Coding of LSP Parameters Using Interframe Moving Average Prediction and Multi-Stage Vector Quantization

Abstract

This paper proposes a new efficient line spectrum pair (LSP) vector quantization method which uses moving average (MA) interframe prediction of the parameters. This method has two advantages over auto-regressive prediction: the degradation of decoded parameters caused by bit errors affects only a few of the following frames, and codeword decoding can start at any frame. Using MA prediction, the final reconstructed LSP vector (quantized vector) of the current frame is represented as a linear combination of current and previous frame code vectors. In this paper we describe ways to achieve a more efficient LSP coder when each frame has four 10-ms subfraines. The spectral distance obtained by this method at 30 bits per 40 ms is better than that obtained using conventional multi-stage VQ without interframe correlation at 40 bits per 40 ms.