微调Wav2Vec2时Librosa音频重采样遇参数错误，求解决

错误原因 #

你遇到的resample() takes 1 positional argument but 3 were given错误，是因为新版Librosa（0.9.0及以上）修改了resample函数的参数结构——不再支持直接传入三个位置参数，必须显式指定orig_sr（原采样率）和target_sr（目标采样率）作为关键字参数。

修复方案（Librosa版本适配） #

修改代码中调用librosa.resample的部分，将位置参数改为关键字参数即可解决：

修复后的重采样代码 #

import librosa
import numpy as np

# List to store new sampling rates
new_sr = []

# Resample each audio signal and store the new sampling rate
for i in range(len(database['psr'])):
    try:
        audio_signal = np.asarray(database['audio'][i])
        original_sr = database['psr'][i]

        if audio_signal.ndim == 1:
            # 显式指定orig_sr和target_sr关键字参数
            resampled_audio = librosa.resample(audio_signal, orig_sr=original_sr, target_sr=16000)
        else:
            resampled_channels = []
            for channel in audio_signal:
                resampled_channel = librosa.resample(channel, orig_sr=original_sr, target_sr=16000)
                resampled_channels.append(resampled_channel)
            resampled_audio = np.array(resampled_channels)

        database['audio'][i] = resampled_audio
        new_sr.append(16000)
    except Exception as e:
        print(f"Error processing audio at index {i}: {e}")

database['newsr'] = new_sr

替代方案：用Torchaudio直接重采样（更适配你的流程） #

既然你已经用torchaudio.load加载音频，推荐直接用Torchaudio的resample函数，避免Librosa版本兼容问题，且全程保持张量操作更高效：

import torch
import torchaudio.transforms as T

new_sr = []
resampler = T.Resample(orig_freq=None, new_freq=16000)

for i in range(len(database['psr'])):
    try:
        # 把numpy数组转回张量
        audio_tensor = torch.from_numpy(database['audio'][i]).unsqueeze(0)
        original_sr = database['psr'][i]
        
        # 更新重采样器的原采样率
        resampler.orig_freq = original_sr
        resampled_tensor = resampler(audio_tensor)
        
        # 转回numpy数组存回database
        database['audio'][i] = resampled_tensor.squeeze(0).numpy()
        new_sr.append(16000)
    except Exception as e:
        print(f"Error processing audio at index {i}: {e}")

database['newsr'] = new_sr

额外提示 #

• 若必须使用旧版Librosa参数格式，可降级Librosa到0.8.x版本：pip install librosa==0.8.1，但不推荐，新版本修复了更多音频处理问题。
• Wav2Vec2要求输入为16kHz单声道音频，你的加载代码已经取了speech_array[0]，确保了单声道输入，无需额外处理。

内容的提问来源于stack exchange，提问作者SAFIQUL ISLAM UZZAL 203-15-144