An LSP based speech quality measure

An objective measure is introduced for predicting the subjective quality of a distorted speech signal. The unique feature of the measure is that it explicitly uses a parametric model of speech perception (rather than a signal fidelity measure) in the design of an objective measure for speech quality. Specifically, the known sensitivity of human auditory perception to perturbations in the speech formants is combined with a simple (statistical) relationship between the line spectrum pair (LSP) coefficients and the vocal tract formants to form a simple objective speech quality measure that performs better than any other known simple measure. A detailed description is given of the LSP measure, the performance results are presented, and its associated implementational advantages are discussed.<
>