k2-fsa / snowfall

Moved to https://github.com/k2-fsa/icefall
Apache License 2.0
143 stars 42 forks source link

Get best matching stats from two nbest objects #236

Closed pkufool closed 2 years ago

pkufool commented 2 years ago

Functions to generate training features for special model for nbest score estimation.

Require https://github.com/k2-fsa/k2/pull/786

pkufool commented 2 years ago

@csukuangfj I tried to generate training data from just one Nbest object, but it is too complicate, we need the init score (before second pass recoring) as one of the features, if with just one Nbest object (after rescoring) we should pass one more parameter, say, old_scores. So, I thought passing two Nbest objects to the function would be the simplest way. To generate more training data(by swapping keys and queries), we have to do second pass rescoring two times, one for each Nbest object.

danpovey commented 2 years ago

The training time for this special is probably going to be very small, so I wouldn't worry too much about it.