Wav2vec2 Large Xlsr Japanese

🚀 Wav2Vec2-Large-XLSR-53-日語版

本項目基於 facebook/wav2vec2-large-xlsr-53 模型，使用 Common Voice 以及東京大學猿渡研究室的日語語音語料庫 JSUT 對日語進行了微調。使用該模型時，請確保輸入的語音採樣率為 16kHz。

🚀 快速開始

本模型可直接使用（無需語言模型），具體使用方法如下文所示。

✨ 主要特性

• 數據集：使用 Common Voice 和東京大學猿渡研究室的日語語音語料庫進行訓練。
• 評估指標：在 Common Voice 日語測試數據上，字錯誤率（WER）為 30.84%，字符錯誤率（CER）為 17.85%。

📦 安裝指南

運行代碼前，需要安裝以下依賴：

!pip install mecab-python3
!pip install unidic-lite
!python -m unidic download

💻 使用示例

基礎用法

# config
wakati = MeCab.Tagger("-Owakati")
chars_to_ignore_regex = '[\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\,\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\、\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\。\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\．\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\「\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\」\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\…\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\？\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\・]'

# load data, processor and model
test_dataset = load_dataset("common_voice", "ja", split="test[:2%]")
processor = Wav2Vec2Processor.from_pretrained("vumichien/wav2vec2-large-xlsr-japanese")
model = Wav2Vec2ForCTC.from_pretrained("vumichien/wav2vec2-large-xlsr-japanese")
resampler = lambda sr, y: librosa.resample(y.numpy().squeeze(), sr, 16_000)

# Preprocessing the datasets.
def speech_file_to_array_fn(batch):
    batch["sentence"] = wakati.parse(batch["sentence"]).strip()
    batch["sentence"] = re.sub(chars_to_ignore_regex,'', batch["sentence"]).strip()
    speech_array, sampling_rate = torchaudio.load(batch["path"])
    batch["speech"] = resampler(sampling_rate, speech_array).squeeze()
    return batch
test_dataset = test_dataset.map(speech_file_to_array_fn)
inputs = processor(test_dataset["speech"][:2], sampling_rate=16_000, return_tensors="pt", padding=True)
with torch.no_grad():
    logits = model(inputs.input_values, attention_mask=inputs.attention_mask).logits
predicted_ids = torch.argmax(logits, dim=-1)
print("Prediction:", processor.batch_decode(predicted_ids))
print("Reference:", test_dataset["sentence"][:2])

高級用法

# 對模型在 Common Voice 日語測試數據上進行評估
import torch
import librosa
import torchaudio
from datasets import load_dataset, load_metric
import MeCab
from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor
import re

#config
wakati = MeCab.Tagger("-Owakati")
chars_to_ignore_regex = '[\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\,\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\、\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\。\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\．\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\「\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\」\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\…\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\？\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\\・]'

# load data, processor and model
test_dataset = load_dataset("common_voice", "ja", split="test")
wer = load_metric("wer")
processor = Wav2Vec2Processor.from_pretrained("vumichien/wav2vec2-large-xlsr-japanese")
model = Wav2Vec2ForCTC.from_pretrained("vumichien/wav2vec2-large-xlsr-japanese")
model.to("cuda")
resampler = lambda sr, y: librosa.resample(y.numpy().squeeze(), sr, 16_000)

# Preprocessing the datasets.
def speech_file_to_array_fn(batch):
    batch["sentence"] = wakati.parse(batch["sentence"]).strip()
    batch["sentence"] = re.sub(chars_to_ignore_regex,'', batch["sentence"]).strip()
    speech_array, sampling_rate = torchaudio.load(batch["path"])
    batch["speech"] = resampler(sampling_rate, speech_array).squeeze()
    return batch
test_dataset = test_dataset.map(speech_file_to_array_fn)

# evaluate function
def evaluate(batch):
    inputs = processor(batch["speech"], sampling_rate=16_000, return_tensors="pt", padding=True)
    with torch.no_grad():
        logits = model(inputs.input_values.to("cuda"), attention_mask=inputs.attention_mask.to("cuda")).logits
    pred_ids = torch.argmax(logits, dim=-1)
    batch["pred_strings"] = processor.batch_decode(pred_ids)
    return batch
result = test_dataset.map(evaluate, batched=True, batch_size=8)
print("WER: {:2f}".format(100 * wer.compute(predictions=result["pred_strings"], references=result["sentence"])))

📚 詳細文檔

測試結果

• 字錯誤率（WER）：30.84%
• 字符錯誤率（CER）：17.85%

訓練信息

訓練使用了 Common Voice 的 train、validation 數據集以及日語語音語料庫 basic5000 數據集。

📄 許可證

本項目採用 Apache-2.0 許可證。