File size: 18,253 Bytes
e4cfdcc
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
---
license: mit
language: fr
library_name: transformers
pipeline_tag: automatic-speech-recognition
thumbnail: null
tags:
- automatic-speech-recognition
- hf-asr-leaderboard
datasets:
- mozilla-foundation/common_voice_13_0
- facebook/multilingual_librispeech
- facebook/voxpopuli
- google/fleurs
- gigant/african_accented_french
metrics:
- wer
model-index:
- name: whisper-large-v3-french-distil-dec8
  results:
  - task:
      name: Automatic Speech Recognition
      type: automatic-speech-recognition
    dataset:
      name: Common Voice 13.0
      type: mozilla-foundation/common_voice_13_0
      config: fr
      split: test
      args:
        language: fr
    metrics:
    - name: WER
      type: wer
      value: 7.62
  - task:
      name: Automatic Speech Recognition
      type: automatic-speech-recognition
    dataset:
      name: Multilingual LibriSpeech (MLS)
      type: facebook/multilingual_librispeech
      config: french
      split: test
      args:
        language: fr
    metrics:
    - name: WER
      type: wer
      value: 3.80
  - task:
      name: Automatic Speech Recognition
      type: automatic-speech-recognition
    dataset:
      name: VoxPopuli
      type: facebook/voxpopuli
      config: fr
      split: test
      args:
        language: fr
    metrics:
    - name: WER
      type: wer
      value: 8.85
  - task:
      name: Automatic Speech Recognition
      type: automatic-speech-recognition
    dataset:
      name: Fleurs
      type: google/fleurs
      config: fr_fr
      split: test
      args:
        language: fr
    metrics:
    - name: WER
      type: wer
      value: 5.40
  - task:
      name: Automatic Speech Recognition
      type: automatic-speech-recognition
    dataset:
      name: African Accented French
      type: gigant/african_accented_french
      config: fr
      split: test
      args:
        language: fr
    metrics:
    - name: WER
      type: wer
      value: 4.18
---

# Whisper-Large-V3-French-Distil-Dec8

Whisper-Large-V3-French-Distil represents a series of distilled versions of [Whisper-Large-V3-French](https://huggingface.co/bofenghuang/whisper-large-v3-french), achieved by reducing the number of decoder layers from 32 to 16, 8, 4, or 2 and distilling using a large-scale dataset, as outlined in this [paper](https://arxiv.org/abs/2311.00430).

The distilled variants reduce memory usage and inference time while maintaining performance (based on the retained number of layers) and mitigating the risk of hallucinations, particularly in long-form transcriptions. Moreover, they can be seamlessly combined with the original Whisper-Large-V3-French model for speculative decoding, resulting in improved inference speed and consistent outputs compared to using the standalone model.

This model has been converted into various formats, facilitating its usage across different libraries, including transformers, openai-whisper, fasterwhisper, whisper.cpp, candle, mlx, etc.

## Table of Contents

- [Performance](#performance)
- [Usage](#usage)
    - [Hugging Face Pipeline](#hugging-face-pipeline)
    - [Hugging Face Low-level APIs](#hugging-face-low-level-apis)
    - [Speculative Decoding](#speculative-decoding)
    - [OpenAI Whisper](#openai-whisper)
    - [Faster Whisper](#faster-whisper)
    - [Whisper.cpp](#whispercpp)
    - [Candle](#candle)
    - [MLX](#mlx)
- [Training details](#training-details)
- [Acknowledgements](#acknowledgements)

## Performance

We evaluated our model on both short and long-form transcriptions, and also tested it on both in-distribution and out-of-distribution datasets to conduct a comprehensive analysis assessing its accuracy, generalizability, and robustness.

Please note that the reported WER is the result after converting numbers to text, removing punctuation (except for apostrophes and hyphens), and converting all characters to lowercase.

All evaluation results on the public datasets can be found [here](https://drive.google.com/drive/folders/1rFIh6yXRVa9RZ0ieZoKiThFZgQ4STPPI?usp=drive_link).

### Short-Form Transcription

![eval-short-form](https://huggingface.co/bofenghuang/whisper-large-v3-french/resolve/main/assets/whisper_fr_eval_short_form.png)

Due to the lack of readily available out-of-domain (OOD) and long-form test sets in French, we evaluated using internal test sets from [Zaion Lab](https://zaion.ai/). These sets comprise human-annotated audio-transcription pairs from call center conversations, which are notable for their significant background noise and domain-specific terminology.

