Marius C Silaghi
This paper addresses the problem of detecting keywords in unconstrained speech. The proposed algorithms search for the speech segment maximizing the average observation probability along the most likely path in the hypothesized keyword model. As known, this approach (sometimes referred to as sliding model method) requires a relaxation of the begin/endpoints of the Viterbi matching, as well as a time normalization of the resulting score. This makes solutions complex (i.e., LN2/2 basic operations for keyword HMM models with L states and utterances with N frames). We present here two alternative (quite simple and efficient) solutions to this problem. a) First we provide a method that finds the optimal segmentation according to the criteria of maximizing the average observation probability. It uses Dynamic Programming as a step, but does not require scoring for all possible begin/endpoints. While the worst case remains O(LN2), this technique converged in at most 3(L+2)N basic operations in each experiment for two very different applications. b) The second proposed algorithm does not provide a segmentation but can be used for the decision problem of whether the utterance should be classified as containing the keyword or not (provided a predefined threshold on the acceptable average observation probability). This allows the algorithm to be even faster, with fix cost of (L+2)N.
Content Area: 14. Natural Language Processing & Speech Recognition
Subjects: 18. Speech Understanding; 15.7 Search
Submitted: May 10, 2005