README.md
19.3 KB · 438 lines · markdown Raw
1 ---
2 language:
3 - en
4 - zh
5 - de
6 - es
7 - ru
8 - ko
9 - fr
10 - ja
11 - pt
12 - tr
13 - pl
14 - ca
15 - nl
16 - ar
17 - sv
18 - it
19 - id
20 - hi
21 - fi
22 - vi
23 - he
24 - uk
25 - el
26 - ms
27 - cs
28 - ro
29 - da
30 - hu
31 - ta
32 - no
33 - th
34 - ur
35 - hr
36 - bg
37 - lt
38 - la
39 - mi
40 - ml
41 - cy
42 - sk
43 - te
44 - fa
45 - lv
46 - bn
47 - sr
48 - az
49 - sl
50 - kn
51 - et
52 - mk
53 - br
54 - eu
55 - is
56 - hy
57 - ne
58 - mn
59 - bs
60 - kk
61 - sq
62 - sw
63 - gl
64 - mr
65 - pa
66 - si
67 - km
68 - sn
69 - yo
70 - so
71 - af
72 - oc
73 - ka
74 - be
75 - tg
76 - sd
77 - gu
78 - am
79 - yi
80 - lo
81 - uz
82 - fo
83 - ht
84 - ps
85 - tk
86 - nn
87 - mt
88 - sa
89 - lb
90 - my
91 - bo
92 - tl
93 - mg
94 - as
95 - tt
96 - haw
97 - ln
98 - ha
99 - ba
100 - jw
101 - su
102 tags:
103 - audio
104 - automatic-speech-recognition
105 - hf-asr-leaderboard
106 widget:
107 - example_title: Librispeech sample 1
108 src: https://cdn-media.huggingface.co/speech_samples/sample1.flac
109 - example_title: Librispeech sample 2
110 src: https://cdn-media.huggingface.co/speech_samples/sample2.flac
111 model-index:
112 - name: whisper-base
113 results:
114 - task:
115 name: Automatic Speech Recognition
116 type: automatic-speech-recognition
117 dataset:
118 name: LibriSpeech (clean)
119 type: librispeech_asr
120 config: clean
121 split: test
122 args:
123 language: en
124 metrics:
125 - name: Test WER
126 type: wer
127 value: 5.008769117619326
128 - task:
129 name: Automatic Speech Recognition
130 type: automatic-speech-recognition
131 dataset:
132 name: LibriSpeech (other)
133 type: librispeech_asr
134 config: other
135 split: test
136 args:
137 language: en
138 metrics:
139 - name: Test WER
140 type: wer
141 value: 12.84936273212057
142 - task:
143 name: Automatic Speech Recognition
144 type: automatic-speech-recognition
145 dataset:
146 name: Common Voice 11.0
147 type: mozilla-foundation/common_voice_11_0
148 config: hi
149 split: test
150 args:
151 language: hi
152 metrics:
153 - name: Test WER
154 type: wer
155 value: 131
156 pipeline_tag: automatic-speech-recognition
157 license: apache-2.0
158 ---
159
160 # Whisper
161
162 Whisper is a pre-trained model for automatic speech recognition (ASR) and speech translation. Trained on 680k hours
163 of labelled data, Whisper models demonstrate a strong ability to generalise to many datasets and domains **without** the need
164 for fine-tuning.
165
166 Whisper was proposed in the paper [Robust Speech Recognition via Large-Scale Weak Supervision](https://arxiv.org/abs/2212.04356)
167 by Alec Radford et al from OpenAI. The original code repository can be found [here](https://github.com/openai/whisper).
168
169 **Disclaimer**: Content for this model card has partly been written by the Hugging Face team, and parts of it were
170 copied and pasted from the original model card.
171
172 ## Model details
173
174 Whisper is a Transformer based encoder-decoder model, also referred to as a _sequence-to-sequence_ model.
175 It was trained on 680k hours of labelled speech data annotated using large-scale weak supervision.
176
177 The models were trained on either English-only data or multilingual data. The English-only models were trained
178 on the task of speech recognition. The multilingual models were trained on both speech recognition and speech
179 translation. For speech recognition, the model predicts transcriptions in the *same* language as the audio.
