语言: 日语
数据集:

• common_voice
• jsut
  评估指标:
• wer（词错误率）
• cer（字符错误率）
  标签:
• 音频
• 自动语音识别
• 语音
• xlsr微调周
  许可证: apache-2.0
  模型索引:
• 名称: Japanese XLSR Wav2Vec2 Large 53
  结果:
  • 任务:
    名称: 语音识别
    类型: automatic-speech-recognition
    数据集:
    名称: Common Voice ja
    类型: common_voice
    参数: ja
    指标:
    • 名称: 测试WER
      类型: wer
      值: 51.72
    • 名称: 测试CER
      类型: cer
      值: 24.89

Wav2Vec2-Large-XLSR-53-Japanese

基于facebook/wav2vec2-large-xlsr-53模型，使用Common Voice和JSUT数据集对日语进行了微调。
使用此模型时，请确保语音输入采样率为16kHz。

使用方法

该模型可直接使用（无需语言模型），如下所示：

import torch  
import torchaudio  
from datasets import load_dataset  
from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor  

test_dataset = load_dataset("common_voice", "ja", split="test[:2%]")  

processor = Wav2Vec2Processor.from_pretrained("qqhann/w2v_hf_jsut_xlsr53")  
model = Wav2Vec2ForCTC.from_pretrained("qqhann/w2v_hf_jsut_xlsr53")  

resampler = torchaudio.transforms.Resample(48_000, 16_000)  

# 数据预处理  
# 需要将音频文件读取为数组  
def speech_file_to_array_fn(batch):  
    speech_array, sampling_rate = torchaudio.load(batch["path"])  
    batch["speech"] = resampler(speech_array).squeeze().numpy()  
    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("预测结果:", processor.batch_decode(predicted_ids))  
print("参考文本:", test_dataset["sentence"][:2])  

评估

可按如下方式在Common Voice的日语测试数据上评估模型性能：

!pip install torchaudio  
!pip install datasets transformers  
!pip install jiwer  
!pip install mecab-python3  
!pip install unidic-lite  
!python -m unidic download  
!pip install jaconv  

import torch  
import torchaudio  
from datasets import load_dataset, load_metric  
from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor  
import re  
import MeCab  
from jaconv import kata2hira  
from typing import List  

# 日语预处理  
tagger = MeCab.Tagger("-Owakati")  
chars_to_ignore_regex = '[\。\、\「\」\,\?\.\!\-\;\:\"\“\%\‘\”\�]'  

def text2kata(text):  
    node = tagger.parseToNode(text)  
    word_class = []  
    while node:  
        word = node.surface  
        wclass = node.feature.split(',')  
        if wclass[0] != u'BOS/EOS':  
            if len(wclass) <= 6:  
                word_class.append((word))  
            elif wclass[6] == None:  
                word_class.append((word))  
            else:  
                word_class.append((wclass[6]))  
        node = node.next  
    return ' '.join(word_class)  

def hiragana(text):  
    return kata2hira(text2kata(text))  

test_dataset = load_dataset("common_voice", "ja", split="test")  
wer = load_metric("wer")  
resampler = torchaudio.transforms.Resample(48_000, 16_000) # JSUT已为16kHz  

processor = Wav2Vec2Processor.from_pretrained("qqhann/w2v_hf_jsut_xlsr53")  
model = Wav2Vec2ForCTC.from_pretrained("qqhann/w2v_hf_jsut_xlsr53")  
model.to("cuda")  

# 数据预处理  
def speech_file_to_array_fn(batch):  
    batch["sentence"] = hiragana(batch["sentence"]).strip()  
    batch["sentence"] = re.sub(chars_to_ignore_regex, '', batch["sentence"]).lower()  
    speech_array, sampling_rate = torchaudio.load(batch["path"])  
    batch["speech"] = resampler(speech_array).squeeze().numpy()  
    return batch  

test_dataset = test_dataset.map(speech_file_to_array_fn)  

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)  

def cer_compute(predictions: List[str], references: List[str]):  
    p = [" ".join(list(" " + pred.replace(" ", ""))).strip() for pred in predictions]  
    r = [" ".join(list(" " + ref.replace(" ", ""))).strip() for ref in references]  
    return wer.compute(predictions=p, references=r)  

print("WER: {:2f}".format(100 * wer.compute(predictions=result["pred_strings"], references=result["sentence"])))  
print("CER: {:2f}".format(100 * cer_compute(predictions=result["pred_strings"], references=result["sentence"])))  

测试结果: 51.72%

训练过程

训练使用了私有收集的JSUT日语数据集。