convert.py

import pandas as pd
import os
import torch
import torchaudio
import soundfile as sf
from tqdm import tqdm

target_sample_rate = 16000

df = pd.read_parquet('data.parquet')

# Rename columns
df = df.rename(columns={
    'query_audio': 'query_audio_path',
    'doc_audio': 'document_audio_path',
    'query_duration': 'query_audio_duration',
    'doc_duration': 'document_audio_duration',
})

# Get base directory (where convert.py is located)
base_dir = os.path.dirname(os.path.abspath(__file__)) if '__file__' in globals() else os.getcwd()

def check_and_resample_audio(audio_path, base_dir):
    """
    检查音频文件的采样率，如果不是16kHz则重采样
    
    Args:
        audio_path: 音频文件的相对路径（在parquet中存储的路径）
        base_dir: 基础目录路径
    
    Returns:
        (actual_sample_rate, needs_resample): 实际采样率和是否需要重采样
    """
    if pd.isna(audio_path) or not audio_path or not audio_path.strip():
        return None, False
    
    # 构建完整路径
    full_path = os.path.join(base_dir, audio_path)
    
    if not os.path.exists(full_path):
        print(f"警告: 音频文件不存在: {full_path}")
        return None, False
    
    try:
        # 读取音频信息
        info = sf.info(full_path)
        actual_sample_rate = info.samplerate
        
        # 检查是否需要重采样
        if actual_sample_rate != target_sample_rate:
            # 需要重采样
            print(f"重采样: {audio_path} ({actual_sample_rate}Hz -> {target_sample_rate}Hz)")
            
            # 加载音频
            waveform, sample_rate = torchaudio.load(full_path)
            
            # 转换为单声道（如果是立体声）
            if waveform.shape[0] > 1:
                waveform = torch.mean(waveform, dim=0, keepdim=True)
            
            # 重采样
            resampler = torchaudio.transforms.Resample(sample_rate, target_sample_rate)
            waveform_resampled = resampler(waveform)
            
            # 转换为numpy并确保在[-1, 1]范围内
            waveform_np = waveform_resampled.squeeze().numpy()
            waveform_np = waveform_np.clip(-1.0, 1.0)
            
            # 保存为16kHz, 16-bit WAV
            sf.write(full_path, waveform_np, target_sample_rate, subtype='PCM_16')
            
            return target_sample_rate, True
        else:
            # 采样率正确，不需要重采样
            return actual_sample_rate, False
            
    except Exception as e:
        print(f"错误: 处理音频文件 {full_path} 时出错: {e}")
        return None, False

# 处理query音频文件
print("检查并重采样query音频文件...")
query_sample_rates = []
query_resampled_count = 0

for idx, audio_path in enumerate(tqdm(df['query_audio_path'], desc="处理query音频")):
    sample_rate, was_resampled = check_and_resample_audio(audio_path, base_dir)
    query_sample_rates.append(sample_rate if sample_rate else target_sample_rate)
    if was_resampled:
        query_resampled_count += 1

df['query_audio_sample_rate'] = query_sample_rates

# 处理document音频文件
print("检查并重采样document音频文件...")
doc_sample_rates = []
doc_resampled_count = 0

for idx, audio_path in enumerate(tqdm(df['document_audio_path'], desc="处理document音频")):
    sample_rate, was_resampled = check_and_resample_audio(audio_path, base_dir)
    doc_sample_rates.append(sample_rate if sample_rate else target_sample_rate)
    if was_resampled:
        doc_resampled_count += 1

df['document_audio_sample_rate'] = doc_sample_rates

# 保存更新后的parquet
df.to_parquet('data.parquet', index=False)

print(f'\n完成!')
print(f'Query音频: 重采样了 {query_resampled_count}/{len(df)} 个文件')
print(f'Document音频: 重采样了 {doc_resampled_count}/{len(df)} 个文件')
print(f'新列: {df.columns.tolist()}')
print(f'形状: {df.shape}')
print(df.head(2))