Why do I obtain a certain dimension in audio features vectors?

I am doing a features extraction from an audio file of 30s, windowed with 5s windows using the following code:

%% Features extraction windowing the signal

duration_audio = N/fs; % Durata totale del segnale in secondi

window_length = 5; % Lunghezza della finestra in secondi

number_windows = duration_audio / window_length;

% Inizializza una matrice per le caratteristiche normalizzate

features_matrix = [];

% Loop per estrarre e normalizzare le caratteristiche da ciascuna finestra temporale

for i = 1:number_windows

% Calcola l'indice di inizio e fine della finestra corrente

ind_start = (i - 1) * window_length * fs + 1;

ind_end = i * window_length * fs;

% Estrai la finestra corrente dal segnale

window_audio = x(ind_start:ind_end);

% Calcola le caratteristiche spettrali

spectral_rolloff = spectralRolloffPoint(window_audio, fs);

spectral_centroid = spectralCentroid(window_audio, fs);

spectral_kurtosis = spectralKurtosis(window_audio, fs);

spectral_entropy = spectralEntropy(window_audio, fs);

spectral_flatness = spectralFlatness(window_audio, fs);

spectral_crest = spectralCrest(window_audio, fs);

spectral_flux = spectralFlux(window_audio, fs);

mfccs = mfcc(window_audio, fs);

% Concatena le caratteristiche in una riga della matrice

features_line = [spectral_rolloff, spectral_centroid, spectral_kurtosis, spectral_entropy, spectral_flatness, spectral_crest, spectral_flux, mfccs];

% Aggiungi la riga normalizzata alla matrice delle caratteristiche del periodo di riferimento

features_matrix = [features_matrix; features_line];

end

Every spectral feature is a vector of dimensions 498x1, do any of you know why 498 rows?

Thank you!

Answer 1

[Alan]

Hi Isabella,

From what I understand, you want to know why each spectral feature has a dimension of 498.

That is because the spectral features are calculated after computing the spectrogram of each audio slice (of 5 seconds as specified in the code). While calculating the spectrogram, each of these slices are further sliced up into “windows” to compute the Fourier Transform. By default, a Hamming Window of 1024 samples length with 50 samples overlapping is used. The term “windows” in this context is different from the “window_length” used in the shared code.

The larger the window size and smaller the overlap length, the longer the spectrogram. A longer spectrogram leads to longer feature vectors. The “window_length” parameter from the code will also proportionally affect the length of the feature vector.

To play with these parameters and quickly view outputs, I would highly recommend using the “Extract Audio Features” task in a live script file: https://www.mathworks.com/help/audio/ref/extractaudiofeatures.html

In the task, expand “Specify Window Properties”, and modify the window size and overlap length, and you will observe that the number of rows for each feature will change!

I would also recommend directly passing the audio signal into the ”Extract Audio Features” task instead of splitting the signal to chunks of 5 seconds and combining the feature matrices, as the task will automatically output a single matrix with combined features for the entire audio signal.

I hope this helped.