ASR decoding confidence scores (sentence and word) from lattice

[ntu.spe...@gmail.com]

Dear all, 

I have used chain model to get a good ASR system and use online2-wav-nnet3-latgen-faster to decode input utterance, the asr result is good. Thanks for sharing this awesome project.

Based on my understanding, online2-wav-nnet3-latgen-faster is used to produce a lattice, where some posteriors are stored. So my question is how to get word-level and sentence-level confidence score from produced lattice? For example, assuming decoded transcription is "today is sunshine", so how to get the confidence score of each word and the whole sentence?

Many thanks

Jiajia

[Daniel Povey]

Look into lattice-mbr-decode, it can output a format that is similar to/interpretable as word-level confidence scores.

There is a function SentenceLevelConfidence in lat/confidence.h that can give you something usable as a sentence-level confidence score, but I don't believe there is a binary for it.

None of these confidences are very well calibrated; they just reflect the information in the lattice.

Dan

--
Go to http://kaldi-asr.org/forums.html find out how to join
---
You received this message because you are subscribed to the Google Groups "kaldi-help" group.
To unsubscribe from this group and stop receiving emails from it, send an email to kaldi-help+unsubscribe@googlegroups.com.
For more options, visit https://groups.google.com/d/optout.