GSoC Project status update 2: Speech Recognition in Open Moko
erland at lewin.nu
Mon Jun 16 22:34:23 CEST 2008
2008/6/16 saurabh gupta <saurabhgupta1403 at gmail.com>:
> So, I have written fixed point codes in which all calculations are handled
> through fixed point. I have used the 16:16 notation for fixed point and used
> 32 bit integer to represent a fixed point.
FWIW, when I wrote an embedded speech recognition engine, I used 16 bits for
most of the data in the models etc. If you could manage to store values in
16 bits rather than 32, then I your memory requirements could be halved, and
probably performance increase significantly, since the cpu to memory
bandwidth would be lower, and I think that might be a performance
bottleneck. However, it required careful tracking of what range of values
were actually required in various stages of the algorithm.
Multiplication and division procedures for fixed point are implemented
> through macros. As the project progresses, the stress will be given to
> optimize the codes
I hope you won't have to do much if any division because division is quite
slow. If there are variables you need to divide by, try storing 1/x instead
of x, and then doing multiplication instead of division if you can manage
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the community