noise reduction of voice sample

I have a voice samples. i will try to reduce the noises but my code is not working.

% Load the audio file

[y, Fs] = audioread('voice_sample.wav');

% Perform spectral subtraction

y_clean = spectralSubtraction(y, Fs);

% Write the cleaned audio to a new file

audiowrite('voice_sample_cleaned.wav', y_clean, Fs);

% Function for spectral subtraction

function y_clean = spectralSubtraction(y, Fs)

% Compute the short-time Fourier transform (STFT) of the audio signal

win_size = round(0.03*Fs); % Window size of 30ms

hop_size = round(0.01*Fs); % Hop size of 10ms

nfft = 2^nextpow2(win_size); % Number of FFT points

[S, f, t] = spectrogram(y, win_size, hop_size, nfft, Fs);

% Compute the power spectrum of the noisy speech

Pyy = abs(S).^2;

% Compute the power spectrum of the noise

noise_start = 1; % Starting frequency bin of the noise

noise_end = round(2000 / (Fs/nfft)); % Ending frequency bin of the noise (assumed to be below 2 kHz)

Pnn = mean(Pyy(:,noise_start:noise_end), 2);

% Compute the power spectrum of the speech

Pyy_est = max(Pyy - Pnn, 0);

% Reconstruct the cleaned audio signal

y_clean = real(ifft(S .* sqrt(Pyy_est ./ Pyy), 'symmetric'));

end

Answer 1

[Drishti]

Hi Sanusha,

I understand that you are trying to reduce noise from the audio signal, but the provided code is not working.

On debugging the code, I found that ‘y_clean’ variable is of dimension ‘1024 x 1147’ which signifies that ‘1147’ audio channels are present in the provided audio signal.

After referring to the MathWorks Documentation of ‘audiowrite’ function, I discovered that ‘audiowrite’ function does not support large number of audio channels like ‘1147’, hence the provided code throws an error as ‘Unsupported number of channels.’

For more information, you can refer to the MATLAB Documentation of ‘audiowrite’ function.

https://www.mathworks.com/help/releases/R2024a/matlab/ref/audiowrite.html#btiacgz-1-y:~:text=The%20maximum%20number%20of%20channels%20depends%20on%20the%20file%20format.

This issue can be resolved by converting the ‘y_clean’ channel into mono channel form by averaging the channels to maintain the balance of all available channels.

You can refer to the implementation below for better understanding:

% Ensure y_clean is a column vector if the input is mono
if size(y_clean, 2) > 1
    y_clean = mean(y_clean, 2);
end

I hope this resolves the issue.