Projected subset least squares for robust linear prediction of speech

A projected subset least squares method is presented as a new method for robust linear prediction of speech. The proposed algorithm combines conventional principal component analysis techniques with subset least squares method to perform robust linear regression. The subset least squares is a univariate estimator which applies the generalized maximum likelihood principles to obtain a proper set of inliers from contaminated data. We then use the chosen subset when performing least squares fit. In contrast, the conventional linear prediction procedure weights all prediction residuals equally. In comparison to conventional linear prediction algorithms, our method yields a more efficient estimate of the linear prediction coefficients for speech. Testing on natural human speech demonstrates that formant estimation from contaminated data can be greatly improved.