This paper describes the techniques for non-keyword modeling and rejection in a Korean keyword spotting system based on continuous hidden Markov models (CHMM). For non-keyword modeling, we propose a filler model which remains in the optimum path for minimum duration. It consists of several states which share an identical output probability distribution function and the number of states is the same as that of the minimum duration. Also, we investigate the performance of the proposed filler model focused on both reducing the computational complexity and improving the detection rate. In order to reject false alarms, we introduce false models similar to keyword models and employ them to determine the existence of the keyword embedded in the utterance in post processing stage. The experimental results showed that our new filler requires lower computational complexity while it provides slightly better performance than monophone fillers. Also, the false models rejected 75 false alarms out of 100 at 5.9 false alarms per keyword per hour resulting in 79.1 % detection rate. Keywords: rejection, continuous hidden Markov model, filler, false alarms, detection rate.
Bibliographic reference. Lee, Seung-Bae / Kim, Lag-Yong / Kim, Min-Seong / Lee, Jong-Seok / Kang, Shin-Wook (1995): "Minimum duration constrained non-keyword modeling and rejection for word spotting", In EUROSPEECH-1995, 2111-2114.