Poor accuracy for spanish language

[oscarjiv91]

Hello again. I’m trying to recognize when an specific word is said. In this case, a name (Carlos, David, etc). I’m having good results with english, but in spanish I have a lot of false positives. I’m using Rejecto, and tried the vadThrestold between 2.0 and 4.0. Is there a way to improve the accuracy? Thanks

[Halle Winkler]

Hello,

The Spanish vadThreshold has to be even higher to aggressively exclude incidental noise – I think it needs to be something like 4.1-4.4.

[oscarjiv91]

Thanks for your quick reply.

The thing is that I need to recognize the words from like 2 meters away, and if I set the vadThreshold with a very high value, it does not recognize the word I want. Adapting the acoustic model would improve the accuracy?

[oscarjiv91]

or training the acoustic model?

[Halle Winkler]

There should be a sensitivity setting which is low enough to work from that distance and high enough to exclude incidental noise.

[oscarjiv91]

This “sensitivity setting” is the vadThreshold? or another variable used by openears?

[Halle Winkler]

vadThreshold, correct. This issue:

I have a lot of false positives

isn’t an accuracy issue but a sensitivity issue. It doesn’t need to be fixed by improving accuracy but by increasing vadThreshold (and probably using Rejecto if the vocabulary is small) to the point that incidental noise isn’t prompting non-null hypotheses to be returned.

[oscarjiv91]

Ok thanks Halle! Will test it more with higher values. I’m using Rejecto btw :)

[Halle Winkler]

Super, I would just operate from the expectation that there is going to be a combination of vadThreshold and Rejecto that will give the right results, and experiment from there. Keep in mind that Rejecto can also have a larger or smaller weighting which will increase and decrease its effects. Rejecto is more concerned with rejecting speech at the same loudness as the desired speech but that is not speech represented in the app vocabulary, and vadThreshold is somewhat more concerned with rejecting incidental sounds and non-user speech, but it’s a known issue that Spanish needs higher baseline vadThreshold settings than English.

[oscarjiv91]

Yes, I’m always setting the Rejecto weight to 2.0. Setting the vadThreshold to 3.0 in spanish, I get more false positives than expected when somebody is talking (especially when there are words with similar vowels), even if there is no background noises. I guess this is an accuracy issue, right?

[Halle Winkler]

With the Spanish model, vadThreshold has to be set much higher (as above, something around 4.0 to 4.3). You should turn Rejecto off and find the ideal setting for vadThreshold first, and then use Rejecto to exclude out-of-vocabulary speech that remains. This isn’t an accuracy issue, since there is no way to be accurate with utterances which aren’t in the vocabulary – they are simply unknown. This is a combination of oversensitivity and out-of-vocabulary recognition, for which vadThreshold and Rejecto are respectively used.

[oscarjiv91]

I appreciate your help Halle! This was very helpful. Thanks

[Halle Winkler]

You’re welcome!

[Author]

Posts