语言: 德语
数据集:

• common_voice
• 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: 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 德语
    类型: common_voice
    参数: de
    指标:
    • 名称: 测试WER
      类型: wer
      值: 15.80

Wav2Vec2-Large-XLSR-53-德语

基于facebook/wav2vec2-large-xlsr-53在Common Voice德语数据集上微调。
使用该模型时，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：

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

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

processor = Wav2Vec2Processor.from_pretrained("marcel/wav2vec2-large-xlsr-53-german")  
model = Wav2Vec2ForCTC.from_pretrained("marcel/wav2vec2-large-xlsr-53-german")   

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

# 数据预处理  
# 需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

该模型可按如下方式在Common Voice德语测试数据上评估：

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

test_dataset = load_dataset("common_voice", "de", split="test")  
wer = load_metric("wer")  

processor = Wav2Vec2Processor.from_pretrained("marcel/wav2vec2-large-xlsr-53-german")  
model = Wav2Vec2ForCTC.from_pretrained("marcel/wav2vec2-large-xlsr-53-german")   
model.to("cuda")  

chars_to_ignore_regex = '[\,\?\.\!\-\;\:\"\“\%\”\�\カ\æ\無\ན\カ\臣\ѹ\…\«\»\ð\ı\„\幺\א\ב\比\ш\ע\)\ứ\в\œ\ч\+\—\ш\‚\נ\м\ń\乡\$\=\ש\ф\支\(\°\и\к\̇]'  
substitutions = {  
    'e' : '[\ə\é\ě\ę\ê\ế\ế\ë\ė\е]',  
    'o' : '[\ō\ô\ô\ó\ò\ø\ọ\ŏ\õ\ő\о]',  
    'a' : '[\á\ā\ā\ă\ã\å\â\à\ą\а]',  
    'c' : '[\č\ć\ç\с]',  
    'l' : '[\ł]',  
    'u' : '[\ú\ū\ứ\ů]',  
    'und' : '[\&]',  
    'r' : '[\ř]',  
    'y' : '[\ý]',  
    's' : '[\ś\š\ș\ş]',  
    'i' : '[\ī\ǐ\í\ï\î\ï]',  
    'z' : '[\ź\ž\ź\ż]',  
    'n' : '[\ñ\ń\ņ]',  
    'g' : '[\ğ]',  
    'ss' : '[\ß]',  
    't' : '[\ț\ť]',  
    'd' : '[\ď\đ]',  
    "'": '[\ʿ\་\’\`\´\ʻ\`\‘]',  
    'p': '\р'  
}  
resampler = torchaudio.transforms.Resample(48_000, 16_000)  

# 数据预处理  
def speech_file_to_array_fn(batch):  
    batch["sentence"] = re.sub(chars_to_ignore_regex, '', batch["sentence"]).lower()  
    for x in substitutions:  
        batch["sentence"] = re.sub(substitutions[x], x, batch["sentence"])  
        speech_array, sampling_rate = torchaudio.load(batch["path"])  
    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)  

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

该模型也可分10%片段评估以降低资源消耗（待测试）：

import jiwer  
# ...（其余代码与完整评估相同）  

H, S, D, I = 0, 0, 0, 0  
for i in range(10):  
    print("test["+str(10*i)+"%:"+str(10*(i+1))+"%]")  
    test_dataset = load_dataset("common_voice", "de", split="test["+str(10*i)+"%:"+str(10*(i+1))+"%]")  
    test_dataset = test_dataset.map(speech_file_to_array_fn)  
    result = test_dataset.map(evaluate, batched=True, batch_size=8)  
    predictions = result["pred_strings"]  
    targets = result["sentence"]  
    chunk_metrics = jiwer.compute_measures(targets, predictions)  
    H += chunk_metrics["hits"]  
    S += chunk_metrics["substitutions"]  
    D += chunk_metrics["deletions"]  
    I += chunk_metrics["insertions"]  
WER = float(S + D + I) / float(H + S + D)  
print("WER: {:2f}".format(WER*100))  

测试结果: 15.80%

训练

使用Common Voice训练集的前50%和验证集的12%进行训练（前12%数据训练30轮，剩余数据训练3轮）。