I have just finished reviewing the code and realized that there was nothing particularly incorrect with the original implementation. However, I did notice that it lacked any visualization of the spectrogram and detected peaks, which made it challenging to diagnose the issue with the peak thresholding. To address this, I went ahead and added in the necessary visualization data. This allowed me to conclude that the peak thresholding was functioning correctly, and the issue with the code may reside in the threshold value selection procedure.
% read audio file
[y, fs] = audioread('tonev2018.mp3');
% preprocessing: resizing of signal
y = mean(y, 2);
y = y - mean(y);
% preprocessing: resampling
new_fs = 8000;
y = resample(y, new_fs, fs);
% spectrogram
window = 0.064 * new_fs;
noverlap = 0.032 * new_fs;
nfft = window;
[S, F, T] = spectrogram(y, window, noverlap, nfft, new_fs);
S = log(abs(S));
% peak detection
gs = 9;
peaks = zeros('like', S);
for shiftx = ceil(-gs/2) : ceil(gs/2)
for shifty = ceil(-gs/2) : ceil(gs/2)
CS = circshift(S, [shifty, shiftx]);
peaks = peaks + ((S - CS) > 0);
end
end
P = peaks == max(peaks(:));
% thresholding of peaks
tpeaks = [];
for i = 1 : 16 : size(P, 2)
istart = i;
iend = i + 15;
if iend > size(P, 2)
iend = size(P, 2);
end
sec = P(:, istart:iend);
flatten = reshape(sec, 1, size(sec, 1)*size(sec, 2));
sorted = sort(flatten, 'descend');
threshold = sorted(30);
sec(sec < threshold) = 0;
tpeaks = [tpeaks, sec];
end
% plot spectrogram
figure;
imagesc(T, F, S);
set(gca, 'YDir', 'normal');
xlabel('Time (s)');
ylabel('Frequency (Hz)');
title('Spectrogram');
% plot detected peaks
figure;
imagesc(T, F, tpeaks);
set(gca, 'YDir', 'normal');
xlabel('Time (s)');
ylabel('Frequency (Hz)');
title('Detected Peaks');
Got this result as output.
do let me know if it was helpful
