Python
328
star
61

100-times-faster-nlp

๐Ÿš€100 Times Faster Natural Language Processing in Python - iPython notebook
HTML
325
star
62

education-toolkit

Educational materials for universities
Jupyter Notebook
307
star
63

controlnet_aux

Python
306
star
64

optimum-intel

๐Ÿค— Optimum Intel: Accelerate inference with Intel optimization tools
Jupyter Notebook
295
star
65

datablations

Scaling Data-Constrained Language Models
Jupyter Notebook
293
star
66

unity-api

C#
284
star
67

open-muse

Open reproduction of MUSE for fast text2image generation.
Python
284
star
68

audio-transformers-course

The Hugging Face Course on Transformers for Audio
MDX
247
star
69

hub-docs

Docs of the Hugging Face Hub
221
star
70

lighteval

LightEval is a lightweight LLM evaluation suite that Hugging Face has been using internally with the recently released LLM data processing library datatrove and LLM training library nanotron.
Python
208
star
71

quanto

A pytorch Quantization Toolkit
Python
201
star
72

simulate

๐ŸŽข Creating and sharing simulation environments for embodied and synthetic data research
Python
185
star
73

ratchet

A cross-platform browser ML framework.
Rust
184
star
74

optimum-benchmark

A unified multi-backend utility for benchmarking Transformers, Timm, Diffusers and Sentence-Transformers with full support of Optimum's hardware optimizations & quantization schemes.
Python
183
star
75

hf_transfer

Rust
181
star
76

olm-datasets

Pipeline for pulling and processing online language model pretraining data from the web
Python
169
star
77

instruction-tuned-sd

Code for instruction-tuning Stable Diffusion.
Python
167
star
78

optimum-neuron

Easy, fast and very cheap training and inference on AWS Trainium and Inferentia chips.
Jupyter Notebook
163
star
79

llm-swarm

Manage scalable open LLM inference endpoints in Slurm clusters
Python
156
star
80

OBELICS

Code used for the creation of OBELICS, an open, massive and curated collection of interleaved image-text web documents, containing 141M documents, 115B text tokens and 353M images.
Python
147
star
81

workshops

Materials for workshops on the Hugging Face ecosystem
Jupyter Notebook
146
star
82

cosmopedia

Python
138
star
83

api-inference-community

Python
131
star
84

diffusion-fast

Faster generation with text-to-image diffusion models.
Python
127
star
85

diarizers

Python
106
star
86

optimum-habana

Easy and lightning fast training of ๐Ÿค— Transformers on Habana Gaudi processor (HPU)
Python
106
star
87

sharp-transformers

A Unity plugin for using Transformers models in Unity.
C#
104
star
88

competitions

Python
101
star
89

hf-hub

Rust client for the huggingface hub aiming for minimal subset of features over `huggingface-hub` python package
Rust
93
star
90

olm-training

Repo for training MLMs, CLMs, or T5-type models on the OLM pretraining data, but it should work with any hugging face text dataset.
Python
87
star
91

fuego

[WIP] A ๐Ÿ”ฅ interface for running code in the cloud
Python
84
star
92

tune

Python
83
star
93

datasets-viewer

Viewer for the ๐Ÿค— datasets library.
Python
82
star
94

optimum-graphcore

Blazing fast training of ๐Ÿค— Transformers on Graphcore IPUs
Python
78
star
95

frp

FRP Fork
Go
73
star
96

paper-style-guide

72
star
97

block_movement_pruning

Block Sparse movement pruning
Python
70
star
98

amused

Python
68
star
99

doc-builder

The package used to build the documentation of our Hugging Face repos
Python
67
star
100

data-measurements-tool

Developing tools to automatically analyze datasets
Python
67
star