Documentation for matlab_speech_features
This is documentation for matlab_speech_features, a library for speech feature extraction. Code is available at https://github.com/jameslyons/matlab_speech_features. If you find any errors, feel free to make a pull request or leave a comment at the bottom of the page.
Download matlab_speech_features.zip
msf_mfcc - Mel Frequency Cepstral Coefficients §
function feat = msf_mfcc(speech,fs,varargin)
given a speech signal, splits it into frames and computes Mel frequency cepstral coefficients for each frame. For a tutorial on MFCCs, see MFCC tutorial.
- speech - the input speech signal, vector of speech samples
- fs - the sample rate of 'speech', integer
optional arguments supported include the following 'name', value pairs from the 3rd argument on:
- 'winlen' - length of window in seconds. Default: 0.025 (25 milliseconds)
- 'winstep' - step between successive windows in seconds. Default: 0.01 (10 milliseconds)
- 'nfilt' - the number filterbanks to use. Default: 26
- 'lowfreq' - the lowest filterbank edge. In Hz. Default: 0
- 'highfreq' - the highest filterbank edge. In Hz. Default: fs/2
- 'nfft' - the FFT size to use. Default: 512
- 'ncep' - the number of cepstral coeffients to use. Default: 13
- 'liftercoeff' - liftering coefficient, 0 is no lifter. Default: 22
- 'appendenergy' - if true, replaces 0th cep coeff with log of total frame energy. Default: true
Example usage:
mfccs = msf_mfcc(signal,16000,'nfilt',40,'ncep',12);
msf_lpc - Linear Prediction Coefficients §
function feat = msf_lpc(speech,fs,varargin)
given a speech signal, splits it into frames and computes Linear Prediction Coefficients for each frame.
- speech - the input speech signal, vector of speech samples
- fs - the sample rate of 'speech', integer
optional arguments supported include the following 'name', value pairs from the 3rd argument on:
- 'winlen' - length of window in seconds. Default: 0.025 (25 milliseconds)
- 'winstep' - step between successive windows in seconds. Default: 0.01 (10 milliseconds)
- 'order' - the number of coefficients to return. Default: 12
Example usage:
lpcs = msf_lpc(signal,16000,'order',10);
msf_rc - Reflection Coefficients §
function feat = msf_rc(speech,fs,varargin)
given a speech signal, splits it into frames and computes Reflection Coefficients for each frame.
- speech - the input speech signal, vector of speech samples
- fs - the sample rate of 'speech', integer
optional arguments supported include the following 'name', value pairs from the 3rd argument on:
- 'winlen' - length of window in seconds. Default: 0.025 (25 milliseconds)
- 'winstep' - step between successive windows in seconds. Default: 0.01 (10 milliseconds)
- 'order' - the number of coefficients to return. Default: 12
Example usage:
rcs = msf_rc(signal,16000,'order',10);
msf_logfb - Log Filterbank Energies §
function feat = msf_logfb(speech,fs,varargin)
given a speech signal, splits it into frames and computes log filterbank energies for each frame.
- speech - the input speech signal, vector of speech samples
- fs - the sample rate of 'speech', integer
optional arguments supported include the following 'name', value pairs from the 3rd argument on:
- 'winlen' - length of window in seconds. Default: 0.025 (25 milliseconds)
- 'winstep' - step between successive windows in seconds. Default: 0.01 (10 milliseconds)
- 'nfilt' - the number filterbanks to use. Default: 26
- 'lowfreq' - the lowest filterbank edge. In Hz. Default: 0
- 'highfreq' - the highest filterbank edge. In Hz. Default: fs/2
- 'nfft' - the FFT size to use. Default: 512
Example usage:
logfbs = msf_logfb(signal,16000,'nfilt',40,'ncep',12);
msf_filterbank - return a mel-spaced filterbank §
function fbank = msf_filterbank(nfilt,fs,lowfreq,highfreq,nfft)
returns a mel-spaced filterbank for use with filterbank energies, mfccs, sscs etc.
- nfilt - the number filterbanks to use.
- fs - the sample rate of 'speech', integer
- lowfreq - the lowest filterbank edge. In Hz.
- highfreq - the highest filterbank edge. In Hz.
- nfft - the FFT size to use.
Example usage:
lpcs = msf_filterbank(26,16000,0,16000,512);
msf_lsf - Line Spectral Frequencies §
function feat = msf_lsf(speech,fs,varargin)
given a speech signal, splits it into frames and computes Line Spectral Frequencies for each frame.
