ValueError：使用 MFCC 提取特征时，无法使用“常量”或“空”以外的模式扩展空轴 0

问题描述

我正在研究语音识别系统，但在提取信号特征时遇到问题

这是我读取音频文件和标签并提取特征的预处理块：

audio_path = r"C:\Users\Salma\Downloads\TORGO"
files = os.listdir(audio_path)
all_wave = []
all_label = []
mfcc_features = []
ignored = {"Notes","phn_arrayMic","amps","rawpos","wavall","pos","alignment","Cpcmd","EMA","VIDEO","wav_headMic","normpos","log"}

for file in files:
    for subfile in os.listdir(audio_path + '/'):
        
        for session in [s for s in os.listdir(audio_path + '/' + subfile +'/')if s not in ignored]:
            for subsession in [x for x in os.listdir(audio_path + '/'+ subfile +'/'+ session + '/')if x not in ignored]:
                for data in [y for y in os.listdir(audio_path + '/'+ subfile +'/'+ session + '/' +subsession +'/') if y not in ignored]:
                    
                    waves = [f for f in os.listdir(audio_path + '/' + subfile +'/'+ session +'/' + subsession +'/' + data + '/')if f.endswith('.wav')]
                    labels = [L for L in os.listdir(audio_path + '/' + subfile +'/'+ session +'/' + subsession + '/' + data + '/')if L.endswith('.txt')]

                    for wav in waves:
                        samples,sample_rate = librosa.load(audio_path + '/' + subfile +'/'+ session + '/'+subsession +'/'+ data +'/'+ wav,sr = 16000)
                        samples = librosa.resample(samples,sample_rate,16000)
                        mfcc = librosa.feature.mfcc(samples,n_mfcc=13)
                        mfcc = mfcc.T                       
                        all_wave.append(samples)
                        mfcc_features.append(mfcc.tolist())
                    for label in labels:
                        file_path = audio_path + '/' + subfile + '/' + session + '/' + subsession + '/' + data +'/'+ label
                        l = open(file_path,"r")
                        for target in l:
                        
                            all_label.append(target.strip())

如果我删除了 mfcc = librosa.feature.mfcc，代码工作正常

这里是返回错误：

ValueError                                Traceback (most recent call last)
<ipython-input-13-fd45bf295b2c> in <module>
     12                         samples,sr = 16000)
     13                         samples = librosa.resample(samples,16000)
---> 14                         mfcc = librosa.feature.mfcc(samples,n_mfcc=13)
     15                         mfcc = mfcc.T
     16                         all_wave.append(samples)

C:\Users\Salma\Anaconda3\lib\site-packages\librosa\feature\spectral.py in mfcc(y,sr,S,n_mfcc,dct_type,norm,lifter,**kwargs)
   1850 
   1851     if S is None:
-> 1852         S = power_to_db(melspectrogram(y=y,sr=sr,**kwargs))
   1853 
   1854     M = scipy.fftpack.dct(S,axis=0,type=dct_type,norm=norm)[:n_mfcc]

C:\Users\Salma\Anaconda3\lib\site-packages\librosa\feature\spectral.py in melspectrogram(y,n_fft,hop_length,win_length,window,center,pad_mode,power,**kwargs)
   2003         window=window,2004         center=center,-> 2005         pad_mode=pad_mode,2006     )
   2007 

C:\Users\Salma\Anaconda3\lib\site-packages\librosa\core\spectrum.py in _spectrogram(y,pad_mode)
   2517                     center=center,2518                     window=window,-> 2519                     pad_mode=pad_mode,2520                 )
   2521             )

C:\Users\Salma\Anaconda3\lib\site-packages\librosa\core\spectrum.py in stft(y,dtype,pad_mode)
    226             )
    227 
--> 228         y = np.pad(y,int(n_fft // 2),mode=pad_mode)
    229 
    230     elif n_fft > y.shape[-1]:

<__array_function__ internals> in pad(*args,**kwargs)

C:\Users\Salma\Anaconda3\lib\site-packages\numpy\lib\arraypad.py in pad(array,pad_width,mode,**kwargs)
    817                 raise ValueError(
    818                     "can't extend empty axis {} using modes other than "
--> 819                     "'constant' or 'empty'".format(axis)
    820                 )
    821         # passed,don't need to do anything more as _pad_simple already

ValueError: can't extend empty axis 0 using modes other than 'constant' or 'empty'

所以这里可能是哪里出了问题，提前致谢

解决方法

我遇到了同样的问题，.m4a 音频持续时间为 280 秒，但只有 160KB。 我觉得这个音频很可能坏了，你最好检查一下音频。

,

我认为您可以轻松实现 MFCC 并且 here 具有完整细节的完美实现