Closed Yymax-max closed 2 years ago
It should be likelihood_[i] = -(weight.Value1()+weight.Value2()) if we want the sum of acoustic score and lm score.
It should be likelihood_[i] = -(weight.Value1()+weight.Value2()) if we want the sum of acoustic score and lm score.
thank you for your answer ,i approve your answer, but why ctc_wfst_beamsearch.cc is likelihood[i] = -weight.Value2(),-weight.Value2() is likely acoustic score , only add it ,what the mean of the tlg
https://github.com/wenet-e2e/wenet/pull/1116
Please see the PR, we use both acoustic and graph cost now.
for (int i = 0; i < nbest; i++) { kaldi::LatticeWeight weight; std::vector alignment;
fst::GetLinearSymbolSequence(nbestlats[i], &alignment, &outputs[i],
&weight);
ConvertToInputs(alignment, &inputs[i], ×[i]);
RemoveContinuousTags(&outputs[i]);
likelihood[i] = -weight.Value2();
}
in ctc_wfst_beamsearch.cc why likelihood[i] = -weight.Value2() is not likelihood_[i] = -(weight.Value1()+weight.Value2())?