Sphinx on WindowPhone

[Loc Le Duc]

Hello everyone,

I'm building a recognizer in my WindowPhone. I'm flowing the project of Toine de Boer in this link:

https://github.com/cmusphinx/pocketsphinx-wp-demo
It works very well and it's so amazing.
Now, I want to create "Adapting the default acoustic model" project in Winphone.
So my question is:
1) Can I use SpinxTrain in Winphone to adapt the default acoustic model?
2) If not, Have any idea to improve the accuracy when i use the recognizer in the noise environment or an independent user can use my recognizer?

Thanks and Regards,
DucLoc.

[Nickolay V. Shmyrev]

We do not recommend to run sphinxtrain on the phone, adaptation is very easy to make wrong and basically this feature is not supported yet. We might implement MLLR adaptation one day in pocketsphinx itself, for now it requires some more attention.

Recognizer should work pretty well in noisy environment, it has noise reduction features included, however, we might recommend you to use noise-cancelling headset in case you want to operate in significant level of noise.

[Toine db]

Hi DucLoc,

Great to hear there are more people working on WP and Sphinx :-)

What would be the goal of adapting the current acoustic model?

Kind Regards,
Toine

[Loc Le Duc]

Hi Nickolay V. Shmyrev

Sorry for my late answer, To use noise reduction features in sphinx, I need to add

-remove_noise yes

To file feat.params.
Right?

Thanks and Regards,
DucLoc.

[Nickolay V. Shmyrev]

Noise is removed by default, you do not need to do anything special.

[Loc Le Duc]

Thank you vey much.

Thanks and Regards,
DucLoc.

[Loc Le Duc]

Hi Toine db.

Sorry for my late answer.
First, I want to say thank you, because your projects on WP is so amazing.

I do my thesis at the university. I create a voice recognizer application on my phone, and use it to control the house. When i test with my voice, the accuracy is ok, it's approximate 90-95%. But when my friends and my teacher use it, the accuracy is just 70-80%.
I want other people can use my system without waste to much time to train the acoustic model. So I think the adapting model is suitable.

But as Nickolay V. Shmyrev say the adapt features in Sphinxtrain cannot use on WP.
Do you have any suggestion to improve the accuracy when other people use the recognizer on phone?

Thanks and Regards,
DucLoc.

[Toine db]

Hi DucLoc,

Good to hear that you made fun stuff with some part of the example!!!

If you have adjustments or imporvements to the Github source, please commit it with a pull request.... or send it to me. I would love to work together if we can make the solution better that way.

I do not have a advise to improve the recognition, I myself am working on a kind of same issue.

[Loc Le Duc]

Hi Toine db.

Thank you vey much. i'll try something to improve the accuracy for other people. If it work, i'll tell you.

Thanks and Best Regards,
DucLoc.

[Loc Le Duc]

Hi Toine db.

When I run Spinx in Ubuntu, the terminal output will have the new value of cmminit. The value of cmninit is used to array string results.
Now in the windows phone, I think I need to change

"-logfn", logPath,

in the SpeechRecognizer.cpp to create the log file. But I donot know where is it location in my phone, and how can i read it?
Can you help me?

Thanks and Best Regards,
DucLoc.

[Nickolay V. Shmyrev]

But I donot know where is it location in my phone, and how can i read it?

You can use relative location, something like "error.log", then it will be created in your app folder on the phone.

But I donot know where is it location in my phone, and how can i read it?

You can access the filesystem with a special tool:

http://wptools.codeplex.com

You can also use

https://msdn.microsoft.com/en-us/library/windows/apps/hh286408(v=vs.105).aspx

but it's command line only.

[Toine db]

Hi DocLuc,
Sorry for the late reply on this, I do not get any notification when new messages come on this forum.

Do you still need help?