Formant estimation system based on weighted least-squares lattice filters

A formant estimation system using the adaptive weighted least-squares lattice (WLSL) algorithm and novel formant labelling techniques is presented. In the WLSL, the likelihood variable, which can be considered as a statistical measure of the non-Gaussian component of the speech signal, is used to deweight time intervals in the speech waveform which correspond to glottal excitation. A short analysis window coupled with optimal frame position placement, determined by the local minima of both the likelihood variable and the residual is used to emulate glottis-closure, closed-phase analysis. The algorithm, which can also be considered as a special form of robust linear prediction analysis, offers an improved performance (i.e. a less biased formant frequency) in comparison to the frame-based linear prediction analysis. After formant candidates have been frame extracted from the spectral estimates for each of the waveform, a clustering procedure is first used to produce line segments of possible formants. A rule-based labelling mechanism is then applied to these segments to provide final formant trace estimates. Experimental results show the labelling algorithm proposed offers improved formant labelling accuracy.