Enhanced CELP coding with discrete spectral modeling

During the coding of speech with the code excited linear prediction (CELP) method, the linear prediction (LP) filter coefficients are usually calculated by standard autocorrelation or covariance methods. These methods minimize the mean squared error between the speech signal and predicted value. The perceptual quality of the coded speech is indicated by the spectral distortion measured over a set of discrete frequencies related to the formant of the speech. We introduce a new spectral modeling approach to refine the LP coefficients obtained by the autocorrelation method and use the refined parameters for CELP coding. Simulation results show that enhanced harmonic structure is obtained in the speech spectrum coded by CELP coding with refined LP parameters. Discussions of informal listening tests are also given.