180 For speech translation, the model predicts transcriptions to a *different* language to the audio.
181
182 Whisper checkpoints come in five configurations of varying model sizes.
183 The smallest four are trained on either English-only or multilingual data.
184 The largest checkpoints are multilingual only. All ten of the pre-trained checkpoints
185 are available on the [Hugging Face Hub](https://huggingface.co/models?search=openai/whisper). The
186 checkpoints are summarised in the following table with links to the models on the Hub:
187
188 | Size | Parameters | English-only | Multilingual |
189 |----------|------------|------------------------------------------------------|-----------------------------------------------------|
190 | tiny | 39 M | [✓](https://huggingface.co/openai/whisper-tiny.en) | [✓](https://huggingface.co/openai/whisper-tiny) |
191 | base | 74 M | [✓](https://huggingface.co/openai/whisper-base.en) | [✓](https://huggingface.co/openai/whisper-base) |
192 | small | 244 M | [✓](https://huggingface.co/openai/whisper-small.en) | [✓](https://huggingface.co/openai/whisper-small) |
193 | medium | 769 M | [✓](https://huggingface.co/openai/whisper-medium.en) | [✓](https://huggingface.co/openai/whisper-medium) |
194 | large | 1550 M | x | [✓](https://huggingface.co/openai/whisper-large) |
195 | large-v2 | 1550 M | x | [✓](https://huggingface.co/openai/whisper-large-v2) |
196
197 # Usage
198
199 To transcribe audio samples, the model has to be used alongside a [`WhisperProcessor`](https://huggingface.co/docs/transformers/model_doc/whisper#transformers.WhisperProcessor).
200
201 The `WhisperProcessor` is used to:
202 1. Pre-process the audio inputs (converting them to log-Mel spectrograms for the model)
203 2. Post-process the model outputs (converting them from tokens to text)
204
205 The model is informed of which task to perform (transcription or translation) by passing the appropriate "context tokens". These context tokens
206 are a sequence of tokens that are given to the decoder at the start of the decoding process, and take the following order:
207 1. The transcription always starts with the `<|startoftranscript|>` token
208 2. The second token is the language token (e.g. `<|en|>` for English)
209 3. The third token is the "task token". It can take one of two values: `<|transcribe|>` for speech recognition or `<|translate|>` for speech translation
210 4. In addition, a `<|notimestamps|>` token is added if the model should not include timestamp prediction
211
212 Thus, a typical sequence of context tokens might look as follows:
213 ```
214 <|startoftranscript|> <|en|> <|transcribe|> <|notimestamps|>
215 ```
216 Which tells the model to decode in English, under the task of speech recognition, and not to predict timestamps.
217
218 These tokens can either be forced or un-forced. If they are forced, the model is made to predict each token at
219 each position. This allows one to control the output language and task for the Whisper model. If they are un-forced,
220 the Whisper model will automatically predict the output langauge and task itself.
221
222 The context tokens can be set accordingly:
223
224 ```python
225 model.config.forced_decoder_ids = WhisperProcessor.get_decoder_prompt_ids(language="english", task="transcribe")
226 ```
227
228 Which forces the model to predict in English under the task of speech recognition.
229
230 ## Transcription
231
232 ### English to English
233 In this example, the context tokens are 'unforced', meaning the model automatically predicts the output language
234 (English) and task (transcribe).
