Update README.md
Browse files
README.md
CHANGED
|
@@ -8,4 +8,329 @@ language:
|
|
| 8 |
- ja
|
| 9 |
pipeline_tag: automatic-speech-recognition
|
| 10 |
library_name: transformers
|
| 11 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 8 |
- ja
|
| 9 |
pipeline_tag: automatic-speech-recognition
|
| 10 |
library_name: transformers
|
| 11 |
+
tags:
|
| 12 |
+
- audio
|
| 13 |
+
- automatic-speech-recognition
|
| 14 |
+
- hf-asr-leaderboard
|
| 15 |
+
---
|
| 16 |
+
|
| 17 |
+
# Kotoba-Whisper-Bilingual (v1.0)
|
| 18 |
+
|
| 19 |
+
_Kotoba-Whisper-Bilingual_ is a collection of distilled [Whisper](https://arxiv.org/abs/2212.04356) models trained for
|
| 20 |
+
- **Japanese ASR**
|
| 21 |
+
- **English ASR**
|
| 22 |
+
- **Speech-to-text translation (Japanese -> English)**
|
| 23 |
+
- **Speech-to-text translation (English -> Japanese)**
|
| 24 |
+
|
| 25 |
+
developed through the collaboration bewteen
|
| 26 |
+
[Asahi Ushio](https://asahiushio.com) and [Kotoba Technologies](https://twitter.com/kotoba_tech).
|
| 27 |
+
Following the original work of distil-whisper ([Robust Knowledge Distillation via Large-Scale Pseudo Labelling](https://arxiv.org/abs/2311.00430)),
|
| 28 |
+
we employ OpenAI's [Whisper large-v3](https://huggingface.co/openai/whisper-large-v3) as the teacher model for Japanese and English ASR, while we translate the
|
| 29 |
+
transcription into English and Japanese by [chatgpt](https://openai.com/chatgpt/) to obtain training dataset for speech-to-text translation.
|
| 30 |
+
We employ [ReazonSpeech](https://huggingface.co/datasets/japanese-asr/ja_asr.reazon_speech_all) for Japanese ASR and Japanese speech to English text translation,
|
| 31 |
+
and [Multilingual LibriSpeech](https://huggingface.co/datasets/japanese-asr/en_asr.mls) for English ASR and English speech to Japanese text translation.
|
| 32 |
+
Kotoba-whisper-bilingual's loss objective consists of cross-entropy on both of ASR and translation tasks, while KL divergence loss only for ASR task.
|
| 33 |
+
The student model consists the full encoder of the teacher large-v3 model and the decoder with two layers initialized from the first and last layer of the large-v3 model.
|
| 34 |
+
|
| 35 |
+
Kotoba-Whisper is **6.3x faster than large-v3**, while retaining as low error rate as the large-v3.
|
| 36 |
+
|
| 37 |
+
## Evaluation
|
| 38 |
+
|
| 39 |
+
### Speech2Text Translation (Japanese->English)
|
| 40 |
+
|
| 41 |
+
| model | [CoVoST2 (Ja->En)](https://huggingface.co/datasets/japanese-asr/ja2en.s2t_translation)| [Fleurs (Ja->En)](https://huggingface.co/datasets/japanese-asr/ja2en.s2t_translation) |
|
| 42 |
+
|:----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|-------------------------------------------------------------------------------------------------------:|------------------------------------------------------------------------------------------------------:|
|
| 43 |
+
| [kotoba-tech/kotoba-whisper-bilingual-v1.0](https://huggingface.co/kotoba-tech/kotoba-whisper-bilingual-v1.0) | 73.9 | 98.7 |
|
| 44 |
+
| [japanese-asr/ja-cascaded-s2t-translation](https://huggingface.co/japanese-asr/ja-cascaded-s2t-translation) ([facebook/nllb-200-3.3B](https://huggingface.co/facebook/nllb-200-3.3B)) | 64.3 | 67.1 |
|
| 45 |
+
