Speech Coding Algorithm with Dynamic Weighted Inter-Frame Linear Prediction

Abstract

A new multi-frame joint quantization algorithm with dynamic weighted inter-frame linear prediction based on mixed excitation linear prediction (MELP) is proposed in this paper. In encoding stage, a super-frame consists of three adjacent single-frames. Fourier magnitudes and aperiodic jitter flag are eliminated. The other parameters are jointly quantized. LSF of the first and third frame are quantized as a 20-dimensional vector. According to the BPVs of super-frame, pitch is quantized with codebooks of dynamic size. In decoding stage, parameters are indexed from corresponding codebook. The LSF of middle frame are predicted from the first and third frame. The weighted factors keep changing in accordance with the BPVs of adjacent five frames. Results show that the reconstruction accuracy of LSF is significantly improved using dynamic weighted inter-frame linear prediction. Meanwhile the coding bit rate is reduced to 0.6 kbps.