235
236 ```python
237 >>> from transformers import WhisperProcessor, WhisperForConditionalGeneration
238 >>> from datasets import load_dataset
239
240 >>> # load model and processor
241 >>> processor = WhisperProcessor.from_pretrained("openai/whisper-base")
242 >>> model = WhisperForConditionalGeneration.from_pretrained("openai/whisper-base")
243 >>> model.config.forced_decoder_ids = None
244
245 >>> # load dummy dataset and read audio files
246 >>> ds = load_dataset("hf-internal-testing/librispeech_asr_dummy", "clean", split="validation")
247 >>> sample = ds[0]["audio"]
248 >>> input_features = processor(sample["array"], sampling_rate=sample["sampling_rate"], return_tensors="pt").input_features
249
250 >>> # generate token ids
251 >>> predicted_ids = model.generate(input_features)
252 >>> # decode token ids to text
253 >>> transcription = processor.batch_decode(predicted_ids, skip_special_tokens=False)
254 ['<|startoftranscript|><|en|><|transcribe|><|notimestamps|> Mr. Quilter is the apostle of the middle classes and we are glad to welcome his gospel.<|endoftext|>']
255
256 >>> transcription = processor.batch_decode(predicted_ids, skip_special_tokens=True)
257 [' Mr. Quilter is the apostle of the middle classes and we are glad to welcome his gospel.']
258 ```
259 The context tokens can be removed from the start of the transcription by setting `skip_special_tokens=True`.
260
261 ### French to French
262 The following example demonstrates French to French transcription by setting the decoder ids appropriately.
263
264 ```python
265 >>> from transformers import WhisperProcessor, WhisperForConditionalGeneration
266 >>> from datasets import Audio, load_dataset
267
268 >>> # load model and processor
269 >>> processor = WhisperProcessor.from_pretrained("openai/whisper-base")
270 >>> model = WhisperForConditionalGeneration.from_pretrained("openai/whisper-base")
271 >>> forced_decoder_ids = processor.get_decoder_prompt_ids(language="french", task="transcribe")
272
273 >>> # load streaming dataset and read first audio sample
274 >>> ds = load_dataset("common_voice", "fr", split="test", streaming=True)
275 >>> ds = ds.cast_column("audio", Audio(sampling_rate=16_000))
276 >>> input_speech = next(iter(ds))["audio"]
277 >>> input_features = processor(input_speech["array"], sampling_rate=input_speech["sampling_rate"], return_tensors="pt").input_features
278
279 >>> # generate token ids
280 >>> predicted_ids = model.generate(input_features, forced_decoder_ids=forced_decoder_ids)
281 >>> # decode token ids to text
282 >>> transcription = processor.batch_decode(predicted_ids)
283 ['<|startoftranscript|><|fr|><|transcribe|><|notimestamps|> Un vrai travail intéressant va enfin être mené sur ce sujet.<|endoftext|>']
284
285 >>> transcription = processor.batch_decode(predicted_ids, skip_special_tokens=True)
286 [' Un vrai travail intéressant va enfin être mené sur ce sujet.']
287 ```
288
289 ## Translation
290 Setting the task to "translate" forces the Whisper model to perform speech translation.
291
292 ### French to English
293
294 ```python
295 >>> from transformers import WhisperProcessor, WhisperForConditionalGeneration
296 >>> from datasets import Audio, load_dataset
297
298 >>> # load model and processor
299 >>> processor = WhisperProcessor.from_pretrained("openai/whisper-base")
300 >>> model = WhisperForConditionalGeneration.from_pretrained("openai/whisper-base")
301 >>> forced_decoder_ids = processor.get_decoder_prompt_ids(language="french", task="translate")
302
303 >>> # load streaming dataset and read first audio sample
304 >>> ds = load_dataset("common_voice", "fr", split="test", streaming=True)
305 >>> ds = ds.cast_column("audio", Audio(sampling_rate=16_000))
306 >>> input_speech = next(iter(ds))["audio"]
307 >>> input_features = processor(input_speech["array"], sampling_rate=input_speech["sampling_rate"], return_tensors="pt").input_features
308
309 >>> # generate token ids
310 >>> predicted_ids = model.generate(input_features, forced_decoder_ids=forced_decoder_ids)
311 >>> # decode token ids to text
312 >>> transcription = processor.batch_decode(predicted_ids, skip_special_tokens=True)
313 [' A very interesting work, we will finally be given on this subject.']
