Mel滤波器组的设计.docx

《Mel滤波器组的设计.docx》由会员分享，可在线阅读，更多相关《Mel滤波器组的设计.docx（5页珍藏版）》请在淘文阁 - 分享文档赚钱的网站上搜索。

1、一、Mel滤波器组的设计我们设置采样率fs=8000Hzfs=8000Hz ,滤波器频率范围的最低频率fl=Ofl=O ,滤波器频率范 围的最高频率 fh=fs/2=8000/2=4000fh=fs/2=8000/2=4000 ；设置滤波器个数 M=24M=24, FFT 的长度N=256N=256。根据每个带通滤波器的传递函数可以计算出每个Mel滤波器的值，再 由设置的滤波器个数可以形成一个Mel滤波器组。注意每个带通滤波器的传递函数值是在 Mel频率域上计算的，计算完后再转到频率域上。二、MATLAB 实现fs=8000;fl=O; fh=fs/2;bl=1125*log(l+fl/700

2、);%将频率转换为 Mel 频率bh=1125*log(l+fh/700);p=24;%滤波器个数nfft=256;%FFT 点数B=bh-bl;y=linspace(0,B,p+2);%产生 0 到 B 之间 p+2 个数Fb=700*(exp(y/1125)-l);%将 Mel 频率转换为频率W2=nfft/2+l;%fs/2 内对应的 FFT 点数df=fs/nfft;freq=(0:W2-l)*df;% 采样频率值bank二zeros(24,W2);%生成一个24行W2列的全零数组for k=2:p+l%why 从 2 开始？因为 k-1fl=Fb(k-l); f2=Fb(k+l);

3、fO=Fb(k);nl二floor(fl/df)+l;%f(m.l)在频域中的谱线索引号n2=floor(f2/df)+l;%f(m+l)在频域中的谱线索引号nO=floor(fO/df)+l;%f(m)在频域中的谱线索引号。f(m)是从0开始，而在MATLAB中数组 的索引是从1开始，所以要加L否则会出现index=0的错误for i=l: W2if i=nl & inO & i=nl and in0 and i=n2:bankk-lJ=(n2-i)/(n2-nO) # print(k) # print(bankk-l,:) (freq,bankk-l,:/r)set-up代码import

4、numpyimport .wavfilefrom ck import detsample rate, signal = . (0SR_us_000_0010_8k.wav)# File assumed to be in the samedirectorysignal = signal 0: int (3. 5 * sample_rate) # Keep the first 3. 5 seconds预加重emphasized_signal = d(signal0, signal1: - pre_emphasis * signalframing 代码frame_stride = 0. 01.fra

5、me_length, frame_step = frame_size * sample_rate, frame_stride * sample_rate # Convert from seconds to samplessignal_length = len (emphasized _signal)frame_length = int (round (frame_length)frame_step = int (round (frame_step)num_frames = int(float(signal_length - frame_length) / frame_step)# Make s

6、ure thatwe have at least 1 framepad_signal_length = num_frames * frame_step + frame_lengthz = (pad_signal_length - signal_length)pad_signal = d(emphasized_signal, z) # Pad Signal to make sure that all frames have equal number of samples without truncating any samples from the original signalindices

7、= (e(0, frame_length), (num_frames, 1) + (e(0, num_frames * frame_step, frame_step), (frame_length, 1). Tframes = pad_signale(32, copy=False)加窗分帧frames *= ng(frame_length)# frames *= 0. 54 - 0. 46 * (2 * * n) / (frame length - 1) # Explici t Implemen ta tion *mag_frames = ute (. rfft (frames, NFFT)#

8、 Magnitude of the FFTpow_frames = (1. 0 / NFFT) * (mag_frames) * 2) tt Power Spectrurnfilter bankslow freq_mel = 0high freq_mel = (2595 * 10(1 + (samplerate / 2) / 700)# Convert Hz to Melmel_points = ace(low_freq_mel, high_freq_mel, nfilt + 2) U Equally spaced in Mcl scale hz_points = (700 * (10* (m

9、el_points / 2595) - 1) # Convert Mcl to Hz bin = (NFFT + 1) * hz points / sample rate) fbank = (nfilt, int(NFFT / 2 + 1)for m in range(1, nfilt + 1):f_m_minus = int(binm - 1)# leftf_m = int (binmj)# centerfmplus = int(binm + 1)# rightfor k in range(f_m_minus, f_m):fbankm - 1, k = (k - binm - 1) / (b

10、inm - binm - 1)for k in range(f_m, f_m_plus):fbankm - 1, k = (binm + 1 - k) / (binm + 1 - binm)filter banks = (pow frames, fbank. T)filter_banks = (filter_banks = 0, (float). eps, filter_banks) # Numerical Stabilityfilter_banks = 20 * 10(filter_banks) # dBMel-frequency Cepstral Coefficients (MFCCs)m

11、fcc = det (filter banks, type=2, axis=l, norm=，ortho,) :, 1 : (num ceps + 1) tt Keep 2-13(nframes, ncoeff)= n = e (ncoeff)lift = 1 + (cep_lifter / 2) * ( * n / cep_lifter) mfcc *= liftMean NormalizationThe mean-normalized filter banks:filter_banks -= (filterbanks, axis=0) + le-8)The mean-normalized MFCCsmfcc -= (mfcc, axis=0) + le-8)