Hi, all!

I'm trying to implement the algorithm KWS on the basis of phonemes. Phoneme recognition accuracy is about 60% (training acoustic models produced by means of HTK). Words required for the search, entered from the keyboard and converted into a set of phonemes. The main problem is the large number of false positives. Is there an algorithm for solving this problem (without the use of linguistic models)?

--- (Edited on 8/19/2014 8:05 am [GMT-0500] by Winchester) ---