audioDelta

Compute delta features

Syntax

delta = audioDelta(x)

delta = audioDelta(x,deltaWindowLength)

delta = audioDelta(x,deltaWindowLength,initialCondition)

[delta,finalCondition] = audioDelta(x,___)

Description

delta = audioDelta(x) returns the delta of audio features x.

delta = audioDelta(x,deltaWindowLength) specifies the delta window length.

delta = audioDelta(x,deltaWindowLength,initialCondition) specifies the initial condition of the filter.

[delta,finalCondition] = audioDelta(x,___) also returns the final condition of the filter.

Examples

Delta of Audio Features

Read in an audio file.

[audioIn,fs] = audioread('Counting-16-44p1-mono-15secs.wav');

Create an audioFeatureExtractor object to extract some spectral features over time from the audio. Call extract to extract the audio features.

afe = audioFeatureExtractor('SampleRate',fs, ...
    'spectralCentroid',true, ...
    'spectralSlope',true);

audioFeatures = extract(afe,audioIn);

Call audioDelta to approximate the first derivative of the spectral features over time.

deltaAudioFeatures = audioDelta(audioFeatures);

Plot the spectral features and the delta of the spectral features.

map = info(afe);
tiledlayout(2,1)
nexttile
plot(audioFeatures(:,map.spectralCentroid))
ylabel('Spectral Centroid')
nexttile
plot(deltaAudioFeatures(:,map.spectralCentroid))
ylabel('Delta Spectral Centroid')
xlabel('Frame')

tiledlayout(2,1)
nexttile
plot(audioFeatures(:,map.spectralSlope))
ylabel('Spectral Slope')
nexttile
plot(deltaAudioFeatures(:,map.spectralSlope))
ylabel('Delta Spectral Slope')
xlabel('Frame')

Delta and Delta-Delta of MFCC

The delta and delta-delta of mel frequency cepstral coefficients (MFCC) are often used with the MFCC for machine learning and deep learning applications.

Read in an audio file.

[audioIn,fs] = audioread("Counting-16-44p1-mono-15secs.wav");

Use the designAuditoryFilterBank function to design a one-sided frequency-domain mel filter bank.

analysisWindowLength = round(fs*0.03);
fb = designAuditoryFilterBank(fs,"FFTLength",analysisWindowLength);

Use the stft function to convert the audio signal to a complex, one-sided frequency-domain representation. Convert the STFT to magnitude and apply the frequency-domain filtering.

[S,~,t] = stft(audioIn,fs,"Window",hann(analysisWindowLength,"periodic"),"FrequencyRange","onesided");
auditorySTFT = fb*abs(S);

Call the cepstralCoefficients function to extract the MFCC.

melcc = cepstralCoefficients(auditorySTFT);

Call the audioDelta function to compute the delta MFCC. Call audioDelta again to compute the delta-delta MFCC. Plot the results.

deltaWindowLength = 21;

melccDelta = audioDelta(melcc,deltaWindowLength);
melccDeltaDelta = audioDelta(melccDelta,deltaWindowLength);

coefficientToDisplay = 4;

tiledlayout(3,1)
nexttile
plot(t,melcc(:,coefficientToDisplay+1))
ylabel('Coefficient ' + string(coefficientToDisplay))
nexttile
plot(t,melccDelta(:,coefficientToDisplay+1))
ylabel('Delta')
nexttile
plot(t,melccDeltaDelta(:,coefficientToDisplay+1))
xlabel('Time (s)')
ylabel('Delta-Delta')

Delta of Streaming Signals

You can calculate the delta of streaming signals by passing state in and out of the audioDelta function.