314 ```
315
316 ## Evaluation
317
318 This code snippet shows how to evaluate Whisper Base on [LibriSpeech test-clean](https://huggingface.co/datasets/librispeech_asr):
319
320 ```python
321 >>> from datasets import load_dataset
322 >>> from transformers import WhisperForConditionalGeneration, WhisperProcessor
323 >>> import torch
324 >>> from evaluate import load
325
326 >>> librispeech_test_clean = load_dataset("librispeech_asr", "clean", split="test")
327
328 >>> processor = WhisperProcessor.from_pretrained("openai/whisper-base")
329 >>> model = WhisperForConditionalGeneration.from_pretrained("openai/whisper-base").to("cuda")
330
331 >>> def map_to_pred(batch):
332 >>> audio = batch["audio"]
333 >>> input_features = processor(audio["array"], sampling_rate=audio["sampling_rate"], return_tensors="pt").input_features
334 >>> batch["reference"] = processor.tokenizer._normalize(batch['text'])
335 >>>
336 >>> with torch.no_grad():
337 >>> predicted_ids = model.generate(input_features.to("cuda"))[0]
338 >>> transcription = processor.decode(predicted_ids)
339 >>> batch["prediction"] = processor.tokenizer._normalize(transcription)
340 >>> return batch
341
342 >>> result = librispeech_test_clean.map(map_to_pred)
343
344 >>> wer = load("wer")
345 >>> print(100 * wer.compute(references=result["reference"], predictions=result["prediction"]))
346 5.082316555716899
347 ```
348
349 ## Long-Form Transcription
350
351 The Whisper model is intrinsically designed to work on audio samples of up to 30s in duration. However, by using a chunking
352 algorithm, it can be used to transcribe audio samples of up to arbitrary length. This is possible through Transformers
353 [`pipeline`](https://huggingface.co/docs/transformers/main_classes/pipelines#transformers.AutomaticSpeechRecognitionPipeline)
354 method. Chunking is enabled by setting `chunk_length_s=30` when instantiating the pipeline. With chunking enabled, the pipeline
355 can be run with batched inference. It can also be extended to predict sequence level timestamps by passing `return_timestamps=True`:
356
357 ```python
358 >>> import torch
359 >>> from transformers import pipeline
360 >>> from datasets import load_dataset
361
362 >>> device = "cuda:0" if torch.cuda.is_available() else "cpu"
363
364 >>> pipe = pipeline(
365 >>> "automatic-speech-recognition",
366 >>> model="openai/whisper-base",
367 >>> chunk_length_s=30,
368 >>> device=device,
369 >>> )
370
371 >>> ds = load_dataset("hf-internal-testing/librispeech_asr_dummy", "clean", split="validation")
372 >>> sample = ds[0]["audio"]
373
374 >>> prediction = pipe(sample.copy(), batch_size=8)["text"]
375 " Mr. Quilter is the apostle of the middle classes, and we are glad to welcome his gospel."
376
377 >>> # we can also return timestamps for the predictions
378 >>> prediction = pipe(sample.copy(), batch_size=8, return_timestamps=True)["chunks"]
379 [{'text': ' Mr. Quilter is the apostle of the middle classes and we are glad to welcome his gospel.',
380 'timestamp': (0.0, 5.44)}]
381 ```
382
383 Refer to the blog post [ASR Chunking](https://huggingface.co/blog/asr-chunking) for more details on the chunking algorithm.
384
385 ## Fine-Tuning
386
387 The pre-trained Whisper model demonstrates a strong ability to generalise to different datasets and domains. However,
388 its predictive capabilities can be improved further for certain languages and tasks through *fine-tuning*. The blog
389 post [Fine-Tune Whisper with 🤗 Transformers](https://huggingface.co/blog/fine-tune-whisper) provides a step-by-step
390 guide to fine-tuning the Whisper model with as little as 5 hours of labelled data.
391
392 ### Evaluated Use
393
394 The primary intended users of these models are AI researchers studying robustness, generalization, capabilities, biases, and constraints of the current model. However, Whisper is also potentially quite useful as an ASR solution for developers, especially for English speech recognition. We recognize that once models are released, it is impossible to restrict access to only “intended” uses or to draw reasonable guidelines around what is or is not research.