| [japanese-asr/ja-cascaded-s2t-translation](https://huggingface.co/japanese-asr/ja-cascaded-s2t-translation) ([facebook/nllb-200-1.3B](https://huggingface.co/facebook/nllb-200-1.3B)) | 65.4 | 68.9 |
|
| 46 |
+
| [japanese-asr/ja-cascaded-s2t-translation](https://huggingface.co/japanese-asr/ja-cascaded-s2t-translation) ([facebook/nllb-200-distilled-1.3B](https://huggingface.co/facebook/nllb-200-distilled-1.3B)) | 65.6 | 67.4 |
|
| 47 |
+
| [japanese-asr/ja-cascaded-s2t-translation](https://huggingface.co/japanese-asr/ja-cascaded-s2t-translation) ([facebook/nllb-200-distilled-600M](https://huggingface.co/facebook/nllb-200-distilled-600M)) | 68.2 | 72.2 |
|
| 48 |
+
| [openai/whisper-large-v3](https://huggingface.co/openai/whisper-large-v3) | 71 | 86.1 |
|
| 49 |
+
| [openai/whisper-large-v2](https://huggingface.co/openai/whisper-large-v2) | 66.4 | 78.8 |
|
| 50 |
+
| [openai/whisper-large](https://huggingface.co/openai/whisper-large) | 66.5 | 86.1 |
|
| 51 |
+
| [openai/whisper-medium](https://huggingface.co/openai/whisper-medium) | 70.3 | 97.2 |
|
| 52 |
+
| [openai/whisper-small](https://huggingface.co/openai/whisper-small) | 97.3 | 132.2 |
|
| 53 |
+
| [openai/whisper-base](https://huggingface.co/openai/whisper-base) | 186.2 | 349.6 |
|
| 54 |
+
| [openai/whisper-tiny](https://huggingface.co/openai/whisper-tiny) | 377.2 | 474 |
|
| 55 |
+
|
| 56 |
+
|
| 57 |
+
### Speech2Text Translation (English->Japanese)
|
| 58 |
+
|
| 59 |
+
| model | [CoVoST2 (En->Ja)](https://huggingface.co/datasets/japanese-asr/en2ja.s2t_translation)| [Fleurs (En->JA)](https://huggingface.co/datasets/japanese-asr/en2ja.s2t_translation) |
|
| 60 |
+
|:----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|-------------------------------------------------------------------------------------------------------:|------------------------------------------------------------------------------------------------------:|
|
| 61 |
+
| [kotoba-tech/kotoba-whisper-bilingual-v1.0](https://huggingface.co/kotoba-tech/kotoba-whisper-bilingual-v1.0) | 69.1 | 74.4 |
|
| 62 |
+
| [japanese-asr/en-cascaded-s2t-translation](https://huggingface.co/japanese-asr/en-cascaded-s2t-translation) ([facebook/nllb-200-3.3B](https://huggingface.co/facebook/nllb-200-3.3B)) | 62.4 | 63.5 |
|
| 63 |
+
| [japanese-asr/en-cascaded-s2t-translation](https://huggingface.co/japanese-asr/en-cascaded-s2t-translation) ([facebook/nllb-200-1.3B](https://huggingface.co/facebook/nllb-200-1.3B)) | 64.4 | 67.2 |
|
| 64 |
+
| [japanese-asr/en-cascaded-s2t-translation](https://huggingface.co/japanese-asr/en-cascaded-s2t-translation) ([facebook/nllb-200-distilled-1.3B](https://huggingface.co/facebook/nllb-200-distilled-1.3B)) | 62.4 | 62.9 |
|
| 65 |
+
| [japanese-asr/en-cascaded-s2t-translation](https://huggingface.co/japanese-asr/en-cascaded-s2t-translation) ([facebook/nllb-200-distilled-600M](https://huggingface.co/facebook/nllb-200-distilled-600M)) | 63.4 | 66.2 |
|
| 66 |
+
| [openai/whisper-large-v3](https://huggingface.co/openai/whisper-large-v3) | 178.9 | 209.5 |
|
| 67 |
+
| [openai/whisper-large-v2](https://huggingface.co/openai/whisper-large-v2) | 179.6 | 201.8 |
|
| 68 |
+
| [openai/whisper-large](https://huggingface.co/openai/whisper-large) | 178.7 | 201.8 |
|
| 69 |
+
