Subspace projection based analysis of speech under stressed condition

Abstract

This paper proposes a novel subspace projection based approach for analysis of stressed speech signal. The projection of stressed speech vectors onto the neutral speech subspace can separate speech specific information from stress information. Orthogonality between speech and stress is assumed to separate these two information. The orthogonal relation between speech and stress subspaces is verified using speech and stress recognition techniques with a stressed speech database consisting of four stress conditions namely, neutral, angry, sad and Lombard from 30 words vocabulary. Studies show that the speech and stress specific information are present in their respective subspaces which proves orthogonality between these two subspaces.