[Asterisk-Dev] <<< SPHINX INTEGRATION >>>

[Steve Underwood]

Eric Wieling wrote:

> jeff quade wrote:
>
>> Howdie:
>> Thought Id just post to see if anyone has had any success with 
>> integrating CMUs SPHINX into Asterisk?
>> I see the EAGI in the source/build, and a couple of posts to 
>> Asterisk-Users, but no useable info.
>> Seems like a DEAD issue, waiting for a re-visit.
>> Before I pull the code apart and struggle with deciphering the 
>> source-- I thought someone might have pointers or suggestions.
>
>
> Someone, I don't remember who, on IRC #Asterisk got it working, but 
> only with ULAW/ALAW codec.  Apparently the sound quality on the 
> compressed codec was not quite good enough for Sphinx.  I don't know 
> if he ever resolved the problem.  I imagine you could tune Sphinx to 
> work better with compressed codecs.

None of the speech recognition engines achieves much success if the 
voice has been compressed. In fact, they are even less reliable than 
usual. :-)

Seriously, you can probably get reasonable results with ADPCM. Any more 
sophisticated codec (pretty much anything at a lower bit rate) is 
unlikely to leave the voice close enough to intact to be workable. Even 
ulaw and alaw have issues. Most speech recognisers would much rather 
have at least 16k samples/second input. At that rate unvoiced sounds 
have more character. Things like "s" and "f" sound too similar at 8k to 
tell the difference - we resolve them by context, but computer 
recognisers are very good at that. Telephony recognisers can be 
optimised to reduce this limitation, but its always there.

Regards.,
Steve