Allow zero length references for rnnt loss

[desh2608]

Based on some discussions with @danpovey, it seems the RNNT loss computation itself does not have any inherent requirement of references being non-zero. Allowing this relaxation would be useful for training with possibly empty references, e.g. cases where utterance is just noise or silence, or for multi-talker ASR (e.g. see k2-fsa/icefall#845).

[csukuangfj]

I think you can also remove the constraint

assert T >= S

That constraint is required only for modified transducer, where it can only move horizontally or diagonally.

[desh2608]

I think you can also remove the constraint

assert T >= S

That constraint is required only for modified transducer, where it can only move horizontally or diagonally.

Okay, I will make that constraint conditional on if the rnnt_type is modified.

BTW I suppose if we remove the T>=S assertion, we can also get rid of the corresponding data filtering in icefall recipes.

[csukuangfj]

Thanks. Would you mind also creating a PR in the repo danpovey/fast_rnnt ?

[csukuangfj]

Will merge it tomorrow. Thanks!

[csukuangfj]

Thanks for your contribution!