Create a dsp.AudioFileReader object to read an audio file frame-by-frame. Create an audioDeviceWriter object to write audio to your speaker. Create a timescope object to visualize the change in harmonic ratio over time.

fileReader = dsp.AudioFileReader("FemaleSpeech-16-8-mono-3secs.wav","SamplesPerFrame",32,"PlayCount",3);
deviceWriter = audioDeviceWriter("SampleRate",fileReader.SampleRate);
scope = timescope("SampleRate",fileReader.SampleRate/fileReader.SamplesPerFrame, ...
    "TimeSpanSource","Property", ...
    "TimeSpan",3, ...
    "YLimits",[-1,1], ...
    "Title","Delta of Harmonic Ratio");

While the audio file has unread frames of data:

1. Read a frame from the audio file

2. Calculate the harmonic ratio of that frame

3. Calculate the delta of the harmonic ratio

4. Write the audio frame to your speaker

5. Write the change in the harmonic ratio to your scope

On each call to audioDelta, overwrite the previous state. Initialize the state using an empty array.

z = [];
while ~isDone(fileReader)
    audioIn = fileReader();
    
    hr = harmonicRatio(audioIn,fileReader.SampleRate,"Window",hann(fileReader.SamplesPerFrame,'periodic'),"OverlapLength",0);
    
    [deltaHR, z] = audioDelta(hr,5,z);
    
    deviceWriter(audioIn);
    
    scope(deltaHR)
end
release(scope)

Input Arguments

x — Audio features

vector | matrix | array

Audio features, specified as a vector, matrix, or multi-dimensional array. The delta computation operates along the first dimension. All other dimensions are treated as independent channels.

Data Types: single | double

deltaWindowLength — Window length over which to calculate delta

9 (default) | odd integer greater than 2

Window length over which to calculate delta, specified as an odd integer greater than 2.

Data Types: single | double

initialCondition — Initial condition of filter

[] (default) | vector | matrix | array

Initial condition of the filter used to calculate the delta, specified as a vector, matrix, or multi-dimensional array. The first dimension of initialCondition must equal deltaWindowLength-1. The remaining dimensions of initialCondition must match the remaining dimensions of the input x. The default initial condition, [], is equivalent to initializing the state with all zeros.

Data Types: single | double

Output Arguments

delta — Delta of audio features

vector | matrix | array

Delta of audio features, returned as a vector, matrix, or multi-dimensional array with the same dimensions as the input x.

Data Types: single | double

finalCondition — Final condition of filter

vector | matrix | array

Final condition of filter, returned as a vector, matrix, or multi-dimensional array. The final condition is returned as the same size as the initialCondition.

Data Types: single | double

More About

Delta of Audio Feature

The delta of an audio feature x is a least-squares approximation of the local slope of a region centered on sample x(k), which includes M samples before the current sample and M samples after the current sample.

$$delta=\frac{{\displaystyle \sum _{k=-M}^{M}kx(k)}}{{\displaystyle \sum _{k=-M}^M{k}^2}}$$

The delta window length defines the length of the region from –M to M.

For more information, see [1]. To learn how audioDelta differs from the definition in [1] and how to equate the function with the definition, see Algorithms.

Algorithms

To support causal streaming, audioDelta introduces a reporting delay of the least-squares approximation equal to floor(deltaWindowLength/2).

For example, suppose x is a column vector of length N, and the delta window length is W. Pad the front and back of the signal with floor(W/2) copies of the first and last elements respectively.

x = randn(100,1);
W = 11;
M = floor(W/2);
frontPad = repmat(x(1),M,1);
backPad = repmat(x(end),M,1);
xPadded = [frontPad ; x ; backPad];

To compute the delta features using the book definition (see Delta of Audio Feature) you can use a for-loop:

delta_book = nan(size(x));
k = -M:M;
for n = 1:length(x)
    delta_book(n) = sum(k(:).*xPadded(n:(n+2*M)))/sum(k.^2);
end

To align audioDelta with the book definition, make these adjustments:

delta_audioDeltaAdjusted = audioDelta(xPadded(:),W);
delta_audioDeltaAdjusted(1:W-1) = [];
scaleFix = 0.5; % This is independent of the window length.
delta_audioDeltaAdjusted = delta_audioDeltaAdjusted*scaleFix;

You can compute the delta features of the original signal and compare root mean square errors.

delta_audioDelta = audioDelta(x,W);
audioDelta_RMS = rms(delta_book-delta_audioDelta)
audioDeltaAdjusted_RMS = rms(delta_book-delta_audioDeltaAdjusted)

audioDelta_RMS =

    0.2681

audioDeltaAdjusted_RMS =

   2.0989e-17

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