Will AlignAtt streaming for Canary model support beamsearch?
As of now the AlignAtt implementation is seemingly supporting only greedy search. Simulstreaming (https://github.com/ufal/SimulStreaming) seems to have generalized this to beamsearch of any beamsize.
I am thinking of implementing a beamsearch unless it is supposed to be implemented in near-term.
Hi @mjhanphd,
You're right, the current AlignAtt implementation in NeMo only supports greedy Canary decoding. We have no plans to add beamsearch support in the near future. It would be great if you could contribute.
As far as I understand, the AlignAtt implementation in Simulstreaming only supports one audio file per decoding. In our case, we've implemented a batched greedy decoding for AlignAtt. Supporting an efficient batched beamsearch solution could pose some implementation challenges.