Switching acausal filters for speech modeling

Abstract

This paper shows a unified model of dynamical systems in speech processing that includes speech recognition and pitch modeling. For this purpose, we propose the use of switching acausal filters (SAFs), which exchange multiple acausal filters. These filters are defined by identical linear dynamical systems that exchange the roles of observation value and system input. This paper describes the formulation of recognition, training, and feature generation methods for SAFs, which can be applied to several previously proposed speech models. As an example, we show that an HMM with dynamic features and our F0 control method can be modeled by the proposed formulation. An HMM synthesis method can also be modeled using the formulations. From these results, we demonstrate the unification capability of SAFs.