| [openai/whisper-medium](https://huggingface.co/openai/whisper-medium) | 178.7 | 202 |
|
| 70 |
+
| [openai/whisper-small](https://huggingface.co/openai/whisper-small) | 178.9 | 206.8 |
|
| 71 |
+
| [openai/whisper-base](https://huggingface.co/openai/whisper-base) | 179.5 | 214.2 |
|
| 72 |
+
| [openai/whisper-tiny](https://huggingface.co/openai/whisper-tiny) | 185.2 | 200.5 |
|
| 73 |
+
|
| 74 |
+
|
| 75 |
+
### ASR (Japanese)
|
| 76 |
+
|
| 77 |
+
| model | [CommonVoice 8 (Japanese test set)](https://huggingface.co/datasets/japanese-asr/ja_asr.common_voice_8_0) | [JSUT Basic 5000](https://huggingface.co/datasets/japanese-asr/ja_asr.jsut_basic5000) | [ReazonSpeech (held out test set)](https://huggingface.co/datasets/japanese-asr/ja_asr.reazonspeech_test) |
|
| 78 |
+
|:--------------------------------------------------------------------------------------------------------------------------------------------------|------------------------------------------------------------------------------------------------------------:|----------------------------------------------------------------------------------------:|------------------------------------------------------------------------------------------------------------:|
|
| 79 |
+
| [kotoba-tech/kotoba-whisper-bilingual-v1.0](https://huggingface.co/kotoba-tech/kotoba-whisper-bilingual-v1.0) | 9.8 | 9.3 | 16.8 |
|
| 80 |
+
| [kotoba-tech/kotoba-whisper-v2.0](https://huggingface.co/kotoba-tech/kotoba-whisper-v2.0) | 9.2 | 8.4 | 11.6 |
|
| 81 |
+
| [kotoba-tech/kotoba-whisper-v1.0](https://huggingface.co/kotoba-tech/kotoba-whisper-v1.0) | 9.4 | 8.5 | 12.2 |
|
| 82 |
+
| [openai/whisper-large-v3](https://huggingface.co/openai/whisper-large-v3) | 8.5 | 7.1 | 14.9 |
|
| 83 |
+
| [openai/whisper-large-v2](https://huggingface.co/openai/whisper-large-v2) | 9.7 | 8.2 | 28.1 |
|
| 84 |
+
| [openai/whisper-large](https://huggingface.co/openai/whisper-large) | 10 | 8.9 | 34.1 |
|
| 85 |
+
| [openai/whisper-medium](https://huggingface.co/openai/whisper-medium) | 11.5 | 10 | 33.2 |
|
| 86 |
+
| [openai/whisper-small](https://huggingface.co/openai/whisper-small) | 15.1 | 14.2 | 41.5 |
|
| 87 |
+
| [openai/whisper-base](https://huggingface.co/openai/whisper-base) | 28.6 | 24.9 | 70.4 |
|
| 88 |
+
| [openai/whisper-tiny](https://huggingface.co/openai/whisper-tiny) | 53.7 | 36.5 | 137.9 |
|
| 89 |
+
| [reazon-research/reazonspeech-nemo-v2](https://huggingface.co/reazon-research/reazonspeech-nemo-v2) | 9.1 | 7.4 | 11.2 |
|
| 90 |
+
|
| 91 |
+
|
| 92 |
+
|
| 93 |
+
### ASR (English)
|
| 94 |
+
|
| 95 |
+
| model | [ESB](https://huggingface.co/datasets/japanese-asr/en_asr.esb_eval) (ami) | [ESB](https://huggingface.co/datasets/japanese-asr/en_asr.esb_eval) (earnings22) | [ESB](https://huggingface.co/datasets/japanese-asr/en_asr.esb_eval) (librispeech) | [ESB](https://huggingface.co/datasets/japanese-asr/en_asr.esb_eval) (tedlium) | [ESB](https://huggingface.co/datasets/japanese-asr/en_asr.esb_eval) (voxpopuli) |
|
| 96 |
+
