Forums

transmitting phase spectrum in MELP?

Started by Ilya Druker June 29, 2005
Ladies & Gentlemen

I encountered the problem of transmission of music by MELP codec. MELP
is intended for coding human speech, so it is not surprising that the
quality of the music synthesized by MELP is very low. One of the
problems is failure of the pitch detector to recognize
pitch-doubling/tripling in music. I think that transmission of
phase-spectrum together with Fourier Magnitudes would partially solve
this problem. But how to do that?

Does anybody know anything about phase-spectrum transmission/synthesis
in MELP and what effect it makes on coding music?

Thank you, Ilya Druker



Ilya-

> I encountered the problem of transmission of music by MELP codec. MELP
> is intended for coding human speech, so it is not surprising that the
> quality of the music synthesized by MELP is very low. One of the
> problems is failure of the pitch detector to recognize
> pitch-doubling/tripling in music. I think that transmission of
> phase-spectrum together with Fourier Magnitudes would partially solve
> this problem. But how to do that?
>
> Does anybody know anything about phase-spectrum transmission/synthesis
> in MELP and what effect it makes on coding music?

We made progress here by first improving pitch detection.

-Jeff



I have found the answer to my question here:
"High Quality MELP Coding at Bite-Rates Around 4Kb/s" by J.Stachurski,
A.McCree & V.Viswanathan.

They checked transmitting complex Fourier coefficients (i.e. actually
phase spectrum) and found that it rather degrades the speech quality.
My guess is that this is due to limited resolution of phase spectrum
which actually leads to dis-alignment of successive pitch-periods or
breaking periodicity. They also tried to improve the algorithm by
smoothing successive phase spectra and by limiting maximal phase
variation by predetermined constant - however, the achieved quality
was not so good as with fixed phase spectrum.

It seems, thus, that there is no a good solution to the problem.

Ilya Druker

--- In speechcoding@spee..., "Ilya Druker" <ied1970r@h...> wrote:
> Ladies & Gentlemen
>
> I encountered the problem of transmission of music by MELP codec. MELP
> is intended for coding human speech, so it is not surprising that the
> quality of the music synthesized by MELP is very low. One of the
> problems is failure of the pitch detector to recognize
> pitch-doubling/tripling in music. I think that transmission of
> phase-spectrum together with Fourier Magnitudes would partially solve
> this problem. But how to do that?
>
> Does anybody know anything about phase-spectrum transmission/synthesis
> in MELP and what effect it makes on coding music?
>
> Thank you, Ilya Druker