Recognition Accuracy under noisy environment

This topic has 2 voices, contains 2 replies, and was last updated by darbienapp 278 days ago.

Viewing 3 posts - 1 through 3 (of 3 total)

Author	Posts
Author	Posts
August 14, 2011 at 4:05 am #7476
darbienapp	Is there any parameter to tune the recognition accuracy of PocketSphinx under noisy background environment? 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. However, it seems that when I tested it under a more noisy background environment (silent input level around -75dB), the recognition accuracy is really bad I can not get even 1 out of 20 tries correct anymore. Most results came in with empty string or “(null)”. One note : I tested this noisy environment while inside a moving car, with windows closed. I suppose just the ambient noise is enough to make the background input level up to -75 dB. Also it seems to correctly detected speech start/end correctly under the noisy background. I started the sample project after the car is moving and I think it must have correctly calibrated the silent level.
August 14, 2011 at 8:38 am #7477
Halle	Hi, It is normal for you to get a very large number of of null and empty results when it is noisy — that means that the noise dictionary is working to recognize noises as not being words, which you can then disregard in the callback. The question is whether it recognizes you when you do speak. I think that your language model is actually especially challenging because it consists of two single-syllable words in an environment in which all of the noise occurrences will also be suspected of being single-syllable words. In a noisy environment, “yes” pretty much just sounds like a sibilant “ess” and “no” sounds like “o”, either of which could be an environmental sound. There isn’t a special setting for noisy environments (just the ongoing calibration that occurs) but I will say that in my own app AllEars, which usually has an ARPA model of around 150 complex words, I see pretty decent recognition with background noise considering what a challenging circumstance it is for recognition. You have the opposite problem of this poster so maybe this discussion will give you some ideas: http://www.politepix.com/forums/topic/adjust-sensitivity-for-commands/#post-6837 If you want to bother, you can experiment with decreasing the probability of the noise dictionary when there is a certain amount of background noise.
August 14, 2011 at 3:36 pm #7478
darbienapp	Halle, Thanks for the great suggestions, I’ll definitely give this probability a try! You also brought up a very good point of the single syllable words, I think it’s worth looking into as well.

Viewing 3 posts - 1 through 3 (of 3 total)

You must be logged in to reply to this topic.