New fast and robust stochastic algorithm of two stage vector quantization for joint source-channel speech coding for any transmission channel

A new algorithm of two stage vector quantization for joint source-channel speech coding for any transmission channels is presented. The computational complexity is only slightly higher than the most widely used Multi Stage Vector Quantization algorithm (MSVQ). This new algorithm improves the characteristics and the results of a sequential quantizer of two stages. The base of this algorithm is the modification of the well-known GS-RGSKAε algorithm (Reduced complexity Generalized Stochastic K-means Algorithm -of- Great Speed) for a nonstationary channel. This new algorithm is optimal for the joint construction of two stages. The main features of the proposed algorithm are as follows: Due to its stochastic nature it avoids being trapped in poor local minimums. Initial codebooks are not needed; the codevectors move away from the gravity center of the training vectors towards their final position. Source coding and channel coding are jointly optimized to obtain robust codebooks for different levels of noise in transmission channels. The reduction of calculation time is based on geometric considerations and memory management. This algorithm allows to design the codebook orderly due to its advantageous convergence properties. The results showed that the algorithm needs only 8 to 16% of the number of mathematical operations in comparison with the operations required by others propositions for full search of MSQV, either stationary or non-stationary channels.