Re: Testing german with pocketsphinx, hardly no hits

German

Flat

User: Sebbi
Date: 1/22/2013 6:50 am

Views: 129
Rating: 10

I found a solution.

Changing the phone "aa" in the dict for "JA" to "@" makes the recognition much better. Almost all rawlogs are recognized then.

But I don't understand it. The @ is used in a totally different context in the full dictionary. So to me it looks like it's completely wrong what I did, but it works.

Any ideas?

Re: Testing german with pocketsphinx, hardly no hits

User: nsh
Date: 1/22/2013 7:05 am

Views: 372
Rating: 7

> But I don't understand it. The @ is used in a totally different context in the full dictionary. So to me it looks like it's completely wrong what I did, but it works.

You made it right.

The dictionary used for original model is not really resonable, it was created with espeak I remember. I'd recreate a phonetic dictionary and retrain the acoustic model using the available data. This way you can also build semi-continuous model which is way more reasonable for android device (it's faster to decode).

Re: Testing german with pocketsphinx, hardly no hits

User: Sebbi
Date: 2/4/2013 3:38 am

Views: 309
Rating: 8

Hi, me again.

Could you please be more specific? Sorry, I'm not a linguist I am a Software/Hardware Engineer.

What do you mean with recreating a phonetic dictionary. Do you mean I have do define Symbols for every phonem of the german language?

What do you mean with " retrain [...] using the available data"?

My biggest problem isn't the accuracy of what I have at the moment, it's more that the systems detects words that I did not say. It some kind of "force fits" to the dictionary even if the word is totally unrelated.

Re: Testing german with pocketsphinx, hardly no hits

User: nsh
Date: 2/5/2013 6:18 am

Views: 1352
Rating: 7

> Could you please be more specific? Sorry, I'm not a linguist I am a Software/Hardware Engineer.

I was under the impression software developers can do many things themselves ;) They are not bank clerks

> My biggest problem isn't the accuracy of what I have at the moment, it's more that the systems detects words that I did not say. It some kind of "force fits" to the dictionary even if the word is totally unrelated.

I replied you here, sorry, it's hard to mainain discussion across multiple sites.

https://sourceforge.net/p/cmusphinx/discussion/help/thread/48335932/

Re: Testing german with pocketsphinx, hardly no hits

User: bsnayak
Date: 12/5/2013 3:53 am

Views: 2935
Rating: 0

Moin Moin..

I am starting to work on limited vocabulary acoustic model in german speech recognition. This application would be used on an trainguard system. Although using the trivial english acoustic model, it is working fine. But i am interested to operate it with german vocabularies. It would be helpful if you would share your experience and possibly your acoustic model files and package to move on with this work.

Thanks for your help.

mit Grußen-

[ «Previous Page | 1 2 | Next Page» ]

Previous • Next •


Username	Password