Forums

Detecting spectral envelope

Started by jungledmnc December 19, 2010
Hi,

I'm using the most common audio pitch shifting algorithm via FFT. It works
fine, but not for shifting / keeping formants, especially for vocals. Do
you have some links for well working algorithms/ideas for detecting
spectral envelope used for this purpose?

I tried several ways and ended with an exceptionally simple thing - just
some moving average of a magnitudes of each FFT block. 

Thanks in advance.
On Dec 19, 10:17&#2013266080;am, "jungledmnc" <jungledmnc@n_o_s_p_a_m.gmail.com>
wrote:
> > I'm using the most common audio pitch shifting algorithm via FFT.
how do you know what is the most common audio pitch shifting alg? i would be interested, since i haven't done the research of the market nor digging in and reverse engineering the common pitch-shifting algs. my guess might be that Autotune is the most common audio pitch shifting alg. i don't think it's a frequency-domain method, i think it's a time domain method. if they use the FFT for anything, it's not that it's in the signal processing chain
> It works > fine, but not for shifting / keeping formants, especially for vocals. Do > you have some links for well working algorithms/ideas for detecting > spectral envelope used for this purpose?
do you know about the old Digitech Vocalist? it's a time-domain alg (so no FFT) based on a paper by Keith Lent in Computer Music Journal in 1989. i did an analysis of it in 1993 and 1995 in the AES Journal. want a pdf copy? it's not frequency domain and needs a good pitch detector, so even without FFTing (and iFFT back) the pitch detector is not a low computational burden. but it retains the formant locations (the spectral envelope) without going into the frequency domain.
> I tried several ways and ended with an exceptionally simple thing - just > some moving average of a magnitudes of each FFT block.
sounds about as good as anything.
> Thanks in advance.
FWIW. r b-j
Sorry, I wasn't explaining myself correctly - I meant phase vocoder using
FFT. I think it is the most used one. Time domain methods usually don't
work very well, especially with aperiodical signals (which polyphonic
usually are and vocals aren't so good either). I mean, for telephony and
stuff it may work, but for music, I don't really think so.

It is also quick common to "shift" formants by some amount, so it is needed
to be able to extract the spectral envelope somehow.

jungledmnc

>On Dec 19, 10:17=A0am, "jungledmnc" <jungledmnc@n_o_s_p_a_m.gmail.com> >wrote: >> >> I'm using the most common audio pitch shifting algorithm via FFT. > >how do you know what is the most common audio pitch shifting alg? i >would be interested, since i haven't done the research of the market >nor digging in and reverse engineering the common pitch-shifting algs. > >my guess might be that Autotune is the most common audio pitch >shifting alg. i don't think it's a frequency-domain method, i think >it's a time domain method. if they use the FFT for anything, it's not >that it's in the signal processing chain > >> It works >> fine, but not for shifting / keeping formants, especially for vocals.
Do
>> you have some links for well working algorithms/ideas for detecting >> spectral envelope used for this purpose? > >do you know about the old Digitech Vocalist? it's a time-domain alg >(so no FFT) based on a paper by Keith Lent in Computer Music Journal >in 1989. i did an analysis of it in 1993 and 1995 in the AES >Journal. want a pdf copy? it's not frequency domain and needs a good >pitch detector, so even without FFTing (and iFFT back) the pitch >detector is not a low computational burden. > >but it retains the formant locations (the spectral envelope) without >going into the frequency domain. > >> I tried several ways and ended with an exceptionally simple thing -
just
>> some moving average of a magnitudes of each FFT block. > >sounds about as good as anything. > >> Thanks in advance. > >FWIW. > >r b-j >