Re: LSP Codebook

Very nice. Please post a patch too, to allow others to experiment.

I did a math-free talk at the ARRL East Bay Section Convention (San 
Francisco area) and Kristen McIntyre K6WX, the other speaker, asked if 
we had tried any hidden Marcov models. I'd said not yet. But it sounds 
like you used one to derive this quantizer, Brian?

     Thanks

     Bruce

On 06/21/2011 03:49 PM, David Rowe wrote:
> Sounds interesting Brian.  Yes, please post some samples.  12 bits for
> LSPs is very low.  One way to evaluate LSP quantiser is spectral
> distortion (SD).
>
> Thanks,
>
> David
>
> On Tue, 2011-06-21 at 19:07 +0100, Brian Smith wrote:
>    
>> If anybody's interested, I've done a 12-bit (4096 item) LSP codebook
>> based on the Voxforge corpus (22 hours of speech samples) for
>> quantization. Scoring is based on the sum of the deltas of the (integer)
>> log of the 10 LSP frequencies. Dropping the FPS is still a pain though,
>> was thinking of maybe doing 44.5fps (same as LPC10 and MELP), for
>> 44.5x12=534bps, plus 7-bit pitch at 10fps (can just about get away with
>> this) and 1-bit energy for a 600bps codec. Can post samples. 300bps
>> remains elusive.