jonatasgrosman/wav2vec2-large-xlsr-53-japanese

1---2language: ja3datasets:4- common_voice5metrics:6- wer7- cer8tags:9- audio10- automatic-speech-recognition11- speech12- xlsr-fine-tuning-week13license: apache-2.014model-index:15- name: XLSR Wav2Vec2 Japanese by Jonatas Grosman16  results:17  - task: 18      name: Speech Recognition19      type: automatic-speech-recognition20    dataset:21      name: Common Voice ja22      type: common_voice23      args: ja24    metrics:25       - name: Test WER26         type: wer27         value: 81.8028       - name: Test CER29         type: cer30         value: 20.1631---32 33# Fine-tuned XLSR-53 large model for speech recognition in Japanese34 35Fine-tuned [facebook/wav2vec2-large-xlsr-53](https://huggingface.co/facebook/wav2vec2-large-xlsr-53) on Japanese using the train and validation splits of [Common Voice 6.1](https://huggingface.co/datasets/common_voice), [CSS10](https://github.com/Kyubyong/css10) and [JSUT](https://sites.google.com/site/shinnosuketakamichi/publication/jsut).36When using this model, make sure that your speech input is sampled at 16kHz.37 38This model has been fine-tuned thanks to the GPU credits generously given by the [OVHcloud](https://www.ovhcloud.com/en/public-cloud/ai-training/) :)39 40The script used for training can be found here: https://github.com/jonatasgrosman/wav2vec2-sprint41 42## Usage43 44The model can be used directly (without a language model) as follows...45 46Using the [HuggingSound](https://github.com/jonatasgrosman/huggingsound) library:47 48```python49from huggingsound import SpeechRecognitionModel50 51model = SpeechRecognitionModel("jonatasgrosman/wav2vec2-large-xlsr-53-japanese")52audio_paths = ["/path/to/file.mp3", "/path/to/another_file.wav"]53 54transcriptions = model.transcribe(audio_paths)55```56 57Writing your own inference script:58 59```python60import torch61import librosa62from datasets import load_dataset63from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor64 65LANG_ID = "ja"66MODEL_ID = "jonatasgrosman/wav2vec2-large-xlsr-53-japanese"67SAMPLES = 1068 69test_dataset = load_dataset("common_voice", LANG_ID, split=f"test[:{SAMPLES}]")70 71processor = Wav2Vec2Processor.from_pretrained(MODEL_ID)72model = Wav2Vec2ForCTC.from_pretrained(MODEL_ID)73 74# Preprocessing the datasets.75# We need to read the audio files as arrays76def speech_file_to_array_fn(batch):77    speech_array, sampling_rate = librosa.load(batch["path"], sr=16_000)78    batch["speech"] = speech_array79    batch["sentence"] = batch["sentence"].upper()80    return batch81 82test_dataset = test_dataset.map(speech_file_to_array_fn)83inputs = processor(test_dataset["speech"], sampling_rate=16_000, return_tensors="pt", padding=True)84 85with torch.no_grad():86    logits = model(inputs.input_values, attention_mask=inputs.attention_mask).logits87 88predicted_ids = torch.argmax(logits, dim=-1)89predicted_sentences = processor.batch_decode(predicted_ids)90 91for i, predicted_sentence in enumerate(predicted_sentences):92    print("-" * 100)93    print("Reference:", test_dataset[i]["sentence"])94    print("Prediction:", predicted_sentence)95```96 97| Reference  | Prediction |98| ------------- | ------------- |99| 祖母は、おおむね機嫌よく、サイコロをころがしている。 | 人母は重にきね起くさいがしている |100| 財布をなくしたので、交番へ行きます。 | 財布をなく手端ので勾番へ行きます |101| 飲み屋のおやじ、旅館の主人、医者をはじめ、交際のある人にきいてまわったら、みんな、私より収入が多いはずなのに、税金は安い。 | ノ宮屋のお親じ旅館の主に医者をはじめ交際のアル人トに聞いて回ったらみんな私より収入が多いはなうに税金は安い |102| 新しい靴をはいて出かけます。 | だらしい靴をはいて出かけます |103| このためプラズマ中のイオンや電子の持つ平均運動エネルギーを温度で表現することがある | このためプラズマ中のイオンや電子の持つ平均運動エネルギーを温度で表弁することがある |104| 松井さんはサッカーより野球のほうが上手です。 | 松井さんはサッカーより野球のほうが上手です |105| 新しいお皿を使います。 | 新しいお皿を使います |106| 結婚以来三年半ぶりの東京も、旧友とのお酒も、夜行列車も、駅で寝て、朝を待つのも久しぶりだ。 | 結婚ル二来三年半降りの東京も吸とのお酒も野越者も駅で寝て朝を待つの久しぶりた |107| これまで、少年野球、ママさんバレーなど、地域スポーツを支え、市民に密着してきたのは、無数のボランティアだった。 | これまで少年野球<unk>三バレーなど地域スポーツを支え市民に満着してきたのは娘数のボランティアだった |108| 靴を脱いで、スリッパをはきます。 | 靴を脱いでスイパーをはきます |109 110## Evaluation111 112The model can be evaluated as follows on the Japanese test data of Common Voice.113 114```python115import torch116import re117import librosa118from datasets import load_dataset, load_metric119from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor120 121LANG_ID = "ja"122MODEL_ID = "jonatasgrosman/wav2vec2-large-xlsr-53-japanese"123DEVICE = "cuda"124 125CHARS_TO_IGNORE = [",", "?", "¿", ".", "!", "¡", ";", "；", ":", '""', "%", '"', "�", "ʿ", "·", "჻", "~", "՞",126                   "؟", "،", "।", "॥", "«", "»", "„", "“", "”", "「", "」", "‘", "’", "《", "》", "(", ")", "[", "]",127                   "{", "}", "=", "`", "_", "+", "<", ">", "…", "–", "°", "´", "ʾ", "‹", "›", "©", "®", "—", "→", "。",128                   "、", "﹂", "﹁", "‧", "～", "﹏", "，", "｛", "｝", "（", "）", "［", "］", "【", "】", "‥", "〽",129                   "『", "』", "〝", "〟", "⟨", "⟩", "〜", "：", "！", "？", "♪", "؛", "/", "\\", "º", "−", "^", "'", "ʻ", "ˆ"]130 131test_dataset = load_dataset("common_voice", LANG_ID, split="test")132 133wer = load_metric("wer.py") # https://github.com/jonatasgrosman/wav2vec2-sprint/blob/main/wer.py134cer = load_metric("cer.py") # https://github.com/jonatasgrosman/wav2vec2-sprint/blob/main/cer.py135 136chars_to_ignore_regex = f"[{re.escape(''.join(CHARS_TO_IGNORE))}]"137 138processor = Wav2Vec2Processor.from_pretrained(MODEL_ID)139model = Wav2Vec2ForCTC.from_pretrained(MODEL_ID)140model.to(DEVICE)141 142# Preprocessing the datasets.143# We need to read the audio files as arrays144def speech_file_to_array_fn(batch):145    with warnings.catch_warnings():146        warnings.simplefilter("ignore")147        speech_array, sampling_rate = librosa.load(batch["path"], sr=16_000)148    batch["speech"] = speech_array149    batch["sentence"] = re.sub(chars_to_ignore_regex, "", batch["sentence"]).upper()150    return batch151 152test_dataset = test_dataset.map(speech_file_to_array_fn)153 154# Preprocessing the datasets.155# We need to read the audio files as arrays156def evaluate(batch):157    inputs = processor(batch["speech"], sampling_rate=16_000, return_tensors="pt", padding=True)158 159    with torch.no_grad():160        logits = model(inputs.input_values.to(DEVICE), attention_mask=inputs.attention_mask.to(DEVICE)).logits161 162    pred_ids = torch.argmax(logits, dim=-1)163    batch["pred_strings"] = processor.batch_decode(pred_ids)164    return batch165 166result = test_dataset.map(evaluate, batched=True, batch_size=8)167 168predictions = [x.upper() for x in result["pred_strings"]]169references = [x.upper() for x in result["sentence"]]170 171print(f"WER: {wer.compute(predictions=predictions, references=references, chunk_size=1000) * 100}")172print(f"CER: {cer.compute(predictions=predictions, references=references, chunk_size=1000) * 100}")173```174 175**Test Result**:176 177In the table below I report the Word Error Rate (WER) and the Character Error Rate (CER) of the model. I ran the evaluation script described above on other models as well (on 2021-05-10). Note that the table below may show different results from those already reported, this may have been caused due to some specificity of the other evaluation scripts used.178 179| Model | WER | CER |180| ------------- | ------------- | ------------- |181| jonatasgrosman/wav2vec2-large-xlsr-53-japanese | **81.80%** | **20.16%** |182| vumichien/wav2vec2-large-xlsr-japanese | 1108.86% | 23.40% |183| qqhann/w2v_hf_jsut_xlsr53 | 1012.18% | 70.77% |184 185## Citation186If you want to cite this model you can use this:187 188```bibtex189@misc{grosman2021xlsr53-large-japanese,190  title={Fine-tuned {XLSR}-53 large model for speech recognition in {J}apanese},191  author={Grosman, Jonatas},192  howpublished={\url{https://huggingface.co/jonatasgrosman/wav2vec2-large-xlsr-53-japanese}},193  year={2021}194}195```