Abstract

Construction of dynamic (delta) features of speech, which has been in the past confined to only the preprocessing domain in the hidden Markov modeling (HMM) framework, is generalized and formulated as an integrated speech modeling problem. This generalization allows us to utilize state-dependent weights to transform static speech features into dynamic ones. In this letter, we describe a rigorous theoretical homework that naturally incorporates the generalized dynamic-parameter technique and present a maximum-likelihood-based algorithm for integrated optimization of the conventional HMM parameters and of the time-varying weighting functions that define the dynamic features of speech.

‚Äč