Currently the w2l layer uses fixed rejection settings that have been chosen for the 77MB acoustic model. The settings haven't been optimized much and also don't transfer to other models. It would be helpful to be able to control them when setting up the w2l engine.
The main reason why they don't transfer is that differently sized acoustic models produce emission frames of different durations. And some of the settings use "number of emission frames" as a unit.
Currently the w2l layer uses fixed rejection settings that have been chosen for the 77MB acoustic model. The settings haven't been optimized much and also don't transfer to other models. It would be helpful to be able to control them when setting up the w2l engine.
The main reason why they don't transfer is that differently sized acoustic models produce emission frames of different durations. And some of the settings use "number of emission frames" as a unit.