Extract cepstral coefficients
Mel Frequency Cepstral Coefficients
Read an audio file into the workspace.
[audioIn,fs] = audioread('SpeechDFT-16-8-mono-5secs.wav');
Convert the audio signal to a frequency-domain representation using 30 ms windows with 15 ms overlap. Because the input is real and therefore the spectrum is symmetric, you can use just one side of the frequency domain representation without any loss of information. Convert the complex spectrum to the magnitude spectrum: phase information is discarded when calculating mel frequency cepstral coefficients (MFCC).
windowLength = round(0.03*fs); overlapLength = round(0.015*fs); S = stft(audioIn,"Window",hann(windowLength,"periodic"),"OverlapLength",overlapLength,"FrequencyRange","onesided"); S = abs(S);
Design a one-sided frequency-domain mel filter bank. Apply the filter bank to the frequency-domain representation to create a mel spectrogram.
filterBank = designAuditoryFilterBank(fs,'FFTLength',windowLength); melSpec = filterBank*S;
cepstralCofficients with the mel spectrogram to create MFCC.
melcc = cepstralCoefficients(melSpec);
Gammatone Frequency Cepstral Coefficients
Read an audio signal and convert it to a one-sided magnitude short-time Fourier transform. Use a 50 ms periodic Hamming window with a 10 ms hop.
[audioIn,fs] = audioread('NoisySpeech-16-22p5-mono-5secs.wav'); windowLength = round(0.05*fs); hopLength = round(0.01*fs); overlapLength = windowLength - hopLength; S = stft(audioIn,"Window",hamming(windowLength,'periodic'),"OverlapLength",overlapLength,"FrequencyRange","onesided"); S = abs(S);
Design a one-sided frequency-domain gammatone filter bank. Apply the filter bank to the frequency-domain representation to create a gammatone spectrogram.
filterBank = designAuditoryFilterBank(fs,'FFTLength',windowLength,"FrequencyScale","erb"); gammaSpec = filterBank*S;
cepstralCoefficients with the gammatone spectrogram to create gammatone frequency cepstral coefficients. Use a cubic-root rectification.
gammacc = cepstralCoefficients(gammaSpec,"Rectification","cubic-root");
Custom Cepstral Coefficients
Cepstral coefficients are commonly used as compact representations of audio signals. Generally, they are calculated after an audio signal is passed through a filter bank and the energy in the individual filters is summed. Researchers have proposed various filter banks based on psychoacoustic experiments (such as mel, Bark, and ERB). Using the
cepstralCoefficients function, you can define your own custom filter bank and then analyze the resulting cepstral coefficients.
Read in an audio file for analysis.
[audioIn,fs] = audioread('Counting-16-44p1-mono-15secs.wav');
Design a filter bank that consists of 20 triangular filters with band edges over the range 62.5 Hz to 8000 Hz. Spread the filters evenly in the log domain. For simplicity, design the filters in bins. Most popular auditory filter banks are designed in a continuous domain, such as Hz, mel, or Bark, and then warped back to bins.
numFilters = 20; filterbankStart = 62.5; filterbankEnd = 8000; numBandEdges = numFilters + 2; NFFT = 1024; filterBank = zeros(numFilters,NFFT); bandEdges = logspace(log10(filterbankStart),log10(filterbankEnd),numBandEdges); bandEdgesBins = round((bandEdges/fs)*NFFT) + 1; for ii = 1:numFilters filt = triang(bandEdgesBins(ii+2)-bandEdgesBins(ii)); leftPad = bandEdgesBins(ii); rightPad = NFFT - numel(filt) - leftPad; filterBank(ii,:) = [zeros(1,leftPad),filt',zeros(1,rightPad)]; end
Plot the filter bank.
frequencyVector = (fs/NFFT)*(0:NFFT-1); plot(frequencyVector,filterBank'); xlabel('Hz') axis([0 frequencyVector(NFFT/2) 0 1])
Transform the audio signal using the
stft function, and then apply the custom filter bank. Apply the filter bank to the frequency-domain representation to create a custom auditory spectrogram. Plot the spectrogram.
[S,~,t] = stft(audioIn,fs,"Window",hann(NFFT,'periodic'),"FrequencyRange","twosided"); S = abs(S); spec = filterBank*S; surf(t,bandEdges(2:end-1),10*log10(spec),'EdgeColor','none') view([0,90]) axis([t(1) t(end) bandEdges(2) bandEdges(end-1)]) xlabel('Time (s)') ylabel('Frequency (Hz)') c = colorbar; c.Label.String = 'Power (dB)';
cepstralCoefficients with the custom auditory spectrogram to create custom cepstral coefficients.
ccc = cepstralCoefficients(S);
Extract Cepstral Coefficients from Streaming Audio
fileReader = dsp.AudioFileReader("Ambiance-16-44p1-mono-12secs.wav"); buff = dsp.AsyncBuffer;
Design a two-sided mel filter bank that is compatible with 30 ms windows.
windowLength = round(0.03*fileReader.SampleRate); filterBank = designAuditoryFilterBank(fileReader.SampleRate,"FFTLength",windowLength,"OneSided",false);
In an audio stream loop:
Read a frame of data from the audio file.
Write the frame of data to the buffer.
If enough data is available for a hop, read a 30 ms frame of data from the buffer with a 20 ms overlap between frames.
Transform the data to a magnitude spectrum.
Apply the mel filter bank to create a mel spectrum.
cepstralCoefficientsto return the mel frequency cepstral coefficients (MFCC).
win = hann(windowLength,'periodic'); overlapLength = round(0.02*fileReader.SampleRate); hopLength = windowLength - overlapLength; while ~isDone(fileReader) audioIn = fileReader(); write(buff,audioIn); while buff.NumUnreadSamples > hopLength x = read(buff,windowLength,overlapLength); X = abs(fft(x.*win)); melSpectrum = filterBank*X; melcc = cepstralCoefficients(melSpectrum); end end
S — Spectrogram or auditory spectrogram
matrix | 3-D array
Spectrogram or auditory spectrogram, specified as an L-by-M matrix or L-by-M-by-N 3-D array, where:
L –– Number of frequency bands
M –– Number of frames
N –– Number of channels
Specify optional pairs of arguments as
the argument name and
Value is the corresponding value.
Name-value arguments must appear after other arguments, but the order of the
pairs does not matter.
Before R2021a, use commas to separate each name and value, and enclose
Name in quotes.
NumCoeffs — Number of cepstral coefficients returned
13 (default) | positive integer greater than one
Number of coefficients returned for each window of data, specified as the
comma-separated pair consisting of
'NumCoeffs' and a positive
integer greater than one.
Rectification — Type of nonlinear rectification
'log' (default) |
Type of nonlinear rectification applied prior to the discrete cosine transform,
specified as the comma-separated pair consisting of
coeffs — Cepstral coefficients
matrix | 3-D array
Cepstral coefficients, returned as an M-by-B matrix or M-by-B-by-N array, where:
M –– Number of frames (columns) of the input.
B –– Number of coefficients returned per frame. This is determined by
N –– Number of channels (pages) of the input.
Given a time-frequency representation, the
function performs the following operations on each spectrum, individually, as described in
 Rabiner, Lawrence R., and Ronald W. Schafer. Theory and Applications of Digital Speech Processing. Upper Saddle River, NJ: Pearson, 2010.