HMM parameter optimization using tabu search [speech recognition]

Hidden Markov model (HMM) is regularly trained via mathematic functions optimized by gradient-based methods such as Baum-Welch (BW) algorithm. However, optimization from gradient-based methods usually yields only a local optimum. In this paper, tabu search (TS), an artificial intelligence (AI) technique able to step back from a local optimum and search for other optima, is employed to attack this difficulty. This paper aims to utilize HMM with TS for speaker-independent (SI) continuous speech recognition. The experiment starts from a single speaker experiment in order to design and adjust the algorithm. Then, multi-Gaussian context-dependent (CD) model is applied for SI system. The results show the merit of this new algorithm comparing with the original BW.