Speech Recognition Engines

Flat
loss of precision in speech recognition in a server mode of Julius
User: calel
Date: 7/23/2012 8:11 am
Views: 2621
Rating: 7

Hello to all,

My name is Charles Julius and I'm using for a university project ..

 

I'm Italian and I live in Benevento, Campania.

I developed my thesis on the possibility to use Julius as "support" for deaf students in university courses.

I managed to get a good recognition (per word), 80% with a training of 6 reps for word, using Jiulius in standard mode.

But now I need to use in server mode Julius, to make accessible to all students on voice recognition service.

I created a system of encapsulation for the server and allows jcontrol, along with a variety of tools, rounding to preferred phrases, which makes the recognition of sentences (less than 10 words) to 92% accurate.

It works well with Julius executed normally, just because the good riconosciemento is greater than 70%, however, using Julius in server mode, the good recognition drastically worsens, reduced to less than 40%.

While using the same acoustic model, the same microphone and the same configuration parameters, I get worse performance with Julius server (module).

I performed several searches for it, but I have not found other people who have my problem.

Can you help?

 

It 's maybe a problem jcontrol .. Or, perhaps the client adds noise?

Or maybe in server mode (there is little documentation about it), several computations are performed by the engine?

 

Thanks C.

P.S.: You can also contact me here  https://www.facebook.com/carlo.galliano

 

--- (Edited on 7/23/2012 8:11 am [GMT-0500] by ) ---

PreviousNext