missing words while do-endpointing is true

[wu poyu]

I tried to do speech recognition on a long wav. While I set do-endpointing false, it can decode complete sentence. However, if I try to do endpoint detection, words between segment sometimes miss. I tried to figure out that missing words' wave shows in previous segment (by counting chunk time).

So my question is:

1. what is the possible reason for this situation? I simply guess that it's because of language model (i.e. "<s> how" probability is much higher than "how </s>" , so the words, how, maybe miss in the end of wave), but I'm not sure.

2. Do you suggest to use "EndpointDetected" or just simply break it up into smaller chunks(maybe per 5 minutes)?

Thank you.

[an...@sayint.ai]

You could break them into smaller chunks using VAD

[Daniel Povey]

The --do-endpointing is not really segmentation.  It is for the scenario where you want to just stop the recognition at a certain point (e.g. to demonstrate how you might do that for an interactive application).  So it will just discard the rest of the file.  

The best way is to break it up into smaller chunks- preferably not longer than about a minute.  If you make them slightly overlapping and generate a ctm file relative to the utterance (steps/get_ctm.sh with --use-segments false), then I think you can resolve the overlaps using utils/ctm/resolve_ctm_overlaps.py.

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.
To post to this group, send email to kaldi...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/kaldi-help/6b4aa19d-a1b5-4220-956d-cedd69e01539%40googlegroups.com.

For more options, visit https://groups.google.com/d/optout.