语言: 日语
数据集:

• common_voice
  评估指标:
• wer
  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: wav2vec2-live-japanese
  结果:
  • 任务:
    名称: 语音识别
    类型: automatic-speech-recognition
    数据集:
    名称: Common Voice 日语
    类型: common_voice
    参数: ja
    指标:
    • 名称: 测试WER
      类型: wer
      值: 21.48%
    • 名称: 测试CER
      类型: cer
      值: 9.82%

wav2vec2-live-japanese

https://github.com/ttop32/wav2vec2-live-japanese-translator
基于facebook/wav2vec2-large-xlsr-53在日语平假名上微调的模型，使用了以下数据集：

• common_voice
• JSUT
• CSS10
• TEDxJP-10K
• JVS
• JSSS

推理

# 使用方法  
import torch  
import torchaudio  
from datasets import load_dataset  
from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor  
model = Wav2Vec2ForCTC.from_pretrained("ttop324/wav2vec2-live-japanese")  
processor = Wav2Vec2Processor.from_pretrained("ttop324/wav2vec2-live-japanese")  
test_dataset = load_dataset("common_voice", "ja", split="test")  
# 预处理数据集  
# 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"] = torchaudio.functional.resample(speech_array, sampling_rate, 16000)[0].numpy()      
    return batch  
test_dataset = test_dataset.map(speech_file_to_array_fn)  
inputs = processor(test_dataset[:2]["speech"], 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[:2]["sentence"])  

评估

import torch  
import torchaudio  
from datasets import load_dataset, load_metric  
from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor  
import re  
import pykakasi  
import MeCab  
wer = load_metric("wer")  
cer = load_metric("cer")  
model = Wav2Vec2ForCTC.from_pretrained("ttop324/wav2vec2-live-japanese").to("cuda")  
processor = Wav2Vec2Processor.from_pretrained("ttop324/wav2vec2-live-japanese")  
test_dataset = load_dataset("common_voice", "ja", split="test")  
chars_to_ignore_regex = '[\,\?\.\!\-\;\:\"\“\‘\”\�‘、。．！，・―─~｢｣『』\\\\※\[\]\{\}「」〇？…]'  
wakati = MeCab.Tagger("-Owakati")  
kakasi = pykakasi.kakasi()  
kakasi.setMode("J","H")      # 汉字转平假名  
kakasi.setMode("K","H")      # 片假名转平假名  
conv = kakasi.getConverter()  
FULLWIDTH_TO_HALFWIDTH = str.maketrans(  
    '　０１２３４５６７８９ａｂｃｄｅｆｇｈｉｊｋｌｍｎｏｐｑｒｓｔｕｖｗｘｙｚＡＢＣＤＥＦＧＨＩＪＫＬＭＮＯＰＱＲＳＴＵＶＷＸＹＺ！゛＃＄％＆（）＊＋、ー。／：；〈＝〉？＠［］＾＿‘｛｜｝～',  
    ' 0123456789abcdefghijklmnopqrstuvwxyzABCDEFGHIJKLMNOPQRSTUVWXYZ!"#$%&()*+,-./:;<=>?@[]^_`{|}~',  
)  
def fullwidth_to_halfwidth(s):  
    return s.translate(FULLWIDTH_TO_HALFWIDTH)  
def preprocessData(batch):  
    batch["sentence"] = fullwidth_to_halfwidth(batch["sentence"])  
    batch["sentence"] = re.sub(chars_to_ignore_regex,' ', batch["sentence"]).lower()  # 移除特殊字符  
    batch["sentence"] = wakati.parse(batch["sentence"])                              # 添加空格  
    batch["sentence"] = conv.do(batch["sentence"])                                   # 转换为平假名  
    batch["sentence"] = " ".join(batch["sentence"].split())+" "                         # 移除多余空格  
      
    speech_array, sampling_rate = torchaudio.load(batch["path"])  
    batch["speech"] = torchaudio.functional.resample(speech_array, sampling_rate, 16000)[0].numpy()      
    return batch  
test_dataset = test_dataset.map(preprocessData)  
# 预处理数据集  
# 需要将音频文件读取为数组  
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"])))  
print("CER: {:2f}".format(100 * cer.compute(predictions=result["pred_strings"], references=result["sentence"])))