Robust Linear Prediction for Formant Estimation

The interaction of source excitation and the linear vocal tract filter often causes problems in estimating the formant frequencies for voiced speech. An innovative algorithm based on Robustness Theory is proposed here for linear predictive signal analysis. Rather than minimizing the sum of squared residuals as in conventional linear prediction, this robust linear prediction procedure minimizes the sum of properly weighted residuals. By relaxing the usual minimum variance assumption in estimating the linear predictive coefficients, the algorithm achieves a much better source/filter separation, especially for the cases of non-Gaussian source excitation like in voiced speech.