Robust LP analysis using glottal source HMM with application to high-pitched and noise corrupted speech
Akira Sasou, Kazuyo Tanaka · 2001
Abstract This paper presents a robust feature extraction method effec-tive to speech signal with high fundamental frequency and/orcorrupted by additive white noise. The method represents theglottal source wave using HMM in order to model the non-stationary properties. The nodes of HMM are concatenated in aring state to represent the periodicity of voiced sounds. Themethod can accurately extract glottal source wave and vocaltract characteristicsfrom speech signals evenin highfundamen-tal frequency as ranging up to 750Hz. From identification the-ory, estimation of vocal tract characteristics from speech cor-rupted by additive noise requires glottal source wave that cannot be observed directly, so that it needsto be estimated. There-fore, estimation accuracy of vocal tract characteristics highlydepends on the estimation accuracy of glottal source wave. Weapply the glottal source HMM to extracting the glottal sourcewave from corrupted speech, and confirmed the feasibility ofthe method. 1. Introduction