I have just started playing with voice recognition using the mic array through node.js (MALOS) and everything is working fine, except for the high “ambient” noise in all recordings.
Since I am using rec in order to stream the voice commands to wit.ai (after the hotword trigger, which is handled fine by snowboy), the high ambient noise forces me to use quite high (~10%) silence margins, otherwise silence is never detected and the recording never stops. Although the high margins work, some voice data is almost always cropped from the beginning and the end of the voice commands, which makes recognition erratic.
I have read that if MALOS is disabled the noise gets better, but since I am using MALOS to talk to the unit, this is not an option for me.
Any ideas on how to get the ambient noise to the normal (3%) levels?
Please note that all my test recordings are made in a perfectly silent room, so we are not talking about real ambient noise.