search

alumae / kaldi-gstreamer-server

Real-time full-duplex speech recognition server, based on the Kaldi toolkit and the GStreamer framwork.
BSD 2-Clause "Simplified" License
1.07k stars 341 forks source link

warning while generating word alignments #29

Open rohithkodali opened 9 years ago

rohithkodali commented 9 years ago

Hi,

i'm getting an error when i try to get the word alignments, it is not happening always, it throws a warning and it doesn't provide the word alignment for that particular segment and for the other segments it works fine,

here is the warning

Lattice has input epsiilons and/or non-deterministic (in mohri's sense) --i.e., lattice is not deterministic. Word alignment may be slow and/or blow up ,memory

when i decode the same file with sgmm based models (offline approach) it dint throw any error and provide all word-alignments perfectly. what is the possible cause for the error and how to solve it?

alumae commented 9 years ago

Yes, this issue is known: https://github.com/alumae/gst-kaldi-nnet2-online/issues/22

I don't know currently how to fix it.

On Wed, Sep 23, 2015 at 9:27 AM, rohithkodali notifications@github.com wrote:

Hi,

i'm getting an error when i try to get the word alignments, it is not happening always, it throws a warning and it doesn't provide the word alignment for that particular segment and for the other segments it works fine,

here is the warning

Lattice has input epsiilons and/or non-deterministic (in mohri's sense) --i.e., lattice is not deterministic. Word alignment may be slow and/or blow up ,memory

when i decode the same file with sgmm based models (offline approach) it dint throw any error and provide all word-alignments perfectly. what is the possible cause for the error and how to solve it?

— Reply to this email directly or view it on GitHub https://github.com/alumae/kaldi-gstreamer-server/issues/29.

russlevy commented 7 years ago

I found this was an issue in nnet2 when the utterance was longer than 20s so kaldi split it -- all intermediate utterances did not get word alignments. We couldn't reproduce this when we enabled nnet3 + chain.

star633669 commented 7 years ago

Can we reduce the utterance duration to 10s and then split it? or split it with silence?