Recognize a phone number

Italian

Flat

User: dantavo
Date: 5/31/2014 7:46 am

Views: 9714
Rating: 9

Hi guys

I have two questions ?

1) I'm creating an acoustic model to recognize a phone number a name and a surname.

I suppose this was very simple, but now i can't recognize nothing.

I have a WER = 10% but i have bad result and my dubt is:

For this kind of application is better create an acustic model where I have a number or a name for each wav file or is correct put into the same .wav file all numbers and another with all names?

I follow the second method.

Have any suggestion?

2) If I solve the first problem, for this kind of application is better using a grammar or the ngram is the best way?

Thank you

Re: Recognize a phone number

User: nsh
Date: 5/31/2014 10:45 am

Views: 134
Rating: 6

> For this kind of application is better create an acustic model where I have a number or a name for each wav file or is correct put into the same .wav file all numbers and another with all names?

For training you need to have many audio files with utterances. Each utterance might contain a number or a name, it's not a good idea to put everything into a single file.

> I have a WER = 10% but i have bad result and my dubt is

You are welcome to share the data so you could get more detailed help on this problem

> If I solve the first problem, for this kind of application is better using a grammar or the ngram is the best way?

Grammar is ok only for few words. If you have more than 100 names in the vocabulary, it's better to use ngram model. That will allow you to account for name probability.

Re: Recognize a phone number

User: dantavo
Date: 6/1/2014 7:30 am

Views: 178
Rating: 7

my model is this:
https://www.dropbox.com/sh/5hhiaas2oehial9/AABPUycYKAe8CTcjJqFAgEeUa

but i think the problem come from the wav file because I have in one file all number and in another one all letters.

Sorry but I have also another question:

I have another model containing 15 sentences. using it with the Ngram works very well. I can say few word present in a sentence and few from another and the recognition is well. This same model used with the grammar does not recognize anything. I can't understand why, and especially if there is a way to solve because I think is a little bit strange.

Re: Recognize a phone number

User: nsh
Date: 6/1/2014 7:42 am

Views: 100
Rating: 5

Well, you need more data but I don't think it's a core issue. Using all letters is also ok probably.

How exactly are you using the model? How did you get no results? Do you use pocketsphinx_continuous or what?

Re: Recognize a phone number

User: dantavo
Date: 6/1/2014 7:58 am

Views: 133
Rating: 8

I'm using in sphinx4.

I test this model changing configuration in HelloWorldNgram, in the HelloWorld with grammar and also trying with one the file .wav used for the training in the Transcriber modifying for run in the ngram manner and with grammar.

whit ngram result are wrong always.

with grammar no result or better white space (not null)

I think the problem is the model but I hope there is a method or a setting who can give me best result.

What do you think about?

Re: Recognize a phone number

User: nsh
Date: 6/1/2014 8:07 am

Views: 107
Rating: 8

> I'm using in sphinx4.

Your model is 8khz, its' not easy to configure sphinx4 to recognize 8khz audio. Most likely you did it wrong.

You need to use latest API from 5prealpha version as described here:

http://cmusphinx.sourceforge.net/wiki/tutorialsphinx4

See the latest trascriber demo.

Also you need to add the call

recognizer.setSampleRate(8000);

then it should work.

Re: Recognize a phone number

User: dantavo
Date: 6/1/2014 8:29 am

Views: 172
Rating: 9

Thank you now I'm trying with the lastest version.

But why if i say only one word with ngram sphinx recognize me well and if I do the same thing with grammar this doesn't happen. Is possible this can depend on the decoder settings and in the search manager? What do you think if I try to change this settings is possible use WordPruningSearchManager with grammar?

Re: Recognize a phone number

User: nsh
Date: 6/1/2014 8:38 am

Views: 213
Rating: 10

> But why if i say only one word with ngram sphinx recognize me well and if I do the same thing with grammar this doesn't happen

No idea, you are asking questions without providing the data (audio and code) to reproduce your problem, it's hard to help you in that case since I can't access files on your computer.

> Is possible this can depend on the decoder settings and in the search manager?

Yes, it is possible

> What do you think if I try to change this settings is possible use WordPruningSearchManager with grammar?

It is not a trivial task that requires understanding of sphinx4 internals. It's better to use latest version and leave search managers as they are, accuracy should be good with it.

Re: Recognize a phone number

User: dantavo
Date: 6/1/2014 9:57 am

Views: 4229
Rating: 7

There is my good model not containing numbers or letters only words...you can test the things i said you before about the grammar.

The new version I don't know why but recognize only a word and this is not spelled into the wav file.
Probably, sorry surely I make a mistake :)

https://www.dropbox.com/sh/llx61cjwnkbw60y/AABhkTiVoab1M1GgFtEYghbEa

Previous • Next •


Username	Password