### Long-Form Transcription

![eval-long-form](https://huggingface.co/bofenghuang/whisper-large-v3-french/resolve/main/assets/whisper_fr_eval_long_form.png)

The long-form transcription was run using the 🤗 Hugging Face pipeline for quicker evaluation. Audio files were segmented into 30-second chunks and processed in parallel.

## Usage

### Hugging Face Pipeline

The model can easily used with the 🤗 Hugging Face [`pipeline`](https://huggingface.co/docs/transformers/main_classes/pipelines#transformers.AutomaticSpeechRecognitionPipeline) class for audio transcription.

For long-form transcription (> 30 seconds), you can activate the process by passing the `chunk_length_s` argument. This approach segments the audio into smaller segments, processes them in parallel, and then joins them at the strides by finding the longest common sequence. While this chunked long-form approach may have a slight compromise in performance compared to OpenAI's sequential algorithm, it provides 9x faster inference speed.

```python
import torch
from datasets import load_dataset
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

# Load model
model_name_or_path = "bofenghuang/whisper-large-v3-french-distil-dec8"
processor = AutoProcessor.from_pretrained(model_name_or_path)
model = AutoModelForSpeechSeq2Seq.from_pretrained(
    model_name_or_path,
    torch_dtype=torch_dtype,
    low_cpu_mem_usage=True,
)
model.to(device)

# Init pipeline
pipe = pipeline(
    "automatic-speech-recognition",
    model=model,
    feature_extractor=processor.feature_extractor,
    tokenizer=processor.tokenizer,
    torch_dtype=torch_dtype,
    device=device,
    # chunk_length_s=30,  # for long-form transcription
    max_new_tokens=128,
)

# Example audio
dataset = load_dataset("bofenghuang/asr-dummy", "fr", split="test")
sample = dataset[0]["audio"]

# Run pipeline
result = pipe(sample)
print(result["text"])
```

### Hugging Face Low-level APIs

You can also use the 🤗 Hugging Face low-level APIs for transcription, offering greater control over the process, as demonstrated below:

```python
import torch
from datasets import load_dataset
from transformers import AutoModelForSpeechSeq2Seq, AutoProcessor

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

# Load model
model_name_or_path = "bofenghuang/whisper-large-v3-french-distil-dec8"
processor = AutoProcessor.from_pretrained(model_name_or_path)
model = AutoModelForSpeechSeq2Seq.from_pretrained(
    model_name_or_path,
    torch_dtype=torch_dtype,
    low_cpu_mem_usage=True,
)
model.to(device)

# Example audio
dataset = load_dataset("bofenghuang/asr-dummy", "fr", split="test")
sample = dataset[0]["audio"]

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


# Generate tokens
predicted_ids = model.generate(
    input_features.to(dtype=torch_dtype).to(device), max_new_tokens=128
)

# Detokenize to text
transcription = processor.batch_decode(predicted_ids, skip_special_tokens=True)[0]
print(transcription)
```

### Speculative Decoding

[Speculative decoding](https://huggingface.co/blog/whisper-speculative-decoding) can be achieved using a draft model, essentially a distilled version of Whisper. This approach guarantees identical outputs to using the main Whisper model alone, offers a 2x faster inference speed, and incurs only a slight increase in memory overhead. 

Since the distilled Whisper has the same encoder as the original, only its decoder need to be loaded, and encoder outputs are shared between the main and draft models during inference.

Using speculative decoding with the Hugging Face pipeline is simple - just specify the `assistant_model` within the generation configurations.

```python
import torch
from datasets import load_dataset
from transformers import (
    AutoModelForCausalLM,
    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

# Load model
model_name_or_path = "bofenghuang/whisper-large-v3-french"
processor = AutoProcessor.from_pretrained(model_name_or_path)
model = AutoModelForSpeechSeq2Seq.from_pretrained(
    model_name_or_path,
    torch_dtype=torch_dtype,
    low_cpu_mem_usage=True,
)
model.to(device)

# Load draft model
assistant_model_name_or_path = "bofenghuang/whisper-large-v3-french-distil-dec2"
assistant_model = AutoModelForCausalLM.from_pretrained(
    assistant_model_name_or_path,
    torch_dtype=torch_dtype,
    low_cpu_mem_usage=True,
)
assistant_model.to(device)

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

# Example audio
dataset = load_dataset("bofenghuang/asr-dummy", "fr", split="test")
sample = dataset[0]["audio"]

# Run pipeline
result = pipe(sample)
print(result["text"])
```

