cmusphinx / pocketsphinx

A small speech recognizer
Other
3.9k stars 714 forks source link

FSG recognition with `-bestpath` is often harmful #318

Open dhdaines opened 1 year ago

dhdaines commented 1 year ago

In the past, bestpath was simply disabled for FSG search, because it caused a great slowdown and did not improve accuracy. In addition, it just flat-out breaks word-alignment so must be disabled when using PocketSphinx for alignment. Now it is enabled again, and I'm not sure this is a good thing. We need it in order to get posterior probabilities, though.

One possible source of these problems seems to be the lattice generation from FSG search, which contains some probably incorrect heuristics for creating the start and end nodes.