A detection approach to search-space reduction for HMM state alignment in speaker verification

To support speaker verification (SV) in portable devices and in telephone servers with millions of users, a fast algorithm for hidden Markov model (HMM) alignment is necessary. Currently, the most popular algorithm is the Viterbi (1967) algorithm with beam search to reduce search-space; however, it is difficult to determine a suitable beam width beforehand. A small beam width may miss the optimal path while a large one may slow down the alignment. To address the problem, we propose a nonheuristic approach to reduce the search-space. Following the definition of the left-to-right HMM, we first detect the possible change-points between HMM states in a forward-and-backward scheme, then use the change-points to enclose a subspace for searching. The Viterbi algorithm or any other search algorithm can then be applied to the subspace to find the optimal state alignment. Compared to a full-search algorithm, the proposed algorithm is about four times faster, while the accuracy is still slightly better in an SV task; compared to the beam search algorithm, the proposed algorithm can provide better accuracy with even lower complexity. In short, for an HMM with S states, the computational complexity can be reduced up to a factor of S/3 with slightly better accuracy than in a full-search approach. This paper also discusses how to extend the change-point detection approach to large-vocabulary continuous speech recognition.