The presentation concerns a preliminary investigation of a spectral pole-zero model that is fitted directly to observed log-magnitude spectra. The parameters of pole-zero models are interpretable in terms of (anti-)formant frequencies and bandwidths that may thus be tracked over time. The speech corpus has comprised connected speech tokens with prominent formant/anti-formant pairs owing to hyper-nasality in many speech frames. Results show that the direct fitting of spectral models is feasible. The quality of fit of the spectral contour by a model transfer function is comparable to the quality of fit obtained via cepstral smoothing with an effective number of cepstral coefficients equal to the number of independent model parameters.

Index Terms. Spectral pole-zero models, formant and anti-formant tracking, hyper-nasality