Impedance formulation of l.p.c. using a lossy vocal tract and lip termination

Linear predictive coding (l.p.c.) of speech waveforms is based on a model of the human vocal tract. For reasons of mathematical convenience it is generally assumed that the model is lossless and terminated at the lips with a short circuit. The author believes that the absence of loss in the model leads to extracted vocal tract parameters which are unrealistic, and to a correspondingly poor synthetic-speech quality. In the paper the l.p.c. procedure is reformulated in terms of the vocal tract input impedance instead of its transfer function, thereby permitting loss to be incorporated in the model and its lip termination. Qualitative reasons are advanced to support the author's view that more realistic results should be achieved by this approach. Preliminary investigations using a short sample of real speech have demonstrated that intelligible and clear speech can indeed be generated from the calculated parameters. The results of formal tests and a comparison with conventional l.p.c. will be published when they become available.