Formant frequency estimation with windowless autocorrelation in the presence of noise
Yu Iwai, Tetsuya Shimamura · 2014
In this paper, we propose a formant frequency estimation method in the presence of noise. Usually the autocorrelation function of a noisy speech signal is employed in linear predictive analysis. The autocorrelation function itself, however, concentrates the noise components in the part of lower lags and suppresses the effect of noise in the other part. This property motivates the use of only a higher lag autocorrelation function sequence for the linear predictive analysis. However, it can not give a sufficient performance with the pre-emphasis involved in linear predictive analysis. In the proposed method, a windowless autocorrelation function is used and the pre-emphasis is modified according to the signal-to-noise ratio (SNR). By experiments on natural vowels in the presence of white and pink noise, it is shown that the performance of the proposed method is better than that of the conventional method at low SNR.