Competing Set Based Verification Method in Speech Keyword Recognition

Traditional likelihood ratio test (LRT) keyword verification methods require high computational complexity in calculation of alternative hypothesis score using cohort models. In this paper, we focus on reducing the computational complexity and extend the research to competing set based verification method, which only requires computing over a smaller number of competing models. Bhattacharyya distance is applied to derive the competing set for each sub-syllable model. Experiments on a Mandarin keyword spotting system shows, with appropriate selected competing set, the proposed method can significantly reduce the computational complexity and provides a comparable performance to the original method.