Endpointing with fast_beam_search ends up with very short segment compared to modified_beam_search and greedy_search. Took a look into the code and it seems that the num_trailing_blanks does not seem to get reset when a non-blank token is seen. (Missing a else condition to set p->num_trailing_blanks = 0; as suggested in https://github.com/k2-fsa/sherpa/issues/229)
Endpointing with
fast_beam_search
ends up with very short segment compared tomodified_beam_search
andgreedy_search
. Took a look into the code and it seems that thenum_trailing_blanks
does not seem to get reset when a non-blank token is seen. (Missing a else condition to setp->num_trailing_blanks = 0;
as suggested in https://github.com/k2-fsa/sherpa/issues/229)https://github.com/k2-fsa/sherpa/blob/3d4a3321a8acfa276e6179a4c9aa81f12607996f/sherpa/csrc/online-transducer-fast-beam-search-decoder.cc#L115-L135