search

aalto-speech / AaltoASR

Aalto Automatic Speech Recognition tools
Other
85 stars 37 forks source link

Fix decoder running out of tokens #11

Closed vsiivola closed 10 years ago

vsiivola commented 10 years ago

Please delete this request.

With long noisy silences, the decoder may first prune all silence paths, leaving only phonemes with duration models. Later, MAX_STATE_DURATION removes theses tokens and decoder asserts(false). Happens only in rare cases.

This patch is an ugly fix, keeping at least 10 tokens with long durations alive and flooring the duration_prob so that it is still possible to exit the state through exit transition (otherwise will only self-transition).

vsiivola commented 10 years ago

Please remove this pull request - I forgot to branch the repo correctly. A new pull request to a branch coming up.