语言: te 数据集:

• openslr 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: Anurag Singh XLSR Wav2Vec2 Large 53 泰卢固语 结果:
  • 任务: 名称: 语音识别 类型: automatic-speech-recognition 数据集: 名称: OpenSLR te 类型: openslr 参数: te 指标:
    • 名称: 测试WER 类型: wer 值: 44.98

Wav2Vec2-Large-XLSR-53-泰卢固语

基于facebook/wav2vec2-large-xlsr-53模型，使用OpenSLR SLR66数据集对泰卢固语进行微调。
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
import pandas as pd
# 评估笔记本包含下载数据的流程
df = pd.read_csv("/content/te/test.tsv", sep="\t")
df["path"] = "/content/te/clips/" + df["path"]
test_dataset = Dataset.from_pandas(df)
processor = Wav2Vec2Processor.from_pretrained("anuragshas/wav2vec2-large-xlsr-53-telugu")
model = Wav2Vec2ForCTC.from_pretrained("anuragshas/wav2vec2-large-xlsr-53-telugu") 
resampler = torchaudio.transforms.Resample(48_000, 16_000)
# 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"] = 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

import torch
import torchaudio
from datasets import Dataset, load_metric
from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor
import re
from sklearn.model_selection import train_test_split
import pandas as pd
# 评估笔记本包含下载数据的流程
df = pd.read_csv("/content/te/test.tsv", sep="\t")
df["path"] = "/content/te/clips/" + df["path"]
test_dataset = Dataset.from_pandas(df)
wer = load_metric("wer")
processor = Wav2Vec2Processor.from_pretrained("anuragshas/wav2vec2-large-xlsr-53-telugu")
model = Wav2Vec2ForCTC.from_pretrained("anuragshas/wav2vec2-large-xlsr-53-telugu") 
model.to("cuda")
chars_to_ignore_regex = '[\,\?\.\!\-\_\;\:\"\“\%\‘\”\।\’\'\&]'
resampler = torchaudio.transforms.Resample(48_000, 16_000)
def normalizer(text):
    # 使用自定义规范化处理
    text = text.replace("\\n","\n")
    text = ' '.join(text.split())
    text = re.sub(r'''([a-z]+)''','',text,flags=re.IGNORECASE)
    text = re.sub(r'''%'''," శాతం ", text)
    text = re.sub(r'''(/|-|_)'''," ", text)
    text = re.sub("ై","ై", text)
    text = text.strip()
    return text
def speech_file_to_array_fn(batch):
    batch["sentence"] = normalizer(batch["sentence"])
    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)
# 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"])))

测试结果: 44.98%

训练过程

使用了OpenSLR泰卢固语数据集的70%进行训练。

训练集标注文件点击下载
测试集标注文件点击下载

训练数据准备笔记本查看链接
训练笔记本查看链接
评估笔记本查看链接