如何正确设置帧与步长？（PyAudioAnalysis数组重塑错误问题）

问题场景 #

研究生毕设中使用PyAudioAnalysis分析音频文件，设置帧（frame）与步长（hop）时触发数组重塑错误，代码及运行输出如下：

运行代码

import numpy as np
from pyAudioAnalysis import audioBasicIO
from pyAudioAnalysis import ShortTermFeatures

[Fs, x] = audioBasicIO.read_audio_file("/content/drive/MyDrive/audio.wav")


print("Sample rate (Fs):", Fs)
print("Total samples (x):", len(x))


frame_size = 0.1 * Fs
hop_size = 0.05 * Fs

print("Frame size:", frame_size)
print("Hop size:", hop_size)


try:
    [F, F_names] = ShortTermFeatures.feature_extraction(x, Fs, frame_size, hop_size)
    print("Feature extraction successful")
except ValueError as e:
    print("Error in feature extraction:", e)

运行输出

Sample rate (Fs): 48000
Total samples (x): 478080
Frame size: 4800.0
Hop size: 2400.0
Error in feature extraction: cannot reshape array of size 9600 into shape (480,10)

解决方案 #

错误原因

1. 音频声道问题：报错中的9600 = 4800 * 2，说明目标音频是双声道立体声，而feature_extraction函数默认处理单声道音频，多声道数据会导致内部数组尺寸不匹配。
2. 参数类型问题：帧长和步长代表采样点数量，必须为整数，当前计算得到的是浮点数（如4800.0），会触发内部重塑逻辑错误。

修正后的代码

import numpy as np
from pyAudioAnalysis import audioBasicIO
from pyAudioAnalysis import ShortTermFeatures

[Fs, x] = audioBasicIO.read_audio_file("/content/drive/MyDrive/audio.wav")

# 将立体声转为单声道（取双声道均值）
if len(x.shape) > 1:
    x = np.mean(x, axis=1)

print("Sample rate (Fs):", Fs)
print("Total samples (x):", len(x))

# 强制转换为整数，确保采样点数量为整数
frame_size = int(0.1 * Fs)
hop_size = int(0.05 * Fs)

print("Frame size:", frame_size)
print("Hop size:", hop_size)

try:
    [F, F_names] = ShortTermFeatures.feature_extraction(x, Fs, frame_size, hop_size)
    print("Feature extraction successful")
    print(f"Extracted features shape: {F.shape}")
except ValueError as e:
    print("Error in feature extraction:", e)

验证说明 #

修正后代码完成两个关键处理：将双声道音频合并为单声道，同时将帧长、步长转为整数类型，可解决数组重塑错误，顺利完成特征提取。

内容的提问来源于stack exchange，提问作者박일영