Regularized linear prediction all-pole models

Abstract

For many cases of voiced speech, linear prediction (LP) based all-pole spectral envelopes exhibit unnatural vocal tract transfer functions that underestimate the formant bandwidths. To obtain smoother contoured all-pole spectral envelopes, we employ a regularization measure which discourages nonsmooth behavior of the transfer function. In particular, we demonstrate how a simple regularization scheme can be incorporated into the LP framework without the need for iterative numerical optimization or spectral sampling. Our results indicate that regularized LP all-pole models can provide more accurate vocal tract transfer function modeling than conventional LP, particularly at the formants.