On Sep 22, 2011, at 5:29 AM, Lorenzo Sutton wrote:
On 21/09/2011 22:44, Hans-Christoph Steiner wrote:
I'm trying to come up with a simple voice scrambling technique that leaves voices understandable, but makes them unrecognizable. A key
part of this is to make it very hard to reverse the scrambling to make the voice more recognizable.Ineresting. Also these possible approaches come to mind:
- Trivial, probably off-topic: use a TTS. Only thing that is
recognisable is the software which produced it.
- Use some sort of granulation with random pitch-shifting. Not sure
how easily reversible this would be.
A fun thing might be mix the two: that is start from a TTS and then
feed it into the granular... Here a silly example made with Festival
TTS and granulation in Granita (the first sentence is the 'original'
produced by Fesival) with different settings and degrees of
recognisability:
This paper has some good discussion on the technique:
http://www.efjohnsontechnologies.com/resources/dyn/files/75835/_fn/Selected_...
The hard thing to defeat would be statistical analysis techniques.
Perhaps the idea is to try to modify to voice to sound as much as
possible like a standard voice, like make them all have the same base
pitch, then add a random warble.
.hc
¡El pueblo unido jamás será vencido!