Mixture excitations and finite-state CELP speech coders

Code excited linear prediction (CELP) coding and its derivatives are currently the most frequently used techniques for speech compression at medium-to-low rate range. Until this study, the excitation vectors were always selected from a codebook generated by a Gaussian source or by the ensemble of residual signals collected from the used speech database. To the best of the author´s knowledge, there is no study in which the excitation vectors were formed from a mixture of sources, a notion very successfully used by the speech recognition community within the hidden Markov model (HMM) framework. The authors proposed an improvement to the excitation model of CELP coders by embedding a labeled-state finite state vector quantization (FSVQ) and a mixture density approach in constructing 5-ms-long excitation vectors