cepstralCoefficients

Extract cepstral coefficients

Since R2020b

Syntax

coeffs = cepstralCoefficients(S)

coeffs = cepstralCoefficients(S,Name=Value)

Description

coeffs = cepstralCoefficients(S) returns the cepstral coefficients over time. The input, S, must be a real-valued spectrogram or auditory spectrogram.

coeffs = cepstralCoefficients(S,Name=Value) specifies options using one or more name-value arguments.

For example, coeffs = cepstralCoefficients(S,Rectification="cubic-root") uses cubic-root rectification to calculate the coefficients.

Examples

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;

Call 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;

Call 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)';

Call cepstralCoefficients with the custom auditory spectrogram to create custom cepstral coefficients.

ccc = cepstralCoefficients(S);

Extract Cepstral Coefficients from Streaming Audio

Create a dsp.AudioFileReader object to read in audio frame-by-frame. Create a dsp.AsyncBuffer object to buffer the input into overlapped frames.

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:

1. Read a frame of data from the audio file.

2. Write the frame of data to the buffer.

3. 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.

4. Transform the data to a magnitude spectrum.

5. Apply the mel filter bank to create a mel spectrum.

6. Call cepstralCoefficients to 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

Input Arguments

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 array, where:

• L –– Number of frequency bands

• M –– Number of frames

• N –– Number of channels

Data Types: single | double

Name-Value Arguments

Specify optional pairs of arguments as Name1=Value1,...,NameN=ValueN, where Name is 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.

Example: cepstralCoefficients(S,NumCoeffs=16)

NumCoeffs — Number of cepstral coefficients returned

13 (default) | positive integer greater than 1

Number of coefficients returned for each window of data, specified as a positive integer greater than 1.

Data Types: single | double | int8 | int16 | int32 | int64 | uint8 | uint16 | uint32 | uint64

Rectification — Type of nonlinear rectification

"log" (default) | "cubic-root" | "none"

Type of nonlinear rectification applied prior to the discrete cosine transform, specified as "log", "cubic-root", or "none".

Data Types: char | string

Output Arguments

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 NumCoeffs.

• N –– Number of channels (pages) of the input.

Data Types: single | double

Algorithms

Given an auditory spectrogram, the algorithm to extract N cepstral coefficients from each individual spectrum comprises the following steps.

1. Rectify the spectrum by applying a logarithm, cubic root, or optionally perform no rectification.

2. Apply the discrete cosine transform (DCT-II) to the rectified spectrum.

3. Return the first N coefficients from the cepstral representation.

For more information, see [1].

References

[1] Rabiner, Lawrence R., and Ronald W. Schafer. Theory and Applications of Digital Speech Processing. Upper Saddle River, NJ: Pearson, 2010.

Extended Capabilities

C/C++ Code Generation
Generate C and C++ code using MATLAB® Coder™.

GPU Arrays
Accelerate code by running on a graphics processing unit (GPU) using Parallel Computing Toolbox™.

Version History

Introduced in R2020b