A fast search technique for discriminative keyword spotting

Keyword spotting systems can be divided into two main groups: HMM-based and discriminative-based systems. Some of these systems apply a phonetic search algorithm to the sequence of recognized phones to find position of target keyword in a set of speech utterances. Thus, they need a fast and accurate phonetic search algorithm to find the position of the target keyword. In this paper, we propose a hierarchical search algorithm. In each level of hierarchy, some segments of input speech will be ignored due to their low probability of being target keyword. This tends to a smaller search space and so faster search and lower computational complexity in comparison with the Viterbi algorithm which is usually used in keyword spotting applications as a phonetic search algorithm. We apply the proposed search method to the classification part of the discriminative keyword spotter introduced in our previous works. The experimental results indicate that the hierarchical search algorithm is 100 times faster than the modified Viterbi algorithm when it is used in the discriminative keyword spotting system. On the other hand, FOM of the discriminative keyword spotting system using the proposed hierarchical search algorithms degraded about 2 % in comparison to the case that this system uses a modified version of Viterbi algorithm.