[FFmpeg-devel] [Brainstorming] sound reproduction with a 3 voice soundchip

Tobias Bindhammer tobias.bindhammer
Tue Mar 1 15:49:43 CET 2011

> If you dont care about encoding speed, some form of iterative minimization
> of the difference seems like an interresting option.

Not at all, quality stands above speed first of all.

> "difference" here can be some sum of squared differences or more advanced
> psyco acoustic thing. some FFT ignoring or downweighting phase could be tried
> too
> do until no more improvment can be achived
>     for each set of parameters that you can optimally choose, do choose it
>     optimally
> for example such a set could be frequency+volume+waveformtype
> when you choose a frequency+waveform by bruteforce then volume can be
> found by correlation.

i'm somewhat scared of bruteforceattempts, because of the possible
amount of parameters i can take into the calculation as variables, what
would make the possible number permutations explode. Sure, there are
possibilities to stop certain branches, or one could try to split up the
data into several partitions. However i am not sure how things behave.
Error might increase hevily on a small change, and fit perfect on the
sligzhtest change as well, so iteration might be kind of hard, right? On
the other hand, can i guarantee, that the oscillators work within the
same range? In fact i get anyway only close to that what i calculate, no
calibration happens at all (would be possible however with some effort)

> i dont know what filters are available and how many bits their parameters have
> nor if they can be found optimally quickly or require bruteforce. It might
> or might not make sense to included them in some bruteforce search or maybe
> there are faster ways to find some of these near optimally. Ive not really
> thought about it
> In the end its alot of trial and error to find the best algorithm i think

It has been so far as well :-) But giving a few tries with multiple
waveforms and few parameters to see how bruteforce performs would ba a
good start i think. Maybe best tested with some easy synthetic waveforms
to see if works acurate enough.


Dipl. Ing. Tobias Bindhammer
Institut f?r Verteilte Systeme
Oberer Eselsberg          Phone: + 49 731/502-4235
Universit?t Ulm           Fax  : + 49 731/502-4142
D-89069 Ulm               mailto:tobias.bindhammer at uni-ulm.de

More information about the ffmpeg-devel mailing list