[Freeswitch-dev] VAD, TALK and NOTALK events
kmtdbourse at free.fr
Wed Mar 5 13:43:30 EST 2008
Thanks a lot for your e-mail, I got database 250 and seems very
helpful. I will let you know.
The interesting point in my technique is that:
1) the statistical test is theoretically characterized: I derived
the null na non-null hypothesis (I wonder whether this sounds to you).
2) I have the exact theoretical threshold, this is fixed
theoretically not experimentally
3) I also have what we call the power of the test.
These 3 points make it unique, that is what I think. It works very
well in noisy environment (very noisy).
I think you maybe of help to judge my results, can you?
At 00:44 20/02/2008, you wrote:
> > Know that
> > (OSLEC)
> > pretty well. Wanna test new algorithms but need wav data. Echo Cancel
> > is also still under research process, VAD is part of a good echo
> > canceller as it freezes the algo processing in case of double-talk.
> > Generally a modified Geigel algo or energy based algo are used.
>VAD and double talk are totally different problems.
> > I have new algorithms and want to test them. Unfortunately, I
> > don't have enough data, there are out there very expensive databases
> > such as (AURORA...) but don't think they are really suitable for my tests.
>If you are serious about VAD testing, the TIA has a downloadable
>database of hand tagged audio. If you are serious about echo
>cancellation, you will start by getting your canceller through the G.168
>tests, and then look at the real world test cases that fail. No
>expensive databases are needed in either case. I can't imagine a
>standard database even being very useful for line echo cancelller
>development. You will probably want to work out the difference between
>VAD and double talk first, though.
>Freeswitch-dev mailing list
>Freeswitch-dev at lists.freeswitch.org
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the Freeswitch-dev