How to access to the weight of each obtained output phoneme ?

[GoogleCodeExporter]

What steps will reproduce the problem?
1. After executing the phonetisaurus-g2p command, the trained model will output 
the phoneme sequence corresponding to the input text. In addition, it is also 
included the output score of the obtained phoneme sequence (i.e. zoom    
0.50446 z uh m). 

What is the expected output? What do you see instead?

I would like to know how I can access to the weight of each output phoneme 
(e.g. the weights of the phoneme /z/, /uh/ and /m/) rather than the calculated 
score. is there any available function inside the phonetisuarus yet? if there 
is a way to do it, could you please explain me step-by-step on how to get those 
information ?

What version of the product are you using? On what operating system?

Linux Ubuntu version 10.x

Please provide any additional information below.

Original issue reported on code.google.com by kheangs...@gmail.com on 2 Jun 2014 at 7:47

[GoogleCodeExporter]

hi, yes it is possible in the source 'trunk' version.  this is not well-tested 
or documented yet.  i will prepare a step-by-step example of how to do so 
(using an RNNLM and normal ngram plus online interpolation) in a day or so, and 
let you know.  

Original comment by Josef.Ro...@gmail.com on 2 Jun 2014 at 11:53

[GoogleCodeExporter]

Hi,
Thank you so much for your quick response!
I'm waiting for it.
All the best

Original comment by kheangs...@gmail.com on 2 Jun 2014 at 11:56

[GoogleCodeExporter]

Original comment by Josef.Ro...@gmail.com on 4 Jun 2014 at 8:40

• Changed state: Started

[GoogleCodeExporter]

Hi, 

I think the updates from today should be sufficient to get the first step: 
scores for each individual token.  I have tested it on: Ubuntu/Linaro 
4.6.3-1ubuntu5, and OSX 10.8.5.

You should be able to verify that it works by following the short list of steps 
in the gist here:

https://gist.github.com/JosefNovak/e68f1f86a5af184ef00f

Definitely appreciated if you can verify that it works!

Assuming it does, I will now follow up with the ngram-server, 
online-interpolation and RNNLM examples.

Original comment by Josef.Ro...@gmail.com on 4 Jun 2014 at 3:12

• Changed state: Fixed

[GoogleCodeExporter]

moving back to 'started' so this stays visible a bit longer.

Original comment by Josef.Ro...@gmail.com on 5 Jun 2014 at 6:48

• Changed state: Started

[GoogleCodeExporter]

Hi,

Thank you so much for spending time on my question.
I will test it and let's you know as soon as possible.

Goodluck,

Original comment by kheangs...@gmail.com on 7 Jun 2014 at 11:38

[GoogleCodeExporter]

No problem.  Please don't hesitate to let me know if it does not work for some 
reason.  I am in the process of preparing a major new distribution, so any 
comments, complaints, requests and other feedback are very welcome.

Original comment by Josef.Ro...@gmail.com on 8 Jun 2014 at 9:32

[GoogleCodeExporter]

Hi, I have now also added a temporary tutorial for using RNNLM models, phoneme 
only models, and rescoring.  The example here gives 'real' phoneme scores.  The 
process is: find the best pronunciation with the joint g2p model, then rescore 
with an ARPA or RNNLM model trained only on phoneme sequences.

https://gist.github.com/JosefNovak/7febf69155d8fda569a2

While I have primarily been using this just for G2P stuff, the ngram-server is 
actually equally suitable to doing generic online nbest rescoring and 
interpolation of RNNLMs and various ARPA format models.

Original comment by Josef.Ro...@gmail.com on 8 Jun 2014 at 3:19

[GoogleCodeExporter]

Dear Dr. Novak,

Thank you very much for your great contribution.
I really appreciate your work and I'm sorry for my belated reply because I have 
been so busy with my lab meeting. 

After trying to test your providing shell scripts, the first file (i.e. 
"basic-usage.sh") work well except that for my PC, I have to add some 
parenthesis to control the conditions of if-statement where is located in 
"/usr/local/include/fst/interval-set.h". 

In addition, I also found two typos in the file "advance-usage.sh". 
   - Line 36: /path/to/kenlm/bin/build_binary cmu.8g.phn.arpa.gz cmu.8g.phn.arpa.bin
     it should be : ./kenlm/bin/build_binary cmu.8g.phn.arpa.gz cmu.8g.phn.arpa.bin

   - Line 45: $ python script/ngram-client.py -w BREAKFAST -n 1
     There is no $ at the beginning of the command.

Please check your second script again in order to make sure it work perfectly!
Thank again for your kind support and help !

Best regards,
Seng 

Original comment by kheangs...@gmail.com on 10 Jun 2014 at 12:39