关于scikits.talkbox.feature.mfcc的学习解释

import numpy as np

from scipy.io import loadmat
#使用模块scipy.io的loadmat可以实现python对mat数据的读写。
from scipy.signal import lfilter, hamming
#lfilter数字滤波器  hamming汉明窗
from scipy.fftpack import fft
#快速傅里叶变换之后生成复数
from scipy.fftpack.realtransforms import dct
#离散余弦变换(DCT)，大多书自然信号（声音、图像）的能量都集中在离散余弦变换后的低频部分。
from scikits.talkbox import segment_axis

from mel import hz2mel
#从赫兹转换为梅尔刻度

def trfbank(fs, nfft, lowfreq, linsc, logsc, nlinfilt, nlogfilt):
    """计算用于MFCC的三角滤波器."""
    # 过滤器总数
    nfilt = nlinfilt + nlogfilt

    #------------------------
    # 计算滤波器组
    #------------------------
    # 计算谱中三角滤波器的起点\中间点\终止点
    # 范围
    freqs = np.zeros(nfilt+2)
    freqs[:nlinfilt] = lowfreq + np.arange(nlinfilt) * linsc
    freqs[nlinfilt:] = freqs[nlinfilt-1] * logsc ** np.arange(1, nlogfilt + 3)
    heights = 2./(freqs[2:] - freqs[0:-2])

    # 计算滤波器组系数 (in fft domain, in bins)
    fbank = np.zeros((nfilt, nfft))
    # FFT bins (in Hz)
    nfreqs = np.arange(nfft) / (1. * nfft) * fs
    for i in range(nfilt):
        low = freqs[i]
        cen = freqs[i+1]
        hi = freqs[i+2]

        lid = np.arange(np.floor(low * nfft / fs) + 1,
                        np.floor(cen * nfft / fs) + 1, dtype=np.int)
        lslope = heights[i] / (cen - low)
        rid = np.arange(np.floor(cen * nfft / fs) + 1,
                        np.floor(hi * nfft / fs) + 1, dtype=np.int)
        rslope = heights[i] / (hi - cen)
        fbank[i][lid] = lslope * (nfreqs[lid] - low)
        fbank[i][rid] = rslope * (hi - nfreqs[rid])

    return fbank, freqs
    #返回 滤波器和频率

def mfcc(input, nwin=256, nfft=512, fs=16000, nceps=13):
    """计算Mel频率倒谱系数.

    参数
    ----------
    输入: ndarray
        计算系数的输入

    返回
    -------
    ceps: ndarray
        Mel倒谱系数
    mspec: ndarray
        mel域中的对数谱。

    Notes
    -----
MFCC计算如下：
              *时域预处理（预强调）
              *用Hamming窗加窗计算频谱幅度
              *在谱域中用三角形滤波信号滤波器组，其滤波器在mel比例，并且在mel比例中具有相等的带宽                   
              *计算对数谱的DCT

    References
    ----------
    .. [1] S.B. Davis and P. Mermelstein, "Comparison of parametric
           representations for monosyllabic word recognition in continuously
           spoken sentences", IEEE Trans. Acoustics. Speech, Signal Proc.
           ASSP-28 (4): 357-366, August 1980."""

    # MFCC参数：取自听觉工具箱
    over = nwin - 160
    # 预加重因子（考虑嘴唇水平辐射的-6dB/倍频程衰减）
    prefac = 0.97

    #lowfreq = 400 / 3.
    lowfreq = 133.33
    #highfreq = 6855.4976
    linsc = 200/3.
    logsc = 1.0711703

    nlinfil = 13
    nlogfil = 27
    nfil = nlinfil + nlogfil

    w = hamming(nwin, sym=0)

    fbank = trfbank(fs, nfft, lowfreq, linsc, logsc, nlinfil, nlogfil)[0]

    #------------------
    # 计算MFCC
    #------------------
    extract = preemp(input, prefac)
    framed = segment_axis(extract, nwin, over) * w

    # 计算频谱幅度
    spec = np.abs(fft(framed, nfft, axis=-1))
    # 通过三角形滤波器组过滤频谱
    mspec = np.log10(np.dot(spec, fbank.T))
    # 使用DCT“压缩”系数（频谱->倒谱域）
    ceps = dct(mspec, type=2, norm='ortho', axis=-1)[:, :nceps]

    return ceps, mspec, spec

def preemp(input, p):
    """预加重滤波器."""
    return lfilter([1., -p], 1, input)

if __name__ == '__main__':
    extract = loadmat('extract.mat')['extract']
    ceps = mfcc(extract)