Multiple Dictionaries to match

[looking_glass]

Hello,

Would open ears recognize names if they are listed in the dictionary? And can I have multiple dictionaries to be matched against for every spoken word? I am thinking of having each spoken word match against either a most frequently spoken word or a name dictionary.

Thanks.
PJ

[Halle Winkler]

Welcome,

I actually originally developed OpenEars as part of an app which recognized spoken names from the contact list, and my experience was pretty good. However, the names on my list were either of English origin or German, where the phonemes could be estimated pretty well. Basically, the farther away the names in the contact list are from English words/phonology, the worse OEPocketsphinxController will do with them. However, you could also add names with phonetic transcription to the lookup list if you want to.

Can you elaborate more on the multiple dictionaries question? I’m not sure yet whether you are asking about having more than one language model and switching between them over time (very easy with OpenEars) or doing two separate recognition passes on the same speech input using two different language models (not so easy).

[looking_glass]

For the multiple dictionaries, I was thinking more along the lines of multiple passes on the same speech using two different models.

That way I could theoretically scan for a set of known words in one pass (hopefully using one thread of execution) and use another thread for scanning for names.

Thanks,
-Piyush

[Halle Winkler]

I can’t think of an easy way to do this, but it might be possible to have SaveThatWave output a WAV of the captured speech and then immediately submit it to this method:

- (void) runRecognitionOnWavFileAtPath:(NSString *)wavPath usingLanguageModelAtPath:(NSString *)languageModelPath dictionaryAtPath:(NSString *)dictionaryPath acousticModelAtPath:(NSString *)acousticModelPath languageModelIsJSGF:(BOOL)languageModelIsJSGF;

I’ve never tried this but it is possible that it may work, no promises. You can’t run the same input on two models using the standard recognition methods.

[Author]

Posts