In this work a method for splitting continuous mixture density hidden Markov models (HMM) is presented.The approach combines a model evaluation measure based on the Maximum Mutual Information (MMI) criterion with subsequent standard Maximum Likelihood (ML) training of the HMM parameters.Experiments were performed on the SieTill corpus for telephone line recorded German continuous digit strings.The proposed splitting approach performed better than discriminative training with conventional splitting and as good as discriminative training after the new splitting approach.

A speaker adaptation algorithm using formant frequencies is proposed. The formants extracted from the cepstral means in the reference codebook are iteratively shifted toward the formants of a test speaker. The number of cepstral means selected at each iteration decreases as the iteration increases. The decision of the number of selected cepstral means and a formant based distance measure are formulated. The proposed algorithm was implemented in two schemes and evaluated by speaker-independent, male speaker dependent, and female speaker-dependent recognition experiments. A combined scheme with the Bayesian adaptation obtained 9.7% enhancement for the average recognition accuracy in speaker-independent experiments and 52.6% in speaker-dependent recognition experiments.

Hidden Markov models (HMMs) with explicit state duration density (HMM/SD) can represent the time-varying characteristics of speech signals more accurately. However, such an advantage is reduced in relatively smooth state duration densities or long bounded duration. To solve this problem, the authors propose HMMs with global path constraint (HMM/GPC) where the transition between states occur only within prescribed time slots. HMM/GPC explicitly limits state durations and accurately describes the temporal structure of speech simply and efficiently. HMMs formed by combining HMM/GPC with HMM/SD are also presented (HMM/SD+GPC) and performances are compared. HMM/GPC can be implemented with slight modifications to the conventional Viterbi algorithm. HMM/GPC and HMM/SD+GPC not only show superior performance than the conventional HMM and HMM/SD but also require much less computation.>