Home › Forums › OpenEars › Clarificatios Need about Openears

Tagged: 

This topic has 2 voices, contains 4 replies, and was last updated by  ramshad 128 days ago.

Viewing 5 posts - 1 through 5 (of 5 total)
Author Posts
Author Posts
January 10, 2012 at 9:48 pm #8378

ramshad

I would like to make a blog post to thank Halle for the fabulous Openears Library and clarify more about it.

Details of Openears
**************************
1.I have used 0.911 version of Openears.

2.I have used the English Voxforge acoustic model with Openears(from https://sourceforge.net/projects/cmusphinx/files/Acoustic%20and%20Language%20Models/)

3.I am using Dynamic Language model creator class.

Clarifications need
************************
1.My clients are from California.So they told me a concern about the accuracy.According to my research, i thought the default acoustic model of poketsphinx may not support California accent well.So i have planned to use the English Voxforge acoustic model instead of default one. However English Voxforge acoustic model slow in performance because of the large binary size.It takes only 3-4 second delay to process after detection is done.
Is my thought is correct??

2.I have read Mr.darbienapp comment from this blog as “I created an ARPA model with just two choices (YES|NO) and tested it on my iPhone device. My finding is that the accuracy of the recognition when in quiet environment (silent input level around -105dB) is really excellent! I’m getting accuracy of about 99% as the recognition is rarely wrong.”

what is the mean of ARPA language model?? is it a common name for .languagemodel and .dic or generated using Sphinx Knowledge Base Tool???

I think JSGF grammar will not support dynamic language model creation.

3.Using an ARPA language model,you suggested as “If you have phrases or sentences that you would like to have recognized in their entirety, you can put them all on a line”

is it possible to do this for dynamic language model creation???

4.There are 2 screens in my application.one of the screen need to detect 10-15 words.Other one need to detect only 1-5 words.

So is there any settings to set the listening time dynamically(for screen one:5 seconds,for screen two:20 seconds )??

5.Is there any default method available in openears to identify the application is resumed?

Thanks.

How can i identify input level(eg:105dB)??
January 11, 2012 at 4:39 pm #8380

Halle

Hi Ramshad,

Welcome. I encourage you to use the current version of OpenEars (.913) and to read all of the documentation, since most of your questions are answered there.

Briefly:

1. The default acoustic model should do fine with a California accent. It works best with US accents. I would expect a >15-word model to have high accuracy with the default hmm so I would expect that your issue is due to something else, such as incorrect configuration or something else about your testbed that doesn’t replicate how your clients are using your app. It could also be an issue with the actual model, for instance maybe you are trying to detect isolated syllables, or numbers, and not complete words or phrases. But you should know firsthand how the accuracy is, it shouldn’t be necessary to get the info about that from your clients indirectly. In a pinch, you can get some good speech recordings of some people with the accent you need to check and play it back out of a very good speaker with the phone facing it to see what is happening.
2. ARPA models in the context of OpenEars are explained in the docs on this page: http://www.politepix.com/openears/yourapp#addingmodel and a spec description can be found here: http://www.speech.sri.com/projects/srilm/manpages/ngram-format.5.html. You are correct that in OpenEars, only dynamic model creation for ARPA models is supported.
3. Yes, you can put a sentence or space-delimited series of words into LanguageModelGenerator. This will give them a somewhat higher probability of being detected as a sentence or phrase.
4. When utterances are submitted to the recognizer is 100% controlled by the user. When they pause for the length of kSecondsOfSilenceToDetect in OpenEarsConfig.h, they cause the utterance to be submitted to the recognizer. It isn’t something which occurs after an arbitrary period of time.
5. Why don’t you want to use the normal Objective-C methods of detecting when the application is resumed?
6. Input level code is demonstrated in the sample app.
January 11, 2012 at 7:22 pm #8381

ramshad

Hi Halle,
I just wanted to thank you for taking the time to reply to my problem in such a prompt manner. You really opened my eyes :) .Now I have another clarification about the point 3.

I have put a sentence or space-delimited series of words into LanguageModelGenerator.Let’s say “IN FRONT OF”. However after the Language model creation, the resultant Openears log shows it as separate 3 words.like

IN
FRONT
OF

i didn’t see the word “IN FRONT OF’ in console log!!.
Let me know is there any issues ? is it a correct log?

Could you please Provide me few sample free/paid Iphone/Ipad application names which used Openears.
January 11, 2012 at 7:40 pm #8382

Halle

The log that is printed out at the start of initializing pocketsphinx isn’t a log of the ARPA language model, it’s just a list of words that the app knows. You could only verify that the words have a higher probability by looking at the probabilities in the ARPA model which is created that ends with the suffix .languagemodel. This would require getting acquainted with the meaning of the information in the ARPA model by doing some self-directed research on the format.

Could you please Provide me few sample free/paid Iphone/Ipad application names which used Openears.

I think Google is the best tool for this.
January 11, 2012 at 8:51 pm #8383

ramshad

Hi Halle,

Thanks for the support. :)
Viewing 5 posts - 1 through 5 (of 5 total)

You must be logged in to reply to this topic.