Views
No views yet
<|uyghur|> (Token ID: 51865)pip install transformers torch soundfile numpy1from transformers import WhisperProcessor, WhisperForConditionalGeneration
2import soundfile as sf
3import numpy as np
4
5# 加载模型
6model_name = "anke01/whisper-small-uyghur"
7processor = WhisperProcessor.from_pretrained(model_name)
8model = WhisperForConditionalGeneration.from_pretrained(model_name)
9
10# 加载音频
11waveform, sample_rate = sf.read("audio.wav", dtype='float32')
12
13# 转为单声道
14if len(waveform.shape) > 1:
15 waveform = np.mean(waveform, axis=1)
16
17# 重采样到16kHz
18if sample_rate != 16000:
19 old_len = len(waveform)
20 new_len = int(old_len * 16000 / sample_rate)
21 waveform = np.interp(
22 np.linspace(0, old_len - 1, new_len),
23 np.arange(old_len),
24 waveform
25 )
26
27# 处理音频
28inputs = processor(waveform, sampling_rate=16000, return_tensors="pt")
29
30# 使用维吾尔语token生成
31lang_token_id = processor.tokenizer.convert_tokens_to_ids("<|uyghur|>")
32forced_decoder_ids = [
33 (1, lang_token_id),
34 (2, processor.tokenizer.convert_tokens_to_ids("<|transcribe|>")),
35]
36
37predicted_ids = model.generate(
38 inputs.input_features,
39 forced_decoder_ids=forced_decoder_ids
40)
41
42# 解码
43transcription = processor.batch_decode(predicted_ids, skip_special_tokens=True)[0]
44print(transcription)1# 转录音频文件
2python inference.py audio.wav
3
4# 指定模型路径
5python inference.py audio.wav ./model1# 在数据集上评估
2python evaluate.py data/val.json
3
4# 指定模型路径
5python evaluate.py data/val.json ./model
6
7# 只评估前100条
8python evaluate.py data/val.json ./model 1001# 自动修复模型配置
2python fix_model_config.py <model_path><|uyghur|>).
├── model.safetensors # 模型权重 (925MB)
├── vocab.json # 分词器词汇表
├── tokenizer_config.json # 分词器配置
├── preprocessor_config.json # 音频预处理器配置
├── config.json # 模型配置
├── generation_config.json # 生成配置 (包含语言映射)
├── language_info.json # 语言元数据
├── inference.py # 推理脚本
├── evaluate.py # 评估脚本
├── fix_model_config.py # 配置修复脚本
└── README.md # 本文件thuyg20_00000_*.wav (8.30秒)بىز نەچچە يىل تىرىشىش ئارقىلىق بۇ قاقاس يەرنى ئاخىرى مۇنبەت ئىتىزغا ئايلاندۇردۇق1@misc{whisper-uyghur,
2 title={Whisper Small Fine-tuned for Uyghur},
3 author={Your Name},
4 year={2024},
5 publisher={Hugging Face}
6}