1
$\begingroup$

I am implementing a voice morphing system, but I am having difficulties getting a good quality output signal. The method I am using is LPC analysis and synthesis. In theory, the LPC process should split the signal into a source/filter representation by calculating a residual and a set of filter coefficients.

However, in practice, I am finding that the residual contains data which I believe should be contained in the filter.

What techniques are available for better separating an input signal into source and filter?

Cheers, Ben

$\endgroup$
  • 1
    $\begingroup$ Interesting question. Can you share a link to your sound files original and morphed. The question can be better answered once I hear the distortion you are referring to. $\endgroup$ – Neeks Oct 14 '15 at 6:02
  • $\begingroup$ One thing I've read which might be helpful to separate source and filter is calculating cepstrum. In frequency, filters and singals are meant to be multiplied. If you compute cesptrum, then multiplication becomes addition, and you may get clear separation between filter factor and a signal. Another thing to clear the signal is deconvolution. But this may needs transfer function (impulse response) of the environment (filter) that your signal is affected by. Hope it can help. $\endgroup$ – Celdor Oct 14 '15 at 8:31
0
$\begingroup$

Have you considered using a higher order filter for your LPC analysis? LPC analysis is known to be sensitive to the presence of noise, say additive, so it may be useful to preprocess/denoise the signal to have the best/cleanest signal going in. Are you using a microphone on a mobile device?

If those don't work, you should look into research on spectral envelope extraction, or alternative algorithms for voice-morphing.

$\endgroup$
  • $\begingroup$ I have tried a lot of different filter orders, but the difference is not significant. The sample I am using for my work is a clean recording done through a professional grade microphone so that's not the issue. I was thinking about looking into alternative methods, but I want to be sure that LPC is a dead end first. A lot of literature on the subject suggests that it is a viable method for performing constant speed pitch changes on vocals, but as yet I am unable to verify any of these claims - either through my own work or from any 3rd party research. $\endgroup$ – bennji_of_the_overflow Mar 18 '15 at 22:37
0
$\begingroup$

WORLD splits a signal into an f0 array, a spectral envelope matrix, and a non-harmonic components matrix. You can then recombine these to recover the original signal.

$\endgroup$

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.