Hi,

It is not entirely clear to me what you mean... especially when you say "the acoustic models in the demos have a very small vocabulary". An acoustic model and a vocabulary are two entirely different things. Do you mean that the demos have a very small grammar? A grammar defines the possible commands of an application (a command and control application).

If you want to use a bigger grammar (more possible commands) then you should study the structure of the grammar that comes with the demo and make a larger grammar. That question has been asked before in the VoxForge forums, so you should be able to find the answer to that question by searching the site.

Also, it is not entirely clear which acoustic model you are trying to use. There are several acoustic models that can be downloaded from VoxForge and some are meant to be used with Julius and some are meant to be used with Sphinx. You cannot use an acoustic model meant to be used in Julius directly in Sphinx.

Sorry to answer your questions with a couple of questions, but right now it's not entirely clear what you mean (at least not to me), which makes it harder to help you.

Robin

--- (Edited on 12/29/2009 1:53 pm [GMT-0600] by Robin) ---

Hi Robin, thank you for your answer.

Sorry, you are right, I didn't explain myself well.

When I said that they have a small vocabulary, I meant they are trained with a little amount of words. For example, the dictionary of acoustic model called TIDIGITS_8gau_13dCep_16k_40mel_130Hz_6800Hz.jar only has numbers from 0 to 9, so I suppose that with this acoustic model, the only thing you can recognize are numbers.

The acoustic model I was trying to use was the one in VoxForge for Sphinx (there are only two in english, I downloaded the most recent). The files in it are not the same as the ones in the acoustic models of the demos, so I didn't know exactly how to configure my sphinx application to use it.

I've been using the VoxForge's acoustic models for Julius successfully (in Julius, of course) and I had no problems to configure Julius to use it. Here, in VoxForge, there are very good tutorials for Julius but very little for Sphinx.


Anyway, don't worry. I've been looking more closely at the demos and I found that it is also used another acoustic model trained with a great amount of words, so I will use that one. That acoustic model is WSJ_8gau_13dCep_16k_40mel_130Hz_6800Hz.jar


Thank you for your help

--- (Edited on 12/29/2009 6:16 pm [GMT-0600] by Visitor) ---

Glad to hear that you are making progress.

If the type of speech you want to recognise is more like the type of speech that was used to train this other model (WSJ_8gau_13dCep_16k_40mel_130Hz_6800Hz.jar) then you are completely right and you should switch to that model.

Of course it is not impossible to recognise words that are not present in the speech used to train the acoustic model, but if the vocabulary in the training speech was extremely limited, then the acoustic model will indeed not be very suited for a large vocabulary application.

Good luck!

--- (Edited on 12/30/2009 2:49 pm [GMT-0600] by Robin) ---

> The acoustic models of the Sphinx4 demos are JAR files and the files

>in that JAR are not the same as the ones in the TAR of de acoustic

>model from VoxForge.

See nsh's post here: Sphinx3 model to Sphinx4 model

--- (Edited on 1/5/2010 2:13 pm [GMT-0500] by kmaclean) ---

SpeechLion is a dialog manager based on Sphinx4 which uses the WSJ acoustic model and a bunch of JSGF grammars.  It may be helpful as another example of configuring Sphinx4 (which I have found to be difficult).

http://freshmeat.net/projects/speechlion

--- (Edited on 3/7/2010 1:58 pm [GMT-0600] by rbrewer) ---

The corpus is recorded by myself ,  there is no clear syntax, so I don't know how to write a grammar file. The wdnet is produced by the command ：“HBuild wlist  wdnet”，but when the last step recognition，HVite alway error

HVite -H .\hmms\hmm7\macros -H .\hmms\hmm7\hmmdefs -S test.scp -l * -i .\results\recout_step7.mlf -w wdnet -p 0.0 -s 5.0 .\dict\dict1.new .\lists\monophones1

ERROR[+6511] SaveToMasterfile: Unable to create MLF file .\results\recout_step7.mlf

ERROR[+6511] LSave: Unable to create label file */S0001.rec

ERROR[+6511] LSave: Unable to create label file */S0002.rec

ERROR[+6511] LSave: Unable to create label file */S0003.rec

ERROR[+6511] LSave: Unable to create label file */S0004.rec

ERROR[+6511] LSave: Unable to create label file */S0005.rec

ERROR[+6511] LSave: Unable to create label file */S0006.rec

ERROR[+6511] LSave: Unable to create label file */S0007.rec

ERROR[+6511] LSave: Unable to create label file */S0008.rec

ERROR[+6511] LSave: Unable to create label file */S0009.rec

ERROR[+6511] LSave: Unable to create label file */S00010.rec

Thank you!

--- (Edited on 3/12/2013 7:52 am [GMT-0500] by ) ---

mkdir results first 

--- (Edited on 9/22/2015 9:59 pm [GMT-0500] by Visitor) ---

The corpus is recorded by myself ,  there is no clear syntax, so I don't know how to write a grammar file. The wdnet is produced by the command ：“HBuild wlist  wdnet”，but when the last step recognition，HVite alway error

HVite -H .\hmms\hmm7\macros -H .\hmms\hmm7\hmmdefs -S test.scp -l * -i .\results\recout_step7.mlf -w wdnet -p 0.0 -s 5.0 .\dict\dict1.new .\lists\monophones1

ERROR[+6511] SaveToMasterfile: Unable to create MLF file .\results\recout_step7.mlf

ERROR[+6511] LSave: Unable to create label file */S0001.rec

ERROR[+6511] LSave: Unable to create label file */S0002.rec

ERROR[+6511] LSave: Unable to create label file */S0003.rec

ERROR[+6511] LSave: Unable to create label file */S0004.rec

ERROR[+6511] LSave: Unable to create label file */S0005.rec

ERROR[+6511] LSave: Unable to create label file */S0006.rec

ERROR[+6511] LSave: Unable to create label file */S0007.rec

ERROR[+6511] LSave: Unable to create label file */S0008.rec

ERROR[+6511] LSave: Unable to create label file */S0009.rec

ERROR[+6511] LSave: Unable to create label file */S00010.rec

Thank you!

--- (Edited on 3/12/2013 7:53 am [GMT-0500] by ) ---