|:----------------------------------------------------------------------------------------------------------------|----------------------------------------------------------------------------:|-----------------------------------------------------------------------------------:|------------------------------------------------------------------------------------:|--------------------------------------------------------------------------------:|----------------------------------------------------------------------------------:|
|
| 97 |
+
| [kotoba-tech/kotoba-whisper-bilingual-v1.0](https://huggingface.co/kotoba-tech/kotoba-whisper-bilingual-v1.0) | 16.7 | 15.3 | 2.4 | 4.1 | 8.3 |
|
| 98 |
+
| [openai/whisper-large-v3](https://huggingface.co/openai/whisper-large-v3) | 17.9 | 14.9 | 2.1 | 3.8 | 12.7 |
|
| 99 |
+
| [openai/whisper-large-v2](https://huggingface.co/openai/whisper-large-v2) | 18.9 | 16.7 | 2.3 | 4.9 | 7.7 |
|
| 100 |
+
| [openai/whisper-large](https://huggingface.co/openai/whisper-large) | 18.8 | 14.9 | 2.6 | 4.2 | 7.7 |
|
| 101 |
+
| [openai/whisper-medium](https://huggingface.co/openai/whisper-medium) | 18.3 | 14.9 | 2.5 | 4.3 | 7.9 |
|
| 102 |
+
| [openai/whisper-small](https://huggingface.co/openai/whisper-small) | 23.1 | 17.2 | 3.5 | 5.3 | 10.8 |
|
| 103 |
+
| [openai/whisper-base](https://huggingface.co/openai/whisper-base) | 26.6 | 21 | 6 | 6.1 | 11.3 |
|
| 104 |
+
| [openai/whisper-tiny](https://huggingface.co/openai/whisper-tiny) | 31.9 | 30.5 | 8.2 | 11.7 | 15.1 |
|
| 105 |
+
| [japanese-asr/distil-whisper-bilingual-v1.0](https://huggingface.co/japanese-asr/distil-whisper-bilingual-v1.0) | 20.7 | 18.6 | 2.4 | 6.4 | 10 |
|
| 106 |
+
|
| 107 |
+
|
| 108 |
+
|
| 109 |
+
|
| 110 |
+
|
| 111 |
+
|
| 112 |
+
|
| 113 |
+
- ***Latency***: As kotoba-whisper uses the same architecture as [distil-whisper/distil-large-v3](https://huggingface.co/distil-whisper/distil-large-v3),
|
| 114 |
+
it inherits the benefit of the improved latency compared to [openai/whisper-large-v3](https://huggingface.co/openai/whisper-large-v3)
|
| 115 |
+
(**6.3x faster than large-v3**, see the table below taken from [distil-whisper/distil-large-v3](https://huggingface.co/distil-whisper/distil-large-v3)).
|
| 116 |
+
|
| 117 |
+
| Model | Params / M | Rel. Latency |
|
| 118 |
+
|----------------------------------------------------------------------------------------------|------------|--------------|
|
| 119 |
+
| **[kotoba-tech/kotoba-whisper-v2.0](https://huggingface.co/kotoba-tech/kotoba-whisper-v2.0)**| **756** | **6.3** |
|
| 120 |
+
| **[kotoba-tech/kotoba-whisper-v1.0](https://huggingface.co/kotoba-tech/kotoba-whisper-v1.0)**| **756** | **6.3** |
|
| 121 |
+
| [openai/whisper-large-v3](https://huggingface.co/openai/whisper-large-v3) | 1550 | 1.0 |
|
| 122 |
+
|
| 123 |
+
|
| 124 |
+
## Transformers Usage
|
| 125 |
+
Kotoba-Whisper is supported in the Hugging Face 🤗 Transformers library from version 4.39 onwards. To run the model, first
|
| 126 |
+
install the latest version of Transformers.
|
| 127 |
+
|
| 128 |
+
```bash
|
| 129 |
+
pip install --upgrade pip
|
| 130 |
+
pip install --upgrade transformers accelerate
|
| 131 |
+
```
|
| 132 |
+
|
| 133 |
+
### Short-Form Transcription
|
| 134 |
+
The model can be used with the [`pipeline`](https://huggingface.co/docs/transformers/main_classes/pipelines#transformers.AutomaticSpeechRecognitionPipeline)
|
| 135 |
+
class to transcribe short-form audio files (< 30-seconds) as follows:
|
| 136 |
+
|
| 137 |
+
Download sample audio.
