语言: pa-IN
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: danurahul/wav2vec2-large-xlsr-pa-IN
  结果:
  • 任务:
    名称: 语音识别
    类型: automatic-speech-recognition
    数据集:
    名称: Common Voice pa-IN
    类型: common_voice
    参数: pa-IN
    指标:
    • 名称: 测试WER
      类型: wer
      值: 54.86

Wav2Vec2-Large-XLSR-53-旁遮普语

基于facebook/wav2vec2-large-xlsr-53在旁遮普语Common Voice数据集上微调。
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：

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

test_dataset = load_dataset("common_voice", "pa-IN", split="test[:2%]")   

processor = Wav2Vec2Processor.from_pretrained("danurahul/wav2vec2-large-xlsr-pa-IN")   
model = Wav2Vec2ForCTC.from_pretrained("danurahul/wav2vec2-large-xlsr-pa-IN")   

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

该模型可在Common Voice的旁遮普语测试数据上评估如下：

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

test_dataset = load_dataset("common_voice", "pa-IN", split="test")   

wer = load_metric("wer")  

processor = Wav2Vec2Processor.from_pretrained("danurahul/wav2vec2-large-xlsr-pa-IN")   

model = Wav2Vec2ForCTC.from_pretrained("danurahul/wav2vec2-large-xlsr-pa-IN")   

model.to("cuda")  

chars_to_ignore_regex = '[\\\\\\\\\\\\\\\\,\\\\\\\\\\\\\\\\?\\\\\\\\\\\\\\\\.\\\\\\\\\\\\\\\\!\\\\\\\\\\\\\\\\-\\\\\\\\\\\\\\\\;\\\\\\\\\\\\\\\\:\\\\\\\\\\\\\\\\"\\\\\\\\\\\\\\\\“\\\\\\\\\\\\\\\\%\\\\\\\\\\\\\\\\‘\\\\\\\\\\\\\\\\”\\\\\\\\\\\\\\\\�]'    
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()  
  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)  

# 评估函数  
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"])))  

测试结果: 100%

训练

使用Common Voice的train和validation集进行训练和验证

训练脚本详见:
https://github.com/rahul-art/huggingface_wav2vec2_punjabi/blob/main/Fine_Tune_XLSR_Wav2Vec2_on_Punjabi_ASR_with_%F0%9F%A4%97_Transformers.ipynb