Adaptive estimation of some parameters of speech excitation

The accuracy of the vocal tract parameters (e.g. area function) estimated by LPC-techniques crucially depends on the interaction model between the excitated speech-waves and their propagation in the vocal tract. As LPC decouples excitation and propagation there exist critical speech segments - the open glottis and plosive segments - for which decoupling is not valid. A new algorithm detects these critical speech segments and removes them from speech analysis. With the aid of a dynamic model of articulatory motion during critical segments the vocal tract parameters are interpolated from neighbouring segments. The estimation of the excitation and vocal tract parameters is done in an adaptive loop using a Kalman filter.