Formant contour estimation in noise: comparison of a new zero-crossing based method with other spectral estimators

The authors present a method of spectral analysis for formant estimation, using statistical properties of zero-crossing (ZC) intervals of the signal. The technique provides noise robustness using an initial stage of filter-bank analysis. The method is compared with other existing spectral analysis techniques for their performance in formant estimation using a hand-marked speed database. It is found that the ZC method performs best for estimating the first formant frequency at high noise levels (<0 dB signal-to-noise ratio). The linear-prediction-based methods deteriorate below 15 dB, and the cepstrum results in too many spurious peaks. The ZC method is also better than autoregressive moving average (ARMA) spectral estimation, which has been proposed for noise robustness.<>