A step in the realization of a speech recognition system based on gestural phonology and landmarks.

Abstract

Current state-of-the-art speech recognition systems poorly perform in casual speech as they are not robust against acoustic variations. Variability is an intrinsic problem of phone-based recognition systems as phone-units may be distinctive in the cognitive domain, but their physical characteristics can vary considerably due to coarticulation. Articulatory phonology proposes that articulatory constriction gesture is an invariant action unit and speech can be decomposed into a constellation of gestures. Speech recognition studies have recently employed this framework in the sense that speech acoustic variations can be accounted for by gestural coarticulation and reduction. The VT time-functions are time-varying physical realizations of gestural constellations at distinct vocal tract sites for a given utterance. This study aims to predict VT time-functions from acoustic signals as a component model in a complete gesture-based speech recognition system. We will compare artificial neural-network (ANN) based nonlinear regression models against hierarchical support vector regression (SVR) to obtain VT time-functions from the speech signal. A single ANN can generate all 8 VT time functions, whereas SVRs generate one output at a time, necessitating use of 8 SVRs. We also explore usage of spectral parameters such as MFCC and PLPs against knowledge-based acoustic parameters (APs).