- speech - the input speech signal, vector of speech samples
- fs - the sample rate of 'speech', integer
optional arguments supported include the following 'name', value pairs from the 3rd argument on:
- 'winlen' - length of window in seconds. Default: 0.025 (25 milliseconds)
- 'winstep' - step between successive windows in seconds. Default: 0.01 (10 milliseconds)
- 'order' - the number of coefficients to return. Default: 12
Example usage:
lsfs = msf_lsf(signal,16000,'order',10);
msf_lpcc - Log Area Ratios §
function feat = msf_lar(speech,fs,varargin)
given a speech signal, splits it into frames and computes Log Area Ratios for each frame.
- speech - the input speech signal, vector of speech samples
- fs - the sample rate of 'speech', integer
optional arguments supported include the following 'name', value pairs from the 3rd argument on:
- 'winlen' - length of window in seconds. Default: 0.025 (25 milliseconds)
- 'winstep' - step between successive windows in seconds. Default: 0.01 (10 milliseconds)
- 'order' - the number of coefficients to return. Default: 12
Example usage:
lars = msf_lar(signal,16000,'order',10);
msf_framesig - break a signal into frames §
function win_frames = msf_framesig(signal, frame_len, frame_step, winfunc)
Takes a 1 by N signal, and breaks it up into frames. Each frame starts frame_step samples after the start of the previous frame. Each frame is windowed by wintype.
- to specify window, use e.g. @hamming, @(x)chebwin(x,30), @(x)ones(x,1), etc.
- signal - the input signal, vector of audio samples
- frame_len - length of window in samples.
- frame_step - step between successive windows in seconds. In samples.
- winfunc - A function to be applied to each window.
Example usage with hamming window:
frames = msf_framesig(speech, winlen*fs, winstep*fs, @(x)hamming(x));
msf_lpcc - Linear Prediction Cepstral Coefficients §
function feat = msf_lpcc(speech,fs,varargin)
given a speech signal, splits it into frames and computes Linear Prediction Cepstral Coefficients for each frame.
- speech - the input speech signal, vector of speech samples
- fs - the sample rate of 'speech', integer
optional arguments supported include the following 'name', value pairs from the 3rd argument on:
- 'winlen' - length of window in seconds. Default: 0.025 (25 milliseconds)
- 'winstep' - step between successive windows in seconds. Default: 0.01 (10 milliseconds)
- 'order' - the number of coefficients to return. Default: 12
Example usage:
lpccs = msf_lpcc(signal,16000,'order',10);
msf_ssc - Spectral Subband Centroids §
function feat = msf_ssc(speech,fs,varargin)
given a speech signal, splits it into frames and computes Spectral Subband Centroids for each frame.
- speech - the input speech signal, vector of speech samples
- fs - the sample rate of 'speech', integer
optional arguments supported include the following 'name', value pairs from the 3rd argument on:
- 'winlen' - length of window in seconds. Default: 0.025 (25 milliseconds)
- 'winstep' - step between successive windows in seconds. Default: 0.01 (10 milliseconds)
- 'nfilt' - the number filterbanks to use. Default: 26
- 'lowfreq' - the lowest filterbank edge. In Hz. Default: 0
- 'highfreq' - the highest filterbank edge. In Hz. Default: fs/2
- 'nfft' - the FFT size to use. Default: 512
Example usage:
sscs = msf_ssc(signal,16000,'nfilt',40,'ncep',12);
msf_powspec - Compute power spectrum of audio frames §
function pspec = msf_powspec(speech,fs,varargin)
given a speech signal, splits it into frames and computes the power spectrum for each frame.
- speech - the input speech signal, vector of speech samples
- fs - the sample rate of 'speech', integer
optional arguments supported include the following 'name', value pairs from the 3rd argument on:
- 'winlen' - length of window in seconds. Default: 0.025 (25 milliseconds)
- 'winstep' - step between successive windows in seconds. Default: 0.01 (10 milliseconds)
- 'nfft' - the FFT size to use. Default: 512
Example usage:
lpcs = msf_powspec(signal,16000,'winlen',0.5);
Contents
- msf_mfcc - Mel Frequency Cepstral Coefficients
- msf_lpc - Linear Prediction Coefficients
- msf_rc - Reflection Coefficients
- msf_logfb - Log Filterbank Energies
- msf_filterbank - return a mel-spaced filterbank
- msf_lsf - Line Spectral Frequencies
- msf_lpcc - Log Area Ratios
- msf_framesig - break a signal into frames
- msf_lpcc - Linear Prediction Cepstral Coefficients
- msf_ssc - Spectral Subband Centroids
- msf_powspec - Compute power spectrum of audio frames