0
$\begingroup$

I am playing with spleeter to separate voice from an audio signal.

They go to the frequency domain by computing the Short-time Fourier transform to create a spectrogram. As there are some approximation (windowing and overlap of window) in this calculus, I was wondering if it hurts the quality of the signal when one computes stft and then inverse-stft.

$\endgroup$
1
$\begingroup$

STFT (short-time Fourier Transform) should not impact the quality of the reconstructed signal if properly implemented.

Furthermore, any signal $x[n]$ should be exactly recovered from its STFT analysis coefficients by a proper implementation of the associated inverse STFT;i.e., The forward and inverse pairs should exactly undo each other.

One should keep in mind about the overlap parameter and window functions however;i.e., if the window includes zeros, the inverse may not be possible at some overlaps. If the window does not contain any zeros, then the inverse is always computed.

| improve this answer | |
$\endgroup$
1
$\begingroup$

By design, quality should be preserved: STFT is a redundant and fully inversible transform, when time-frequency data remains complex. Note this is not the case anymore when complex coefficients are turned into absolute values or squared-magnitudes, as common for "spectrograms".

But as long as you preserve the redundant complex values, the novel representation better separates signal's components' and help further processing. And optimal inverses, since many inverses can recover a redundant transformation, can help a lot.

| improve this answer | |
$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.