Closed francisr closed 1 year ago
Maybe the sign is wrong? IIRC k2 uses scores, not costs.
Hmm, quite silly, in this code I was negating the kaldi matrix when reading it into the loglikes
tensor, because I knew about k2 using scores rather than costs, but it works when I read the loglikes as is.
Thanks.
Hi, I'm trying to decode a Kaldi trained model with the k2 online decoder.
Here's what I'm doing:
However the output is non-sense words, not sure what I'm doing wrong.