We propose a new method of finding discriminative state-weights recursively using the generalized probabilistic descent method. This method is implemented with minor modification of the conventional parameter estimation and recognition algorithms by constraining the sum of the state-weights to the number of states in a recognition unit, and can be applied to continuous speech recognition as well as isolated word recognition. We confirm the validity of the method with phoneme-based and word-based state-weighting schemes for three kinds of recognition tasks.

We propose a modified hidden Markov model (MHMM) that incorporates nonparametric state duration and state duration-dependent observation probabilities to reflect state transitions and to have accurate temporal structures in the HMM. In addition, to cope with the problem that results from the use of insufficient amount of training data, we propose to use the modified continuous density hidden Markov model (MCDHMM) with a different number of mixtures for the probabilities of state duration-independent and state duration-dependent observation. We show that this proposed method yields improvement in recognition accuracy in comparison with the conventional CDHMM.