Robust LP analysis using glottal source HMM with application to high-pitched and noise corrupted speech

This paper presents a robust feature extraction method effective to speech signal with high fundamental frequency and/or corrupted by additive white noise. The method represents the glottal source wave using HMM in order to model the nonstationary properties. The nodes of HMM are concatenated in a ring state to represent the periodicity of voiced sounds. The method can accurately extract glottal source wave and vocal tract characteristics from speech signals even in high fundamental frequency as ranging up to 750Hz. From identification theory, estimation of vocal tract characteristics from speech corrupted by additive noise requires glottal source wave that can not be observed directly, so that it needs to be estimated. Therefore, estimation accuracy of vocal tract characteristics highly depends on the estimation accuracy of glottal source wave. We apply the glottal source HMM to extracting the glottal source wave from corrupted speech, and confirmed the feasibility of the method.