|
| 138 |
+
```shell
|
| 139 |
+
wget
|
| 140 |
+
```
|
| 141 |
+
|
| 142 |
+
```python
|
| 143 |
+
import torch
|
| 144 |
+
from transformers import pipeline
|
| 145 |
+
from datasets import load_dataset
|
| 146 |
+
|
| 147 |
+
# config
|
| 148 |
+
torch_dtype = torch.bfloat16 if torch.cuda.is_available() else torch.float32
|
| 149 |
+
device = "cuda:0" if torch.cuda.is_available() else "cpu"
|
| 150 |
+
model_kwargs = {"attn_implementation": "sdpa"} if torch.cuda.is_available() else {}
|
| 151 |
+
|
| 152 |
+
# load model
|
| 153 |
+
pipe = pipeline(
|
| 154 |
+
"automatic-speech-recognition",
|
| 155 |
+
model="kotoba-tech/kotoba-whisper-bilingual-v1.0",
|
| 156 |
+
torch_dtype=torch_dtype,
|
| 157 |
+
device=device,
|
| 158 |
+
model_kwargs=model_kwargs
|
| 159 |
+
)
|
| 160 |
+
|
| 161 |
+
|
| 162 |
+
generate_kwargs = {"language": "ja", "task": "transcribe"}
|
| 163 |
+
|
| 164 |
+
# load sample audio
|
| 165 |
+
dataset = load_dataset("japanese-asr/ja_asr.reazonspeech_test", split="test")
|
| 166 |
+
sample = dataset[0]["audio"]
|
| 167 |
+
|
| 168 |
+
# run inference
|
| 169 |
+
result = pipe(sample, generate_kwargs=generate_kwargs)
|
| 170 |
+
print(result["text"])
|
| 171 |
+
```
|
| 172 |
+
|
| 173 |
+
- To transcribe a local audio file, simply pass the path to your audio file when you call the pipeline (make sure the audio is sampled in 16kHz):
|
| 174 |
+
```diff
|
| 175 |
+
- result = pipe(sample, generate_kwargs=generate_kwargs)
|
| 176 |
+
+ result = pipe("audio.mp3", generate_kwargs=generate_kwargs)
|
| 177 |
+
```
|
| 178 |
+
|
| 179 |
+
- For segment-level timestamps, pass the argument `return_timestamps=True` and return the `"chunks"` output:
|
| 180 |
+
```python
|
| 181 |
+
result = pipe(sample, return_timestamps=True, generate_kwargs=generate_kwargs)
|
| 182 |
+
print(result["chunks"])
|
| 183 |
+
```
|
| 184 |
+
|
| 185 |
+
***Sequential Long-Form:*** Kotoba-whisper is designed to be compatible with OpenAI's sequential long-form transcription algorithm. This algorithm uses a sliding window for buffered
|
| 186 |
+
inference of long audio files (> 30-seconds), and returns more accurate transcriptions compared to the [chunked long-form algorithm](#chunked-long-form).
|
| 187 |
+
As default, if long audio files are passed to the model, it will transcribes with the sequential long-form transcription.
|
| 188 |
+
The sequential long-form algorithm should be used in either of the following scenarios:
|
| 189 |
+
|
| 190 |
+
1. Transcription accuracy is the most important factor, and latency is less of a consideration
|
| 191 |
+
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
|
| 192 |
+
|
| 193 |
+
If you are transcribing single long audio files and latency is the most important factor, you should use the chunked algorithm
|
| 194 |
+
described [below](#chunked-long-form). For a detailed explanation of the different algorithms, refer to Sections 5 of
|
| 195 |
+
the [Distil-Whisper paper](https://arxiv.org/pdf/2311.00430.pdf). The [`pipeline`](https://huggingface.co/docs/transformers/main_classes/pipelines#transformers.AutomaticSpeechRecognitionPipeline)
|
| 196 |
+
class can be used to transcribe long audio files with the sequential algorithm as follows:
|
| 197 |
+
|
| 198 |
+
|
| 199 |
+
### Chunked Long-Form
|
| 200 |
+
This algorithm should be used when a single large audio file is being transcribed and the fastest possible inference is required. In such circumstances,
|
| 201 |
+
the chunked algorithm is up to 9x faster than OpenAI's sequential long-form implementation (see Table 7 of the [Distil-Whisper paper](https://arxiv.org/pdf/2311.00430.pdf)).