### OpenAI Whisper

You can also employ the sequential long-form decoding algorithm with a sliding window and temperature fallback, as outlined by OpenAI in their original [paper](https://arxiv.org/abs/2212.04356).

First, install the [openai-whisper](https://github.com/openai/whisper) package:

```bash
pip install -U openai-whisper
```

Then, download the converted model:

```bash
python -c "from huggingface_hub import hf_hub_download; hf_hub_download(repo_id='bofenghuang/whisper-large-v3-french-distil-dec8', filename='original_model.pt', local_dir='./models/whisper-large-v3-french-distil-dec8')"
```

Now, you can transcirbe audio files by following the usage instructions provided in the repository:

```python
import whisper
from datasets import load_dataset

# Load model
model = whisper.load_model("./models/whisper-large-v3-french-distil-dec8/original_model.pt")

# Example audio
dataset = load_dataset("bofenghuang/asr-dummy", "fr", split="test")
sample = dataset[0]["audio"]["array"].astype("float32")

# Transcribe
result = model.transcribe(sample, language="fr")
print(result["text"])
```

### Faster Whisper

Faster Whisper is a reimplementation of OpenAI's Whisper models and the sequential long-form decoding algorithm in the [CTranslate2](https://github.com/OpenNMT/CTranslate2) format.

Compared to openai-whisper, it offers up to 4x faster inference speed, while consuming less memory. Additionally, the model can be quantized into int8, further enhancing its efficiency on both CPU and GPU.

First, install the [faster-whisper](https://github.com/SYSTRAN/faster-whisper) package:

```bash
pip install faster-whisper
```

Then, download the model converted to the CTranslate2 format:

```bash
python -c "from huggingface_hub import snapshot_download; snapshot_download(repo_id='bofenghuang/whisper-large-v3-french-distil-dec8', local_dir='./models/whisper-large-v3-french-distil-dec8', allow_patterns='ctranslate2/*')"
```

Now, you can transcirbe audio files by following the usage instructions provided in the repository:

```python
from datasets import load_dataset
from faster_whisper import WhisperModel

# Load model
model = WhisperModel("./models/whisper-large-v3-french-distil-dec8/ctranslate2", device="cuda", compute_type="float16")  # Run on GPU with FP16

# Example audio
dataset = load_dataset("bofenghuang/asr-dummy", "fr", split="test")
sample = dataset[0]["audio"]["array"].astype("float32")

segments, info = model.transcribe(sample, beam_size=5, language="fr")

for segment in segments:
    print("[%.2fs -> %.2fs] %s" % (segment.start, segment.end, segment.text))
```

### Whisper.cpp

Whisper.cpp is a reimplementation of OpenAI's Whisper models, crafted in plain C/C++ without any dependencies. It offers compatibility with various backends and platforms.

Additionally, the model can be quantized to either 4-bit or 5-bit integers, further enhancing its efficiency.

First, clone and build the [whisper.cpp](https://github.com/ggerganov/whisper.cpp) repository:

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

# build the main example
make
```

Next, download the converted ggml weights from the Hugging Face Hub:

```bash
# Download model quantized with Q5_0 method
python -c "from huggingface_hub import hf_hub_download; hf_hub_download(repo_id='bofenghuang/whisper-large-v3-french-distil-dec8', filename='ggml-model-q5_0.bin', local_dir='./models/whisper-large-v3-french-distil-dec8')"
```

Now, you can transcribe an audio file using the following command:

```bash
./main -m ./models/whisper-large-v3-french-distil-dec8/ggml-model-q5_0.bin -l fr -f /path/to/audio/file --print-colors
```

### Candle

[Candle-whisper](https://github.com/huggingface/candle/tree/main/candle-examples/examples/whisper) is a reimplementation of OpenAI's Whisper models in the candle format - a lightweight ML framework built in Rust.

First, clone the [candle](https://github.com/huggingface/candle) repository:

