A neural network mapper for stochastic code book parameter encoding in code-excited linear predictive speech processing

摘要

The authors present a novel method of stochastic code book (SCB) searching for code excited linear predictive (CELP) coding by implementing the counterpropagation neural network model. The high performance of CELP is achieved at the expense of very high computational power required to find the SCB parameters. The counterpropagation neural network model is used to replace the exhaustive serial searching process by an open-loop, less computationally demanding code book parameters encoding. A scheme to embed the neural network model into the original CELP coding is presented. The scheme is equivalent to a standard CELP with a 512 word SCB. The system performance is analyzed and compared with the present closed-loop parameter searching method.