Prediction of Gain in LD-CELP Using Hybrid Genetic/PSO-Neural Models

In this paper, the gain in LD-CELP speech coding algorithm ispredicted using three neural models, that are equipped by genetic and particleswarm optimization (PSO) algorithms to optimize the structure and parameters ofneural networks. Elman, multi-layer perceptron (MLP) and fuzzy ARTMAP are thecandidate neural models. The optimized number of nodes in the first and secondhidden layers of Elman and MLP and also the initial weights and biases of thesenets are determined by genetic algorithm (GA) and PSO. In the fuzzy ARTMAP, thechoice parameter, , learning rate, , and vigilance parameter, , are selected byGA and PSO, as well. In this way, the performance of GA and PSO are comparedwhen using different neural architectures in this application. Empirical results showthat when gain is predicted by Elman and MLP neural networks with GA/PSOoptimizedparameters, the segmental signal to noise ratio (SNRseg) and meanopinion score (MOS) are improved as compared to traditional implementationbased on ITU-T G.728 recommendation. On the other hand, fuzzy ARTMAP-basedgain predictor reduces the computational complexity noticeably, with no significantdegradations in SNRseg and MOS.

Keywords:

Speech coding , neural networks , genetic algorithm , particle swarmoptimization