Csound Csound-dev Csound-tekno Search About

Re: LPC - how do you make it work well?

Date1999-02-25 13:06
Fromrasmus ekman
SubjectRe: LPC - how do you make it work well?
Erik Spjut wrote:
> 
> lpreson and lpfreson interpolate between frames but unless kr=sr you'll get
> a buzz or noise at kr. I suspect that's your burble.

Huh? Isn't the burble a rather common side-effect of FFT processing:
It comes from cross-fading the analysis/resynthesis windows, when
the sine-waves in each frame (window) are faded up/out.


Standard FFT is old, primitive and grotty and should be replaced by 
multi-band sine-wave analysis (not the wavelet sine packets used for 
data reduction) right now. I remain amazed that maths-proficient 
computer musicians refuse to shower themselves in glory by developing
some such highly useful tools! 
(spectrum is a start though)


Date1999-02-25 19:25
FromErik Spjut
SubjectRe: LPC - how do you make it work well?
Like Charlie said LPC != FFT. Without getting too technical or flaming too
much, FFTs are like all-zero (or FIR filters), LPC's are all-pole (or IIR)
filters, the exact reciprocal of the FFT method. With either method you
trade off frequency resolution and time resolution by block size and
overlap. All of the other analysis methods (such as wavelets or filter
banks) involve EXACTLY the same tradeoffs. In particle physics it's called
the Heisenberg uncertainty principle. The only change is where/how easily
the tradeoffs are made. I find LPC works well for some things, pvoc works
well for some things, adsyn works well for some things, but they can all be
made to produce garbage. We bathe ourselves in glory where we can.

At 2:06 PM +0100 2/25/99, rasmus ekman wrote:
>Erik Spjut wrote:
>>
>> lpreson and lpfreson interpolate between frames but unless kr=sr you'll get
>> a buzz or noise at kr. I suspect that's your burble.
>
>Huh? Isn't the burble a rather common side-effect of FFT processing:
>It comes from cross-fading the analysis/resynthesis windows, when
>the sine-waves in each frame (window) are faded up/out.
>
>
>Standard FFT is old, primitive and grotty and should be replaced by
>multi-band sine-wave analysis (not the wavelet sine packets used for
>data reduction) right now. I remain amazed that maths-proficient
>computer musicians refuse to shower themselves in glory by developing
>some such highly useful tools!
>(spectrum is a start though)
>
>
>        re


-------------------------------------------------------------------------------
Erik Spjut (spyoot, rhymes with cute) - Associate Professor of Engineering
and  Associate Director for Engineering Computing,  Center for Design Education
Harvey Mudd College, Claremont, CA 91711-5990  USA
Erik_Spjut@hmc.edu      Ph & Voice mail (909) 607-3890      Fax (909) 621-8967