Improved keyword spotting based on keyword/garbage models

We propose two simple methods to improve the performance of a keyword spotting system. In our application, the users are allowed to change the keywords anytime if they want. Thus we focused on phone-based GMM-HMM models since they do not require keyword-specific training data. However, the GMM-HMM based models usually have very high false alarm rate, i.e., a keyword is not present but the system gives a positive decision. We found that we can utilize the uncertainty of the system when a non-keyword is presented. Two simple methods are proposed to incorporate the uncertainty into the confidence measure. Our experiments show that these two methods can substantially reduce the false alarm rate from 75.05% to 5.71%. Meanwhile, the false reject rate increases from 1.04% to 5.71%.