Is rescoring implemented

My question is whether rescoring has been implemented so as to efficiently use NNLM models. Either in the form of rescoring a lattice or (I guess speechbrain doesn’t have lattices? not sure) reranking nbest results.

The motivation for this is because doing fused decoding with an NNLM is very slow and therefore impractical.


We are integrating K2 (see the PR on the repo) for WFST decoding. For now we only support joint decoding with NNLM (BeamSearch with shallow fusion) or n-gram decoding.

1 Like

Okay thank you I’ll take a look at the PR! :slight_smile: