Improvement of Telephone Keyword Spotting Performance Using Linear Programming-Based Score Normalization

Conventional word spotting systems determine hypothesized keywords and their confidence score using a speech recognizer. Acceptance or rejection of these keywords is intended based on comparison of their scores with a specific threshold. It has been proved that confidence score prepared by recognizer is highly dependent on sub-word structure of each keyword. So comparing assigned scores to keywords without considering their sub-word units could causes degradation in overall performance. In this paper a novel method for confidence score normalization is proposed which is based on sub-word units of each keyword and linear programming algorithm. In proposed method, a keyword-dependent correction term is added to the score of the keyword to maximize separation of confidence score histograms of true and false occurrences. Our results show a 2% improvement in FOM compared to baseline system. Also, choosing an appropriate feature vector has been discussed in this paper.