Short-term temporal decomposition and its properties for speech compression

A technique of temporal speech compression, called temporal decomposition, is considered. A mathematical presentation of temporal decomposition is discussed from the point of view of speech production. Temporal decomposition is studied in terms of an iterative solution and its convergence. For vocoder applications, a practical algorithm, short-term temporal decomposition, is proposed. The statistical performance of this algorithm was tested on a multispeaker database. The efficiencies of vector quantization and vector predictability were comparatively studied on data derived from the algorithm and on data from linear predictive coding (LPC) analysis. The results show that data from the algorithm have more efficient information transmission via vector quantization than data from LPC, and that vector prediction coding is feasible for the data from temporal decomposition.<
>