Re: Need to hire a software developer for a voice processing app!
Re: Need to hire a software developer for a voice processing app!
- Subject: Re: Need to hire a software developer for a voice processing app!
- From: Richard Dobson <email@hidden>
- Date: Sat, 20 Feb 2010 11:10:32 +0000
On 19/02/2010 18:20, Thomas David Kehoe wrote:
Smule has a pitch-shifting app with good sound (and a lousy user interface).
I don't know how they did it. Artefactsoft has a delay and pitch-shifting
app (for treating stuttering), the pitch-shifting has horrendous sound, the
delay sounds pretty good (and their user interface is good). I'd like to
know how Smule implemented pitch-shifting.
From their demo, it sounds a bit "granular", so maybe a time-domain
technique (PSOLA etc; indeed, you can get a quite respectable
pitch-shift using classic granular synthesis techniques). the smoothest
shifters tend to need overlapped FFTs (classically, using pvoc), with
phase locking etc.
It all comes down to how many FFTs of sufficient size (in both
directions) the iPhone can sustain. With the FFT the latency gives you
some of the delay for free, so to speak. And if "audiophile" shifting is
not needed (i.e. a little transient smearing can be tolerated, usually
not bad at all with voice) the plain vanilla pvoc pitch shift technique
(small fraction of the cost of the FFT stages) can be perfectly adequate
and very efficient. I even posted an open source (real-time) VST example
years ago which is still rolling around the net. Assuming the iPhone can
carry the dsp, that aspect is all relatively known territory (e.g. it
can all be found in Csound, and <plug> in the CDP system </plug>). What
time-scale is targetted for this project? Until I actually have FFTs etc
implemented on an iPhone (presumably in 8.24 fixed-point), or someone
posts reliable benchmarks, it is difficult to make a firm prediction of
the likely CPU load for the whole thing. Pity Apple's streaming FFT
spectral framework is not available on the iPhone!
A Mac of course can do it without a blink; I have had a stereo real-time
pvoc engine running as a VST for years now, typical CPU load averaging
around 10% for "reasonable" analysis rates (25% overlap) and FFT sizes.
Half that for mono of course.
Wot we really need is a counterpart to CUDA/OpenCL for the iPhone - FFTs
are classed as "embarrassingly parallel", so a massively-parallel f/p
accelerator chip for the iPhone would be "game-changing".
Richard Dobson
_______________________________________________
Do not post admin requests to the list. They will be ignored.
Coreaudio-api mailing list (email@hidden)
Help/Unsubscribe/Update your Subscription:
This email sent to email@hidden