Re: [Cmusphinx-devel] [Fwd: latest on live demos]

[Paul L. <Paul.Lamere@Sun.COM>]

Phil:

I tried things out today with all of your latest updates. It is really 
performing well on my 1.4 ghz linux box. It's quite snappy now.

Thanks much.  I hope Peter and the MIT folk will give it another try.

Paul


Philip Kwok wrote:

> A resend. sourceforge seems to have eaten my emails again...
>
> ------------------------------------------------------------------------
>
> Subject:
> latest on live demos
> From:
> Philip Kwok <Philip.Kwok@Sun.COM>
> Date:
> Tue, 21 Oct 2003 16:28:25 -0400
> To:
> cmusphinx-devel <cmu...@li...>
>
>
> Hi,
>
> Two main changes in the live demo. Please update your live demos 
> accordingly:
>
> 1. The default values for Live CMN shiftWindow has been reduced from 
> 800 to 80, and the windowSize from 500 to 50. This should reduce the 
> latency in live decoding. You don't need to change anything for this. 
> Thanks to Paul for testing this out. Paul, do you mind trying it again 
> on your machine to see if the decoding is as bad. I'm suspecting that 
> the energy thresholds are not tuned for your home machine.
>
> 2. The FrameDropper is implemented, and incorporated into the 
> live-demos. In live hands-free mode, it drops every other frame, and 
> replaces it with the previous frame. This gives faster decoding speed, 
> with no lost of accuracy. In batch-mode, it runs the ripley test in 
> 1.57 xRT vs 1.87 xRT  on my slow machine. I've updated 
> tests/live/Makefile "live-free" tests to use the FrameDropper. To 
> activate it:
>
> edu.cmu.sphinx.frontend.FrameDropper.dropEveryNthFrame=2
> edu.cmu.sphinx.frontend.FrameDropper.replaceNthWithPrevious=true
>
> I've tested these changes on Solaris and Linux. The live demos are 
> running faster.
>
> My next task is to look at other endpointing strategies. I'll try 
> Bent's self-adapting endpointer. The hate the fact that our 
> EnergyEndpointer needs re-calibration of thresholds every time you 
> switch machines.
>
> philip
>