A formant frequency estimator for noisy speech based on correlation and cepstrum

Abstract

A formant estimation scheme combining the advantageous features of correlation and cepstral domains, which is capable of handling the adverse effect of observation noise was investigated. A residue-based least squares optimization technique based on a model-fitting approach was introduced in order to obtain formant frequencies from noisy observations. Simulations were carried out to estimate formant frequencies from synthetic and natural speech signals under noisy conditions. The human vocal-tract system was represented by a P-th order AR system with a transfer function. A formant frequency estimation scheme based on a new ramp cepstrum model was developed which is capable of efficiently handling the noisy environment. The once-repeated ACF was employed which can significantly reduce the effect of noise in the correlation domain. It was observed that the proposed method provides an accurate formant frequency estimate even at a low level of SNR.