Performance of the weighted burg methods of ar spectral estimation for pitch-synchronous analysis of voiced speech

Recemly three different modifications over the Burg method of autoregressive (AR) spectral estimation are proposed by Swingler [5], Kaveh and Lippert [6], and Scott and Nikias [7], where the linear prediction error is weighted prior to its minimization. In the present paper, we study these weighted Burg methods for pitch-synchronous analysis of short segments (duration less than one pitch period) of non-nasalized voiced speech and make their comparative performance evaluation. Errors in estimating the power spectrum, formant frequencies and formant bandwidths are used as criteria for performance evaluation. It is shown that the weighted Burg method of Kaveh and Lippert results in the best performance. These methods are also compared with the autocorrelation and covariance methods and the results are