High resolution formant extraction from linear-prediction phase spectra

Abstract

A new algorithm is proposed to extract formant information from linear-prediction (LP) spectra. The problems of resolving two closely spaced formants and accurately estimating the bandwidth of a formant are particularly addressed. The proposed method detects resonant poles from the peaks of the second derivative of the log-magnitude spectrum or from the second derivative of the group delay of the LP-coefficient polynomial. It then goes on to estimate frequencies and bandwidths of formants by observing the phase slope of the z-transform around the poles. Consequently, no minimum frequency separation is required in order to distinguish between two formants. The accuracy and utility of the proposed method, in reliably resolving closely spaced formants, is illustrated with several examples. Results on the performance of the algorithm on synthetic and natural speech are also included.