Vector quantization of speech parameters in the time-frequency domain

We investigate a time-frequency domain vector quantization approach for low bit rate speech coding. The well known CELP standard provides an efficient representation of a speech signal in units of a frame length (30 ms) using an all-pole speech model. For most efficient speech compression, inter and intraframe correlations of speech parameters should be exploited. This observation previously lead to multiframe coding with interpolation, or LSP prediction. But such linear modeling approaches suffer from the nonlinear speech parameter dynamics. The proposed approach eliminates such problems while exploiting intra and interframe correlations. Computer simulations demonstrate significantly improved compression performance (1.5-2 bits/frame) under a given spectral distortion.<<ETX>>