Modeling Speech Parameter Sequences With Latent Trajectory Hidden Markov Model
Hirokazu Kameoka

This paper proposes a probabilistic generative model of a sequence of vectors called the latent trajectory hidden Markov model (HMM). While a conventionalHMMis only capable of describing piecewise stationary sequences of data vectors, the proposed model is capable of describing continuously timevarying sequences of data vectors, governed by discrete hidden states. This feature is noteworthy in that it can be used to model many kinds of time series data that are continuous in nature such as speech spectra. Given a sequence of observed data, the optimal state sequence can be decoded using the expectation-maximization (EM) algorithm. Given a set of training examples, the underlying model parameters can be trained by either the expectation-maximization algorithm or the variational inference algorithm.