|
| 202 |
+
To enable chunking, pass the `chunk_length_s` parameter to the `pipeline`. For distil-large-v3, a chunk length of 25-seconds
|
| 203 |
+
is optimal. To activate batching over long audio files, pass the argument `batch_size`:
|
| 204 |
+
|
| 205 |
+
```python
|
| 206 |
+
import torch
|
| 207 |
+
from transformers import pipeline
|
| 208 |
+
from datasets import load_dataset
|
| 209 |
+
|
| 210 |
+
# config
|
| 211 |
+
model_id = "kotoba-tech/kotoba-whisper-v2.0"
|
| 212 |
+
torch_dtype = torch.bfloat16 if torch.cuda.is_available() else torch.float32
|
| 213 |
+
device = "cuda:0" if torch.cuda.is_available() else "cpu"
|
| 214 |
+
model_kwargs = {"attn_implementation": "sdpa"} if torch.cuda.is_available() else {}
|
| 215 |
+
generate_kwargs = {"language": "ja", "task": "transcribe"}
|
| 216 |
+
|
| 217 |
+
# load model
|
| 218 |
+
pipe = pipeline(
|
| 219 |
+
"automatic-speech-recognition",
|
| 220 |
+
model=model_id,
|
| 221 |
+
torch_dtype=torch_dtype,
|
| 222 |
+
device=device,
|
| 223 |
+
model_kwargs=model_kwargs,
|
| 224 |
+
chunk_length_s=15,
|
| 225 |
+
batch_size=16
|
| 226 |
+
)
|
| 227 |
+
|
| 228 |
+
# load sample audio (concatenate instances to create a long audio)
|
| 229 |
+
dataset = load_dataset("japanese-asr/ja_asr.reazonspeech_test", split="test")
|
| 230 |
+
sample = {"array": np.concatenate([i["array"] for i in dataset[:20]["audio"]]), "sampling_rate": dataset[0]['audio']['sampling_rate']}
|
| 231 |
+
|
| 232 |
+
# run inference
|
| 233 |
+
result = pipe(sample, generate_kwargs=generate_kwargs)
|
| 234 |
+
print(result["text"])
|
| 235 |
+
```
|
| 236 |
+
|
| 237 |
+
|
| 238 |
+
### Additional Speed & Memory Improvements
|
| 239 |
+
You can apply additional speed and memory improvements to further reduce the inference speed and VRAM
|
| 240 |
+
requirements. These optimisations primarily target the attention kernel, swapping it from an eager implementation to a
|
| 241 |
+
more efficient flash attention version.
|
| 242 |
+
|
| 243 |
+
#### Flash Attention 2
|
| 244 |
+
|
| 245 |
+
We recommend using [Flash-Attention 2](https://huggingface.co/docs/transformers/main/en/perf_infer_gpu_one#flashattention-2)
|
| 246 |
+
if your GPU allows for it. To do so, you first need to install [Flash Attention](https://github.com/Dao-AILab/flash-attention):
|
| 247 |
+
|
| 248 |
+
```
|
| 249 |
+
pip install flash-attn --no-build-isolation
|
| 250 |
+
```
|
| 251 |
+
|
| 252 |
+
Then pass `attn_implementation="flash_attention_2"` to `from_pretrained`:
|
| 253 |
+
|
| 254 |
+
```diff
|
| 255 |
+
- model_kwargs = {"attn_implementation": "sdpa"} if torch.cuda.is_available() else {}
|
| 256 |
+
+ model_kwargs = {"attn_implementation": "flash_attention_2"} if torch.cuda.is_available() else {}
|
| 257 |
+
```
|
| 258 |
+
|
| 259 |
+
|
| 260 |
+
## Model Details
|
| 261 |
+
See [https://huggingface.co/distil-whisper/distil-large-v3#model-details](https://huggingface.co/distil-whisper/distil-large-v3#model-details).
|
| 262 |
+
|
| 263 |
+
|
| 264 |
+
## Training
|
| 265 |
+
Please refer to [https://github.com/kotoba-tech/kotoba-whisper](https://github.com/kotoba-tech/kotoba-whisper) for the model training detail.
|
| 266 |
+
Datasets used in distillation and the whole model variations can be found at [https://huggingface.co/japanese-asr](https://huggingface.co/japanese-asr).
|
| 267 |
+
|
| 268 |
+
|
| 269 |
+
## Evaluation
|
| 270 |
+
The following code-snippets demonstrates how to evaluate the kotoba-whisper model on the Japanese subset of the CommonVoice 8.0.
