• PocketSphinxFixes

Stuff that needs to be fixed in PocketSphinx

1. The "main page" part of the API documentation needs to be written.
2. Frame reporting in hypotheses is broken
   • Needs to report input frame index, not dynamic frame index
3. Fixed-point support for continuous density models

   • This will happen by way of the generic GMM support in SphinxBase. Mini-TODO:

     1. <gau_mix.h> needs to be implemented

     2. Figure out how fixed-point is going to be handled here
4. Decoder initialization is weird and bogus
   • Don't initialize the FSG or the LM search unless they're actually being used
   • Put all decoder state in the same place and merge the kb, uttproc, dict, etc. APIs
   • Begin the long and arduous task of making the decoder re-entrant.
5. Fixed-size buffers are everywhere
   • Use a blockarray-list type scheme for the backpointer table
   • Use a growable array for acoustic features
   • Make the effective maximum utterance length 32767 frames (using int16 as frame counter). Long enough for anyone?

Stuff that would be nice to fix in the future

1. Make the dictionary and lextree dynamically growable