Home · Read · Listen · Forums · Develop · Downloads · About

Italian

Flat
Re: help me creare un modello acustico in italiano
User: dantavo
Date: 12/9/2013 10:19 am
Views: 163
Rating: 11

Thank you very much.

I follow your suggestions and now I run sphinxtrain1.0.8, but i've got this error:

Aligning results to find error rate word_align.pl failed with error code 65280 at /usr/local/lib/sphinxtrain/scripts/decode/slave.pl line 173.

 

Re: help me creare un modello acustico in italiano
User: nsh
Date: 12/9/2013 2:16 pm
Views: 175
Rating: 7

> Aligning results to find error rate word_align.pl failed with error code 65280 at /usr/local/lib/sphinxtrain/scripts/decode/slave.pl line 173.

You probably ran setup again and overwrote lm settings. You need to properly configure
$DEC_CFG_LANGUAGEMODEL  = "$CFG_BASE_DIR/etc/${CFG_DB_NAME}.lm";
You can get additional information from logs in logdir/decode.

Re: help me creare un modello acustico in italiano
User: dantavo
Date: 12/11/2013 10:17 am
Views: 169
Rating: 9

tks a lot for your reply, but the error message is still there.

Just to be clear, I summarize the step I did:

1. Installed Sphinxbase 0.8 and  Sphinxtrain 1.0.8

2. I ran sphinxtrain -t sino setup

  •  It generated sphinx_train.cfg (and I modified as suggested in the tutorialam)
  • feat.params 

3. I copied my files in etc dir

  • sino_train.fileids (you gave me)
  • sino_test.fileids(you gave me)
  • sino_test.transcription (you gave me)
  • sino_train.transcription (you gaveme)
  • sino.dic  (you gave me)
  • sino.phone (you gave me)
  • sino.vocab (I generated in my language model)
  • sino.lm.DMP(I generated in my language model)

4.  I ran sphinxtrain run command and I got the error:

Can't open /home/myuser/sperimentazione/sino/result/sino-1-1.match word_align.pl failed with error code 65280 at /usr/local/bin/sphinxtraing/scripts/decode/slave.pl line 173. 

I check my /logdir/decode/sino-1-1.log contains only two dates.

many thanks in advance.

 

I tried also to copy your sino-1-1.match in my result dir, and it run well, but I think this file should be generated from training process. 

 

 

Re: help me creare un modello acustico in italiano
User: nsh
Date: 12/11/2013 6:31 pm
Views: 156
Rating: 8

Empty log means you didn't install pocketsphinx decoder properly. Make sure you can run pocketphinx_continuous from the command line. Common reason of such errors is that you didn't define LD_LIBRARY_PATH as was suggested by tutorial.

 

Re: help me creare un modello acustico in italiano
User: dantavo
Date: 12/12/2013 4:31 am
Views: 163
Rating: 10

Thank a loooooooooooooot you are my guru :D

Now i will work to increase my database.

I hope to reach my aim although i know that now start the hard work ;) 

Could you tell me where I can found a brief or high level description of the parameters I can set in the sphinx_train.cfg file?

Thank in advance

Re: help me creare un modello acustico in italiano
User: nsh
Date: 12/12/2013 4:11 pm
Views: 306
Rating: 7

> Could you tell me where I can found a brief or high level description of the parameters I can set in the sphinx_train.cfg file?

Common parameters you need to configure are covered in the acoustic model training tutorial.
It's not advised to change other parameters.
All parameters are briefly described in comments in sphinx_train.cfg file.
Some parameters do not have a high-level description because you need to understand the internal algorithms involved in order to change them. For full understanding you need to read a text book "Spoken language processing: a guide to theory and algorithms" or any other book on speech recognition.

Re: help me creare un modello acustico in italiano
User: dantavo
Date: 12/20/2013 2:57 pm
Views: 4093
Rating: 6

thank you, the book you've suggested me is fantastic.

But I've a question, I've noted that thus my training db is very small (less than 1 hours) I have best result when the number of tiedstate (or senones) is 600 than 200.

I don't understand why this happens i tryed 600 because i have 40 phones in my dictionary and 15 people's voices, so i setted tiedstate=600 in the cfg file but is only a test.

I reach WER and SER value better than i use tiedstate =200, and the hellongram show me that the recognition is well.

Could you explain me why?

Previous • Next •
Unless otherwise indicated, © 2006-2024 VoxForge; Legal: Terms and Conditions