[Freeswitch-dev] VAD, TALK and NOTALK events

[Steve Underwood]

KMTDBOURSE wrote:
> Hi,
>
>   Thanks a lot for describing the vad algo it seems rather a basic 
> algo and I am surprised it is not more sophisticated. I hear also from 
> speex users (must be same algo?)  that the vad algo is very poor and 
> should be improved. That is what I am doing, try to improve vad and 
> echo cancellation algorithms. I would appreciate a link where I can 
> find data files (audio) to download on which I can perform tests. Of 
> course I will feedback the freeswitch community in case of improvment.
>
> best
>
> Zelda
There is a pretty good free echo canceller at www.rowetel.com, although 
FreeSwitch can't directly use it as it is GPL software. It does, 
however, mean that free and open workable algorithms for robust EC do exist.

VAD is different. It is still a research topic. You can see this from 
the way the VAD algorithms in things like G.729 have been revised over 
the codec's lifetime, as further research has been done. The algorithms 
G.729 and AMR use are encumbered by patents, otherwise cloning them 
would be excellent start. VAD for speech recognition is relatively 
straightforward. When you can buffer the audio, and look forwards and 
backwards in time, it become much easier to identify the start and end 
of voice bursts. It becomes hard when you can only look back in a low 
latency scenario.

Steve