Complexity Reduction of LD-CELP Speech Coding in Prediction of Gain Using Neural Networks

Reducing the computational complexity is desired in speech coding algorithms. In this paper, three neural gain predictors are proposed which can function as backward gain adaptation module of low delay-code excited linear prediction (LD-CELP) G.728 encoder, recommended by International Telecommunication Union-Telecom sector (ITU-T, formerly CCITT). Elman, multilayer perceptron (MLP) and fuzzy ARTMAP are candidate neural models in this work. Empirical results show that gain prediction by Elman and MLP neural networks improve the mean opinion score (MOS) and segmental signal to noise ratio (SNRseg) as compared to traditional implementation of encoder. However, fuzzy ARTMAP reduces the computational complexity noticeably, without significant degradations in MOS and SNRseg.