语言: tt
数据集:

• common_voice
  指标:
• wer
  标签:
• audio
• automatic-speech-recognition
• speech
• xlsr-fine-tuning-week
  许可证: apache-2.0
  模型索引:
• 名称: Anton Lozhkov 的鞑靼语 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 tt
    类型: common_voice
    参数: tt
    指标:
    • 名称: 测试 WER
      类型: wer
      值: 26.76

Wav2Vec2-Large-XLSR-53-鞑靼语

基于 facebook/wav2vec2-large-xlsr-53 在鞑靼语上使用 Common Voice 数据集进行微调。
使用此模型时，请确保您的语音输入采样率为 16kHz。

使用方法

该模型可以直接使用（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", "tt", split="test[:2%]")  

processor = Wav2Vec2Processor.from_pretrained("anton-l/wav2vec2-large-xlsr-53-tatar")  
model = Wav2Vec2ForCTC.from_pretrained("anton-l/wav2vec2-large-xlsr-53-tatar")  

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

# 预处理数据集  
# 我们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:", processor.batch_decode(predicted_ids))  
print("参考:", test_dataset["sentence"][:2])  

评估

可以在 Common Voice 的鞑靼语测试数据上按如下方式评估模型。

import torch  
import torchaudio  
import urllib.request  
import tarfile  
import pandas as pd  
from tqdm.auto import tqdm  
from datasets import load_metric  
from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor  

# 下载原始数据以节省磁盘空间，而非使用 HF 数据集  
data_url = "https://voice-prod-bundler-ee1969a6ce8178826482b88e843c335139bd3fb4.s3.amazonaws.com/cv-corpus-6.1-2020-12-11/tt.tar.gz"  
filestream = urllib.request.urlopen(data_url)  
data_file = tarfile.open(fileobj=filestream, mode="r|gz")  
data_file.extractall()  

wer = load_metric("wer")  

processor = Wav2Vec2Processor.from_pretrained("anton-l/wav2vec2-large-xlsr-53-tatar")  
model = Wav2Vec2ForCTC.from_pretrained("anton-l/wav2vec2-large-xlsr-53-tatar")  
model.to("cuda")  

cv_test = pd.read_csv("cv-corpus-6.1-2020-12-11/tt/test.tsv", sep='\t')  
clips_path = "cv-corpus-6.1-2020-12-11/tt/clips/"  

def clean_sentence(sent):  
    sent = sent.lower()  
    # 'ё' 等同于 'е'  
    sent = sent.replace('ё', 'е')  
    # 将非字母字符替换为空格  
    sent = "".join(ch if ch.isalpha() else " " for ch in sent)  
    # 去除重复空格  
    sent = " ".join(sent.split())  
    return sent  

targets = []  
preds = []  

for i, row in tqdm(cv_test.iterrows(), total=cv_test.shape[0]):  
    row["sentence"] = clean_sentence(row["sentence"])  
    speech_array, sampling_rate = torchaudio.load(clips_path + row["path"])  
    resampler = torchaudio.transforms.Resample(sampling_rate, 16_000)  
    row["speech"] = resampler(speech_array).squeeze().numpy()  

    inputs = processor(row["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)  

    targets.append(row["sentence"])  
    preds.append(processor.batch_decode(pred_ids)[0])  

print("WER: {:2f}".format(100 * wer.compute(predictions=preds, references=targets)))  

测试结果: 26.76 %

训练

使用了 Common Voice 的 train 和 validation 数据集进行训练。