```bash
git clone https://github.com/huggingface/candle.git
cd candle/candle-examples/examples/whisper
```

Transcribe an audio file using the following command:

```bash
cargo run --example whisper --release -- --model large-v3 --model-id bofenghuang/whisper-large-v3-french-distil-dec8 --language fr --input /path/to/audio/file
```

In order to use CUDA add `--features cuda` to the example command line:

```bash
cargo run --example whisper --release --features cuda -- --model large-v3 --model-id bofenghuang/whisper-large-v3-french-distil-dec8 --language fr --input /path/to/audio/file
```

### MLX

[MLX-Whisper](https://github.com/ml-explore/mlx-examples/tree/main/whisper) is a reimplementation of OpenAI's Whisper models in the [MLX](https://github.com/ml-explore/mlx) format - a ML framework on Apple silicon. It supports features like lazy computation, unified memory management, etc.

First, clone the [MLX Examples](https://github.com/ml-explore/mlx-examples) repository:

```bash
git clone https://github.com/ml-explore/mlx-examples.git
cd mlx-examples/whisper
```

Next, install the dependencies:

```bash
pip install -r requirements.txt
```

Download the pytorch checkpoint in the original OpenAI format and convert it into MLX format (We haven't included the converted version here since the repository is already heavy and the conversion is very fast):

```bash
# Download
python -c "from huggingface_hub import hf_hub_download; hf_hub_download(repo_id='bofenghuang/whisper-large-v3-french-distil-dec8', filename='original_model.pt', local_dir='./models/whisper-large-v3-french-distil-dec8')"
# Convert into .npz
python convert.py --torch-name-or-path ./models/whisper-large-v3-french-distil-dec8/original_model.pt --mlx-path ./mlx_models/whisper-large-v3-french-distil-dec8
```

Now, you can transcribe audio with:

```python
import whisper

result = whisper.transcribe("/path/to/audio/file", path_or_hf_repo="mlx_models/whisper-large-v3-french-distil-dec8", language="fr")
print(result["text"])
```

## Training details

We've collected a composite dataset consisting of over 2,500 hours of French speech recognition data, which incldues datasets such as [Common Voice 13.0](https://huggingface.co/datasets/mozilla-foundation/common_voice_13_0), [Multilingual LibriSpeech](https://huggingface.co/datasets/facebook/multilingual_librispeech), [Voxpopuli](https://huggingface.co/datasets/facebook/voxpopuli), [Fleurs](https://huggingface.co/datasets/google/fleurs), [Multilingual TEDx](https://www.openslr.org/100/), [MediaSpeech](https://www.openslr.org/108/), [African Accented French](https://huggingface.co/datasets/gigant/african_accented_french), etc.

Given that some datasets, like MLS, only offer text without case or punctuation, we employed a customized version of 🤗 [Speechbox](https://github.com/huggingface/speechbox) to restore case and punctuation from a limited set of symbols using the [bofenghuang/whisper-large-v2-cv11-french](bofenghuang/whisper-large-v2-cv11-french) model.

However, even within these datasets, we observed certain quality issues. These ranged from mismatches between audio and transcription in terms of language or content, poorly segmented utterances, to missing words in scripted speech, etc. We've built a pipeline to filter out many of these problematic utterances, aiming to enhance the dataset's quality. As a result, we excluded more than 10% of the data, and when we retrained the model, we noticed a significant reduction of hallucination.

For training, we employed the [script](https://github.com/huggingface/distil-whisper/blob/main/training/run_distillation.py) available in the 🤗 Distil-Whisper repository. The model training took place on the [Jean-Zay supercomputer](http://www.idris.fr/eng/jean-zay/jean-zay-presentation-eng.html) at GENCI, and we extend our gratitude to the IDRIS team for their responsive support throughout the project.

## Acknowledgements

- OpenAI for creating and open-sourcing the [Whisper model](https://arxiv.org/abs/2212.04356)
- 🤗 Hugging Face for integrating the Whisper model and providing the training codebase within the [Transformers](https://github.com/huggingface/transformers) and [Distil-Whisper](https://github.com/huggingface/distil-whisper) repository
- [Genci](https://genci.fr/) for their generous contribution of GPU hours to this project