python 提取MFCC特征、一阶差分和二阶差分

依赖库：
python-speech-features-0.6

import python_speech_features
import numpy as np
import scipy.io.wavfile
from matplotlib import pyplot as plt
import os

'''
signal - 需要用来计算特征的音频信号，应该是一个N*1的数组
samplerate - 我们用来工作的信号的采样率
winlen - 分析窗口的长度，按秒计，默认0.025s(25ms)
winstep - 连续窗口之间的步长，按秒计，默认0.01s（10ms）
numcep - 倒频谱返回的数量，默认13
nfilt - 滤波器组的滤波器数量，默认26
nfft - FFT的大小，默认512
lowfreq - 梅尔滤波器的最低边缘，单位赫兹，默认为0
highfreq - 梅尔滤波器的最高边缘，单位赫兹，默认为采样率/2
preemph - 应用预加重过滤器和预加重过滤器的系数，0表示没有过滤器，默认0.97
ceplifter - 将升降器应用于最终的倒谱系数。 0没有升降机。默认值为22。
appendEnergy - 如果是true，则将第0个倒谱系数替换为总帧能量的对数。 
'''

def get_mfcc(data, fs, winSize, winStep):
    # MFCC
    wav_feature = python_speech_features.mfcc(data, fs,
                  numcep=13, winlen=winSize, winstep=winStep,  # winstep = 0.01
                  nfilt=26, nfft=512, lowfreq=0, highfreq=None, preemph=0.97)
    # 一阶差分 二阶差分 MFCC 
    d_mfcc_feat = python_speech_features.delta(wav_feature, 1)
    d_mfcc_feat2 = python_speech_features.delta(wav_feature, 2)
    feature = np.hstack((wav_feature, d_mfcc_feat, d_mfcc_feat2))
	# 三者叠加 尺寸为 【len：39】
    return feature


def readWavToMFCCnpy(path, destination, win=0.025, step=0.025):
    idx = 0
    for wav in os.listdir(path):
        idx  = 1
        if idx % 100 == 1:
            print("======= ",idx,"MFCC npy done =======")
        sample_rate, signal = scipy.io.wavfile.read(path  "/"  wav)
        mfcck = get_mfcc(signal, sample_rate, win, step)
        np.save(destination   "/"   "win{}step{}-{}.npy"
                .format(win*1000, step*1000, wav.split(".wav")[0]), mfcck)

    print("----------------------finish get_mfcc----------------------")


if __name__ == '__main__':
	# 源wav路径 和 npy存放路径 和 窗长 和 移动长
    source_path = "D:/ASVsproof/2017v2/ASVspoof2017_V2_train"
    des_path = "D:/ASVsproof/2017v2npy/train"
    readWavToMFCCnpy(source_path, des_path, win=0.025, step=0.01)

这篇好文章是转载于：学新通技术网