train.py

# pip install numpy==1.26.4 torchaudio==0.13.1 transformers==4.45.2 datasets==3.0.1 librosa==0.10.2.post1 jiwer==3.0.4 bitsandbytes==0.38.0.post1 accelerate==0.26.0

"""
rm -rf ~/.cache/huggingface/datasets/dlxjj___common_voice_17_0 && 
rm -rf ~/.cache/huggingface/modules/datasets_modules/datasets/dlxjj--common_voice_17_0 && 
rm -rf ~/.cache/huggingface/modules/datasets_modules/datasets/dlxjj--common_voice_17_0.lock && 
rm -rf ~/.cache/huggingface/datasets/downloads && 
rm -rf ~/.cache/huggingface/hub/datasets--dlxjj--common_voice_17_0

"""
	# 更新数据集后先删除缓存

dataset_name = "dlxjj/common_voice_17_0"
language = "Japanese"
language_abbr = "ja"

from datasets import load_dataset, DatasetDict

common_voice = DatasetDict()

common_voice["train"] = load_dataset(dataset_name, language_abbr, split="train", trust_remote_code=True)

print(common_voice["train"][0])

array = common_voice["train"][0]["audio"]["array"]

import torch, numpy as np, torchaudio
 
tensor = torch.tensor(np.array( array, dtype='float32'))

sampling_rate = 48000
torchaudio.save("t.wav", tensor.unsqueeze(0), sampling_rate, bits_per_sample=16)

pass