语音处理-MFCC

原理

代码过程

import numpy as np
from scipy.io import wavfile
from scipy.signal import lfilter
import matplotlib.pyplot as plt

def enframe(x, win, inc=None):
    """
    对语音信号进行分帧
    args：
    x:输入信号
    win:窗口大小
    return:返回的是分帧后的数据
    """
    nx = len(x)
    nlen = win
    if isinstance(win, list) or isinstance(win, np.ndarray):
        nwin = len(win)
        nlen = nwin  # 帧长=窗长
    elif isinstance(win, int):
        nwin = 1
        nlen = win  # 设置为帧长
    if inc is None:
        inc = nlen
    nf = (nx - nlen + inc) // inc#取整除商的部分向下取整
    frameout = np.zeros((nf, nlen))
    indf = np.multiply(inc, np.array([i for i in range(nf)]))
    for i in range(nf):
        frameout[i, :] = x[indf[i]:indf[i] + nlen]
    if isinstance(win, list) or isinstance(win, np.ndarray):
        frameout = np.multiply(frameout, np.array(win))
    return frameout

def melbankm(p, NFFT, fs, fl, fh):
    """
    计算Mel滤波器组
    args:
     p: 滤波器个数
     NFFT: 一帧FFT后的数据长度
     fs: 采样率
     fl: 最低频率
     fh: 最高频率
    return:
    """
    bl = 1125 * np.log(1 + fl / 700)  # 把 Hz 变成 Mel
    bh = 1125 * np.log(1 + fh / 700)
    B = bh - bl  # Mel带宽
    y = np.linspace(0, B, p + 2)  # 将梅尔刻度等间隔
    Fb = 700 * (np.exp(y / 1125) - 1)  # 把 Mel 变成Hz
    W2 = int(NFFT / 2 + 1)
    df = fs / NFFT
    freq = [int(i * df) for i in range(W2)]  # 采样频率值
    bank = np.zeros((p, W2))
    for k in range(1, p + 1):
        f0, f1, f2 = Fb[k], Fb[k - 1], Fb[k + 1]
        n1 = np.floor(f1 / df)
        n2 = np.floor(f2 / df)
        n0 = np.floor(f0 / df)
        for i in range(1, W2):
            if n1 <= i <= n0:
                bank[k - 1, i] = (i - n1) / (n0 - n1)
            elif n0 < i <= n2:
                bank[k - 1, i] = (n2 - i) / (n2 - n0)
            elif i > n2:
                break
        # plt.plot(freq, bank[k - 1, :], 'r')
    # plt.savefig('images/mel.png')
    return bank

def mfcc(data,fs,p,frameSize,inc,nfft=512,n_dct = 12):
    #预处理-加重
    data = lfilter([1,-0.94],[1],data)
    #分帧
    data = enframe(data,frameSize,inc)
    #加窗
    data = np.multiply(data,np.hanning(frameSize))
    #fft
    data = np.fft.rfft(data,nfft)
    #计算能量普
    data = (np.abs(data)**2)/nfft
    #计算通过Mel滤波器的能量
    bank = melbankm(p,nfft,fs,0,0.5*fs)
    ss = np.matmul(data, bank.T)
    #计算DCT倒谱
    M = bank.shape[0]
    m = np.array([i for i in range(M)])
    mfcc = np.zeros((ss.shape[0],n_dct))
    for n in range(n_dct):
        mfcc[:, n] = np.sqrt(2 / M) * np.sum(np.multiply(np.log(ss), np.cos((2 * m - 1) * n * np.pi / 2 / M)), axis=1)
    return mfcc

#加载语音文件
fs,data = wavfile.read('fft.wav')
temp = mfcc(data,fs,16, int(fs // 2 * 0.025), int(fs // 2 * 0.01))
time = np.array([i  for i in range(len(data))])
another = temp
temp = temp[:,1]
plt.subplot(2,1,1)
plt.plot(temp)
plt.subplot(2,1,2)
plt.plot(another)
plt.savefig('MFCC.png')
plt.show()

结果图