Fork me on GitHub

mfcc的特征提取python 代码实现和解析 

 1 #!/usr/bin/python
 2 # -*- coding: UTF-8 -*-
 3 
 4 import numpy
 5 import scipy.io.wavfile
 6 from matplotlib import pyplot as plt
 7 from scipy.fftpack import dct
 8 
 9 sample_rate,signal=scipy.io.wavfile.read('stop.wav')
10 
11 print(sample_rate,len(signal))
12 #读取前3.5s 的数据
13 signal=signal[0:int(3.5*sample_rate)]
14 print(signal)
15 
16 
17 
18 #预先处理
19 pre_emphasis = 0.97
20 emphasized_signal = numpy.append(signal[0], signal[1:] - pre_emphasis * signal[:-1])
21 
22 
23 frame_size=0.025
24 frame_stride=0.1
25 frame_length,frame_step=frame_size*sample_rate,frame_stride*sample_rate
26 signal_length=len(emphasized_signal)
27 frame_length=int(round(frame_length))
28 frame_step=int(round(frame_step))
29 num_frames=int(numpy.ceil(float(numpy.abs(signal_length-frame_length))/frame_step))
30 
31 
32 pad_signal_length=num_frames*frame_step+frame_length
33 z=numpy.zeros((pad_signal_length-signal_length))
34 pad_signal=numpy.append(emphasized_signal,z)
35 
36 
37 indices = numpy.tile(numpy.arange(0, frame_length), (num_frames, 1)) + numpy.tile(numpy.arange(0, num_frames * frame_step, frame_step), (frame_length, 1)).T
38 
39 frames = pad_signal[numpy.mat(indices).astype(numpy.int32, copy=False)]
40 
41 #加上汉明窗
42 frames *= numpy.hamming(frame_length)
43 # frames *= 0.54 - 0.46 * numpy.cos((2 * numpy.pi * n) / (frame_length - 1))  # Explicit Implementation **
44 
45 #傅立叶变换和功率谱
46 NFFT = 512
47 mag_frames = numpy.absolute(numpy.fft.rfft(frames, NFFT))  # Magnitude of the FFT
48 #print(mag_frames.shape)
49 pow_frames = ((1.0 / NFFT) * ((mag_frames) ** 2))  # Power Spectrum
50 
51 
52 
53 low_freq_mel = 0
54 #将频率转换为Mel
55 nfilt = 40
56 high_freq_mel = (2595 * numpy.log10(1 + (sample_rate / 2) / 700))
57 mel_points = numpy.linspace(low_freq_mel, high_freq_mel, nfilt + 2)  # Equally spaced in Mel scale
58 hz_points = (700 * (10**(mel_points / 2595) - 1))  # Convert Mel to Hz
59 
60 bin = numpy.floor((NFFT + 1) * hz_points / sample_rate)
61 
62 fbank = numpy.zeros((nfilt, int(numpy.floor(NFFT / 2 + 1))))
63 
64 for m in range(1, nfilt + 1):
65     f_m_minus = int(bin[m - 1])   # left
66     f_m = int(bin[m])             # center
67     f_m_plus = int(bin[m + 1])    # right
68     for k in range(f_m_minus, f_m):
69         fbank[m - 1, k] = (k - bin[m - 1]) / (bin[m] - bin[m - 1])
70     for k in range(f_m, f_m_plus):
71         fbank[m - 1, k] = (bin[m + 1] - k) / (bin[m + 1] - bin[m])
72 filter_banks = numpy.dot(pow_frames, fbank.T)
73 filter_banks = numpy.where(filter_banks == 0, numpy.finfo(float).eps, filter_banks)  # Numerical Stability
74 filter_banks = 20 * numpy.log10(filter_banks)  # dB
75 
76 num_ceps = 12
77 mfcc = dct(filter_banks, type=2, axis=1, norm='ortho')[:, 1 : (num_ceps + 1)]
78 (nframes, ncoeff) = mfcc.shape
79 
80 n = numpy.arange(ncoeff)
81 cep_lifter =22
82 lift = 1 + (cep_lifter / 2) * numpy.sin(numpy.pi * n / cep_lifter)
83 mfcc *= lift  #*
84 
85 #filter_banks -= (numpy.mean(filter_banks, axis=0) + 1e-8)
86 mfcc -= (numpy.mean(mfcc, axis=0) + 1e-8)
87 
88 print(mfcc.shape)
89 plt.plot(filter_banks)
90 
91 plt.show()

 

测试结果：

 

posted on 2018-10-15 14:13  虚生  阅读(4032)  评论(0)  编辑  收藏  举报