如何创建在MFCC中用于MATLAB中语音识别的三角形（Mel）滤镜库？

Question

尽管可能有内置的功能可用，但我需要创建自己的三角滤波器组。 下面是我的代码。 我在HMatrix（filterbank）中得到NaN值。 这是由于创建矩阵时使用的FreqArray中的“相同”值所致。 我在以下问题上需要帮助：

1. 知道我选择的44100Hz采样频率是否正确吗？
2. 如何选择较低的频率= 300Hz和较高的频率= 8000Hz来计算梅尔滤波器组矩阵？
3. 如何选择合适的帧大小（frame_length）和梅尔滤波器的数量（no_of_coeffs）？

function TriFilterBank()
tic
%-----------------------------INITIALISATION---------------------------%


fs=44100; %frequency at which I have sampled my recorded samples
frame_length=256; %How to choose an appropriate frame-size?
low_freq=300; %lower frequency for calculation of mel frequency filter bank (I'm unable to choose a correct one, and find the criteria for choosing it)
high_freq=8000; %upper frequency for calculation of mel frequency filter bank (I'm unable to choose a correct one, and find the criteria for choosing it)
                % I have also tried with (fs/2)=22050Hz, but nno good results 
no_of_coeffs=20;    % This is no. of Mel-Filter banks to create. how to choose a approriate value for this for speech processing applications?


%--------------------------------------------------PRE-PROCESSING FOR MEL FILTER BANK CREATION-----------------------------------------------%

low_linear=2595*log10(1+(low_freq/700));
high_linear=2595*log10(1+(high_freq/700));
band_length=(high_linear-low_linear)/(no_of_coeffs+1); 

MelArray(no_of_coeffs+2,1)=zeros(); %to store mel frequencies to calculate mel frequency filter bank
LinearArray(no_of_coeffs+2,1)=zeros();  %to store linear frequencies to calculate mel frequency filter bank
FreqArray(no_of_coeffs+2,1)=zeros();    %to store frequency array to calculate mel frequency filter bank
%{
THIS ARRAY MAY HAVE WRONG VALUES DUE TO SELECTION of WRONG PARAMETERS LIKE low_freq, high_freq, frame_length (frame-size), no_of_coeffs (no. of filter banks). THIS IS MAJOR REASON BEHIND GENERATION OF NaN values in HMatrix
%} 
HMatrix(no_of_coeffs,frame_length)=zeros();  %Hmk Matrix/ Filter Bank I'M VERY DOUBTFUL OF THE VALUES GENERATED BY THIS FILTER BANK

MelArray(1)=low_linear;
MelArray(no_of_coeffs+2)=high_linear;
LinearArray(1)=low_freq;
LinearArray(no_of_coeffs+2)=high_freq;
FreqArray(1)=floor((int32(frame_length)+1)*LinearArray(1)/fs);
FreqArray(no_of_coeffs+2)=floor((int32(frame_length)+1)*LinearArray(no_of_coeffs+2)/fs);



for m=1:no_of_coeffs
    MelArray(m+1)=MelArray(m)+band_length;
    LinearArray(m+1)=700*((power( 10,MelArray(m+1)/2595))-1);
    FreqArray(m+1)=floor((int32(frame_length)+1)*LinearArray(m+1)/fs); %The values generated here seem to be doubtful, hence maybe an incorrect filter bank
end


% THE MOST DOUBTFUL/WRONG PART i.e. MEL FREQUENCY FILTER BANK MATRIX CREATION

%---------------------------------------------------------PROBABLE ERRONEOUS PART------------------------------------------------------------%
% I'M GETTING NaN values in this matrix probably due to choosing incorrect parameters for  like upper freq, lower freq, frame-size, no.of filter banks, sampling frequency etc.
% In FreqArray I'm getting two same values, hence it's satisfying none of the below conditions and generating a NaN value.

for k=1:frame_length
    for m=1:no_of_coeffs
        if(k<FreqArray(m))
            HMatrix(m,k)=0;
        elseif (FreqArray(m)<=k && k<=FreqArray(m+1))
            HMatrix(m,k)=(k-FreqArray(m))/(FreqArray(m+1)-FreqArray(m));
        elseif(FreqArray(m+1)<=k && k<=FreqArray(m+2))
             HMatrix(m,k)=(FreqArray(m+2)-k)/(FreqArray(m+2)-FreqArray(m+1));
        elseif (k>FreqArray(m+2))
             HMatrix(m,k)=0;
        end
    end
end
%--------------------------------------------------------------------------------------------------------------------------------------------%
save('TriFilterBank');
toc    
end

该代码基于以下等式：

上面的代码输出的主要部分如下所示，以供参考。

作为参考，我使用了以下网站：

http://practicalcryptography.com/miscellaneous/machine-learning/guide-mel-frequency-cepstral-coefficients-mfccs/

提前致谢！

Answer 1

知道我选择的44100Hz采样频率是否正确吗？

这个频率很好。 无论如何，语音驻留在16khz以下，因此16kHz是更常见的选择。 在您用作参考的博客文章中，它是16kHz。

如何选择较低的频率= 300Hz和较高的频率= 8000Hz来计算梅尔滤波器组矩阵？

此范围不是最佳范围，但对于大多数应用程序来说还可以。 为了获得高质量的声音，范围为20Hz至7600Hz。

如何选择合适的帧大小（frame_length）和梅尔滤波器的数量（no_of_coeffs）？

语音的帧大小通常约为25毫秒，这是在一帧内提供平稳性和正常速率语音分辨率的最佳值。 对于44100 kHz的采样率，它以帧中的大约1128（44100 * 0.025）个元素结束，而不是您选择的256个元素。 如果要具有2的幂，则一帧中需要2048个元素。 这也将是FFT阶数。

梅尔过滤器的数量可以是15-40，在许多系统中使用20是一个很好的值，发现它在实验中很有用。

最好阅读现有的实现，有许多您从本教程中学不到的东西，其中一个很好的是VoiceBox 。