1
$\begingroup$

I have written a vocal pitch shifting algorithm which uses LPC to separate the signal into its spectral envelope (predicted by the coefficients) and the pitch information (the residual error).

My understanding (and intuition) was that the residual signal would be unintelligble, but if I listen to the residual created by algorithm it is easily identifiable as a very low volume, slightly distorted version of the original signal.

Is this correct, or should I be getting what I originally expected?

Cheers, Ben

$\endgroup$
1
$\begingroup$

It sounds like you're doing things correctly.

Think of LPC as a compression method. A lossy method at that, if you discarded the residual and replaced it with a synthetic pulse train (which might consider using for pitch shifting).

In a perfect world where the LPC coefficients were a perfect match to the formant filtering applied by the throat/mouth/nose, the residual would be exactly the glottal impulse.

However, since the LPC is just an approximation, the residual contains additional information.

$\endgroup$

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.