395
396 The models are primarily trained and evaluated on ASR and speech translation to English tasks. They show strong ASR results in ~10 languages. They may exhibit additional capabilities, particularly if fine-tuned on certain tasks like voice activity detection, speaker classification, or speaker diarization but have not been robustly evaluated in these areas. We strongly recommend that users perform robust evaluations of the models in a particular context and domain before deploying them.
397
398 In particular, we caution against using Whisper models to transcribe recordings of individuals taken without their consent or purporting to use these models for any kind of subjective classification. We recommend against use in high-risk domains like decision-making contexts, where flaws in accuracy can lead to pronounced flaws in outcomes. The models are intended to transcribe and translate speech, use of the model for classification is not only not evaluated but also not appropriate, particularly to infer human attributes.
399
400
401 ## Training Data
402
403 The models are trained on 680,000 hours of audio and the corresponding transcripts collected from the internet. 65% of this data (or 438,000 hours) represents English-language audio and matched English transcripts, roughly 18% (or 126,000 hours) represents non-English audio and English transcripts, while the final 17% (or 117,000 hours) represents non-English audio and the corresponding transcript. This non-English data represents 98 different languages.
404
405 As discussed in [the accompanying paper](https://cdn.openai.com/papers/whisper.pdf), we see that performance on transcription in a given language is directly correlated with the amount of training data we employ in that language.
406
407
408 ## Performance and Limitations
409
410 Our studies show that, over many existing ASR systems, the models exhibit improved robustness to accents, background noise, technical language, as well as zero shot translation from multiple languages into English; and that accuracy on speech recognition and translation is near the state-of-the-art level.
411
412 However, because the models are trained in a weakly supervised manner using large-scale noisy data, the predictions may include texts that are not actually spoken in the audio input (i.e. hallucination). We hypothesize that this happens because, given their general knowledge of language, the models combine trying to predict the next word in audio with trying to transcribe the audio itself.
413
414 Our models perform unevenly across languages, and we observe lower accuracy on low-resource and/or low-discoverability languages or languages where we have less training data. The models also exhibit disparate performance on different accents and dialects of particular languages, which may include higher word error rate across speakers of different genders, races, ages, or other demographic criteria. Our full evaluation results are presented in [the paper accompanying this release](https://cdn.openai.com/papers/whisper.pdf).
415
416 In addition, the sequence-to-sequence architecture of the model makes it prone to generating repetitive texts, which can be mitigated to some degree by beam search and temperature scheduling but not perfectly. Further analysis on these limitations are provided in [the paper](https://cdn.openai.com/papers/whisper.pdf). It is likely that this behavior and hallucinations may be worse on lower-resource and/or lower-discoverability languages.
417
418
419 ## Broader Implications
420
421 We anticipate that Whisper models’ transcription capabilities may be used for improving accessibility tools. While Whisper models cannot be used for real-time transcription out of the box – their speed and size suggest that others may be able to build applications on top of them that allow for near-real-time speech recognition and translation. The real value of beneficial applications built on top of Whisper models suggests that the disparate performance of these models may have real economic implications.
422
423 There are also potential dual use concerns that come with releasing Whisper. While we hope the technology will be used primarily for beneficial purposes, making ASR technology more accessible could enable more actors to build capable surveillance technologies or scale up existing surveillance efforts, as the speed and accuracy allow for affordable automatic transcription and translation of large volumes of audio communication. Moreover, these models may have some capabilities to recognize specific individuals out of the box, which in turn presents safety concerns related both to dual use and disparate performance. In practice, we expect that the cost of transcription is not the limiting factor of scaling up surveillance projects.
424
425
426 ### BibTeX entry and citation info
427 ```bibtex
428 @misc{radford2022whisper,
429 doi = {10.48550/ARXIV.2212.04356},
430 url = {https://arxiv.org/abs/2212.04356},
431 author = {Radford, Alec and Kim, Jong Wook and Xu, Tao and Brockman, Greg and McLeavey, Christine and Sutskever, Ilya},
432 title = {Robust Speech Recognition via Large-Scale Weak Supervision},
433 publisher = {arXiv},
434 year = {2022},
435 copyright = {arXiv.org perpetual, non-exclusive license}
436 }
437 ```
438