语音识别第4讲：语音特征参数MFCC https://zhuanlan.zhihu.com/p/88625876/
Speech Processing for Machine Learning: Filter banks, Mel-Frequency Cepstral Coefficients (MFCCs) and What's In-Between https://haythamfayek.com/2016/04/21/speech-processing-for-machine-learning.html
python
#梅尔频率倒谱系数(MFCC)的原理讲解及python实现
# https://www.cnblogs.com/LXP-Never/p/10918590.html
import numpy
import numpy as np
import scipy.io.wavfile
from scipy.fftpack import dct
sample_rate, signal = scipy.io.wavfile.read(r'H:\p227_003.wav')
signal = signal[0:int(3.5 * sample_rate)] # 我们只取前3.5s
pre_emphasis = 0.97
#1、预加重 (Pre-Emphasis)
emphasized_signal = numpy.append(signal[0], signal[1:] - pre_emphasis * signal[:-1])
frame_size = 0.025
frame_stride = 0.01
overlap=0.015
#2、分帧 (Framing)
frame_length, frame_step = frame_size * sample_rate, frame_stride * sample_rate # 从秒转换为采样点
signal_length = len(emphasized_signal)
frame_length = int(round(frame_length))
frame_step = int(round(frame_step))
# 确保我们至少有1帧
num_frames = int(numpy.ceil(float(numpy.abs(signal_length - frame_length)) / frame_step)) + 1 #原文这里错了
pad_signal_length = (num_frames-1) * frame_step + frame_length #原文这里错了
z = numpy.zeros((pad_signal_length - signal_length))
# 填充信号，确保所有帧的采样数相等，而不从原始信号中截断任何采样
pad_signal = numpy.append(emphasized_signal, z)
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
frames = pad_signal[indices.astype(numpy.int32, copy=False)]
#3、加窗 (Window)
frames *= numpy.hamming(frame_length)
# frames *= 0.54 - 0.46 * numpy.cos(( numpy.pi * n ) / (frame_length - 1)) # 内部实现
NFFT=512
# 二、FFT (Fourier-Transform）
mag_frames = numpy.absolute(numpy.fft.rfft(frames, NFFT)) # fft的幅度(magnitude)
# 三、功率谱 (Power Spectrum)
pow_frames = ((1.0 / NFFT) * ((mag_frames) ** 2)) # 功率谱
# 四、滤波器组 (Filter Banks）
nfilt = 40
low_freq_mel = 0
high_freq_mel = (2595 * np.log10(1 + (sample_rate / 2) / 700)) # 求最高hz频率对应的mel频率
# 我们要做40个滤波器组，为此需要42个点，这意味着在们需要low_freq_mel和high_freq_mel之间线性间隔40个点
mel_points = np.linspace(low_freq_mel, high_freq_mel, nfilt + 2) # 在mel频率上均分成42个点
hz_points = (700 * (10 ** (mel_points / 2595) - 1)) # 将mel频率再转到hz频率
# bin = sample_rate/2 / NFFT/2=sample_rate/NFFT # 每个频点的频率数
# bins = hz_points/bin=hz_points*NFFT/ sample_rate # hz_points对应第几个fft频点
bins = np.floor((NFFT + 1) * hz_points / sample_rate)
fbank = np.zeros((nfilt, int(np.floor(NFFT / 2 + 1))))
for m in range(1, nfilt + 1):
f_m_minus = int(bins[m - 1]) # 左
f_m = int(bins[m]) # 中
f_m_plus = int(bins[m + 1]) # 右
for k in range(f_m_minus, f_m):
fbank[m - 1, k] = (k - bins[m - 1]) / (bins[m] - bins[m - 1])
for k in range(f_m, f_m_plus):
fbank[m - 1, k] = (bins[m + 1] - k) / (bins[m + 1] - bins[m])
filter_banks = np.dot(pow_frames, fbank.T)
filter_banks = np.where(filter_banks == 0, np.finfo(float).eps, filter_banks) # 数值稳定性
filter_banks = 20 * np.log10(filter_banks) # dB
#五、梅尔频率倒谱系数（MFCCs）
num_ceps = 12
mfcc = dct(filter_banks, type=2, axis=1, norm='ortho')[:, 1 : (num_ceps + 1)] # 保持在2-13
cep_lifter =22
(nframes, ncoeff) = mfcc.shape
n = numpy.arange(ncoeff)
lift = 1 + (cep_lifter / 2) * numpy.sin(numpy.pi * n / cep_lifter)
mfcc *= lift
# 六、均值归一化（Mean Normalization）
filter_banks -= (numpy.mean(filter_banks, axis=0) + 1e-8)
mfcc -= (numpy.mean(mfcc, axis=0) + 1e-8)