语言: 尼泊尔语
数据集:

• OpenSLR
• Common Voice
  评估指标:
• 词错误率 (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-xlsr-nepali
  结果:
  • 任务:
    名称: 语音识别
    类型: 自动语音识别
    数据集:
    名称: OpenSLR ne
    类型: OpenSLR
    参数: ne
    评估指标:
    • 名称: 测试WER
      类型: wer
      值: 05.97

Wav2Vec2-Large-XLSR-53-尼泊尔语

基于facebook/wav2vec2-large-xlsr-53模型，使用Common Voice和OpenSLR ne数据集对尼泊尔语进行微调。

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  

!wget https://www.openslr.org/resources/43/ne_np_female.zip  
!unzip ne_np_female.zip  
!ls ne_np_female  

colnames=['path','sentence']   
df  = pd.read_csv('/content/ne_np_female/line_index.tsv',sep='\\t',header=None,names = colnames)  
df['path'] = '/content/ne_np_female/wavs/'+df['path'] +'.wav'  

train, test = train_test_split(df, test_size=0.1)  

test.to_csv('/content/ne_np_female/line_index_test.csv')  

test_dataset = load_dataset('csv', data_files='/content/ne_np_female/line_index_test.csv',split = 'train')  

processor = Wav2Vec2Processor.from_pretrained("gagan3012/wav2vec2-xlsr-nepali")  
model = Wav2Vec2ForCTC.from_pretrained("gagan3012/wav2vec2-xlsr-nepali")   

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):  
\tspeech_array, sampling_rate = torchaudio.load(batch["path"])  
\tbatch["speech"] = resampler(speech_array).squeeze().numpy()  
\treturn 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():  
\tlogits = 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])  

结果

预测结果: ['पारानाको ब्राजिली राज्यमा रहेको राजधानी', 'देवराज जोशी त्रिभुवन विश्वविद्यालयबाट शिक्षाशास्त्रमा स्नातक हुनुहुन्छ']

参考文本: ['पारानाको ब्राजिली राज्यमा रहेको राजधानी', 'देवराज जोशी त्रिभुवन विश्वविद्यालयबाट शिक्षाशास्त्रमा स्नातक हुनुहुन्छ']

评估

该模型可在Common Voice的尼泊尔语测试数据上进行如下评估：

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

!wget https://www.openslr.org/resources/43/ne_np_female.zip  
!unzip ne_np_female.zip  
!ls ne_np_female  

colnames=['path','sentence']   
df  = pd.read_csv('/content/ne_np_female/line_index.tsv',sep='\\t',header=None,names = colnames)  
df['path'] = '/content/ne_np_female/wavs/'+df['path'] +'.wav'  

train, test = train_test_split(df, test_size=0.1)  

test.to_csv('/content/ne_np_female/line_index_test.csv')  

test_dataset = load_dataset('csv', data_files='/content/ne_np_female/line_index_test.csv',split = 'train')  
wer = load_metric("wer")  

processor = Wav2Vec2Processor.from_pretrained("gagan3012/wav2vec2-xlsr-nepali")  
model = Wav2Vec2ForCTC.from_pretrained("gagan3012/wav2vec2-xlsr-nepali")   
model.to("cuda")  

chars_to_ignore_regex = '[\\,\\?\\.\\!\\-\\;\\:\\"\\“]'    
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):  
\tbatch["sentence"] = re.sub(chars_to_ignore_regex, '', batch["sentence"]).lower()  
\tspeech_array, sampling_rate = torchaudio.load(batch["path"])  
\tbatch["speech"] = resampler(speech_array).squeeze().numpy()  
\treturn batch  

test_dataset = test_dataset.map(speech_file_to_array_fn)  

# 评估函数  
def evaluate(batch):  
\tinputs = processor(batch["speech"], sampling_rate=16_000, return_tensors="pt", padding=True)  

\twith torch.no_grad():  
\t\tlogits = model(inputs.input_values.to("cuda"), attention_mask=inputs.attention_mask.to("cuda")).logits  

\tpred_ids = torch.argmax(logits, dim=-1)  
\tbatch["pred_strings"] = processor.batch_decode(pred_ids)  
\treturn 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: 05.97%

训练

训练脚本可在此处找到。