|
| 271 |
+
First, we need to install the required packages, including 🤗 Datasets to load the audio data, and 🤗 Evaluate to
|
| 272 |
+
perform the WER calculation:
|
| 273 |
+
|
| 274 |
+
```bash
|
| 275 |
+
pip install --upgrade pip
|
| 276 |
+
pip install --upgrade transformers datasets[audio] evaluate jiwer
|
| 277 |
+
```
|
| 278 |
+
|
| 279 |
+
Evaluation can then be run end-to-end with the following example:
|
| 280 |
+
|
| 281 |
+
```python
|
| 282 |
+
import torch
|
| 283 |
+
from transformers import pipeline
|
| 284 |
+
from datasets import load_dataset
|
| 285 |
+
from evaluate import load
|
| 286 |
+
from transformers.models.whisper.english_normalizer import BasicTextNormalizer
|
| 287 |
+
|
| 288 |
+
# model config
|
| 289 |
+
model_id = "kotoba-tech/kotoba-whisper-v2.0"
|
| 290 |
+
torch_dtype = torch.bfloat16 if torch.cuda.is_available() else torch.float32
|
| 291 |
+
device = "cuda:0" if torch.cuda.is_available() else "cpu"
|
| 292 |
+
model_kwargs = {"attn_implementation": "sdpa"} if torch.cuda.is_available() else {}
|
| 293 |
+
generate_kwargs = {"language": "japanese", "task": "transcribe"}
|
| 294 |
+
normalizer = BasicTextNormalizer()
|
| 295 |
+
|
| 296 |
+
# data config
|
| 297 |
+
dataset_name = "japanese-asr/ja_asr.reazonspeech_test"
|
| 298 |
+
audio_column = 'audio'
|
| 299 |
+
text_column = 'transcription'
|
| 300 |
+
|
| 301 |
+
# load model
|
| 302 |
+
pipe = pipeline(
|
| 303 |
+
"automatic-speech-recognition",
|
| 304 |
+
model=model_id,
|
| 305 |
+
torch_dtype=torch_dtype,
|
| 306 |
+
device=device,
|
| 307 |
+
model_kwargs=model_kwargs,
|
| 308 |
+
batch_size=16
|
| 309 |
+
)
|
| 310 |
+
|
| 311 |
+
# load the dataset and sample the audio with 16kHz
|
| 312 |
+
dataset = load_dataset(dataset_name, split="test")
|
| 313 |
+
transcriptions = pipe(dataset['audio'])
|
| 314 |
+
transcriptions = [normalizer(i['text']).replace(" ", "") for i in transcriptions]
|
| 315 |
+
references = [normalizer(i).replace(" ", "") for i in dataset['transcription']]
|
| 316 |
+
|
| 317 |
+
# compute the CER metric
|
| 318 |
+
cer_metric = load("cer")
|
| 319 |
+
cer = 100 * cer_metric.compute(predictions=transcriptions, references=references)
|
| 320 |
+
print(cer)
|
| 321 |
+
```
|
| 322 |
+
|
| 323 |
+
The huggingface links to the major Japanese ASR datasets for evaluation are summarized at [here](https://huggingface.co/collections/japanese-asr/japanese-asr-evaluation-dataset-66051a03d6ca494d40baaa26).
|
| 324 |
+
For example, to evaluate the model on JSUT Basic5000, change the `dataset_name`:
|
| 325 |
+
|
| 326 |
+
```diff
|
| 327 |
+
- dataset_name = "japanese-asr/ja_asr.reazonspeech_test"
|
| 328 |
+
+ dataset_name = "japanese-asr/ja_asr.jsut_basic5000"
|
| 329 |
+
```
|
| 330 |
+
|
| 331 |
+
## Acknowledgements
|
| 332 |
+
* [OpenAI](https://openai.com/) for the Whisper [model](https://huggingface.co/openai/whisper-large-v3).
|
| 333 |
+
* Hugging Face 🤗 [Transformers](https://github.com/huggingface/transformers) for the model integration.
|
| 334 |
+
* Hugging Face 🤗 for the [Distil-Whisper codebase](https://github.com/huggingface/distil-whisper).
|
| 335 |
+
* [Reazon Human Interaction Lab](https://research.reazon.jp/) for the [ReazonSpeech dataset](https://huggingface.co/datasets/reazon-research/reazonspeech).
|
| 336 |
+
|