end of utterance detection for kaldi online2-tcp-nnet3-decode-faster

[Narjes Bozorgi]

Hello

I have a pre-trained nnet3 model that I am using for online2-tcp-nnet3-decode-faster. At the client side, I am using python-client with following format:

TCP-client:

open(file)

socket.sendall(packets)

data=socket.recv(packets)

txt=data.decode("utf-8")

The problem is  that the end of utterance at each refresh period is not detectable.

For both AsPIRE and our model the outputs from sox or python -client program doesn't include " or ' as Kaldi-asr.org explained:

The TCP protocol simply takes RAW signal on input (16-bit signed integer encoding at chosen sampling frequency) and outputs simple text using the following logic:

• each refresh period (output-freq argument) the current state of decoding is output
• each line is terminated by ''
• once an utterance boundary is detected due to endpointing a '
  ' char is output

Is there any option or any configuration needs to be set to activate this protocol?

[Hossein Hadian]

Have you set --endpoint.silence-phones correctly when running the server? 

It might help if you show the full command you use to run the server.

--Hossein

--
visit http://kaldi-asr.org/forums.html to find out how to join.
---
You received this message because you are subscribed to the Google Groups "kaldi-developers" group.
To unsubscribe from this group and stop receiving emails from it, send an email to kaldi-develope...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/kaldi-developers/b7053d16-f030-4046-a29e-98896efeb150n%40googlegroups.com.

[Narjes Bozorgi]

Thank you. Yes I am using --endpoint.silence-phones at online-config file. The server command is:

online2-tcp-nnet3-decode-faster --samp-freq=8000 --frames-per-chunk=20 --output-period=1 --read-timeout=1 --extra-left-initial=0 --frame-subsampling-factor=3  --min-active=200 --max-active=7000 --beam=15.0 lattice-beam=6.0  --acoustic-scale=1.0 --port-num=5050 --config=<pth>/online.conf  final.mdl HCLG.fst  words.txt