Add delay_penalty to rnnt_loss; Add constrained rnnt; Fix monotonic_lower_bound to make it more efficient.

k2-fsa / fast_rnnt

A torch implementation of a recursion which turns out to be useful for RNN-T.

Other

136 stars 22 forks source link

Closed pkufool closed 1 year ago

pkufool commented 1 year ago

Fix #14

pkufool commented 1 year ago

Update changes from k2:

add delay_penalty in rnnt loss (https://github.com/k2-fsa/k2/pull/976)
Add constrained-rnnt (https://github.com/k2-fsa/k2/pull/1068)
Replace monotonic_lower_bound with efficient implementation (https://github.com/k2-fsa/k2/pull/1079)