A new excitation model for linear predictive speech coding at low bit rates

Systems for high-quality low-bit-rate speech coding based on linear prediction, such as CELP, multipulse LPC, and self-excited vocoders (SEV) are among the best techniques for coding audio at medium-to-low bit rates (9.6 kb/s to 4.8 kb/s). For rates of 4.8 kb/s and below, vector quantization (or block coding methods) applied to the excitation signal has been shown to be very promising. An approach is presented that is similar to the CELP and SEV systems in that it uses an analysis-by-synthesis scheme but differs greatly in its method of representing the excitation information of the audio signal. In particular, the excitation system is based on a multistate model where the state transition parameters are determined by an external classifier. Experiments are reported that demonstrate the feasibility of the class structuring approach used and provide some basic information about the system´s adaptation processes