cmusphinx / pocketsphinx

A small speech recognizer
Other
3.87k stars 713 forks source link

Do not attempt to align phones to impossibly short durations (fixes #339) #354

Closed dhdaines closed 1 year ago

dhdaines commented 1 year ago

While we explcitly disable bestpath in the command-line for phone and state alignment due to its tendency to put fake </s> with a duration of a single frame at the end of the utterance, this was not the case in the Python API, either because I forgot, but most likely for Reasons (some people apparently use bestpath with FSG search or alignment, don't ask me why).

This is a more robust fix, which will simply refuse to force-align something to an impossible duration. We may also wish to fix the upstream problem of the fake </s> but that is more complex to do. This at least prevents crashing!