Bandwidth Extesion Of Audio Signals Based On Neural Network

Based on conventional and existing nonlinear audio bandwidth extensiontechniques, a blind audio bandwidth extension algorithm from wideband (0~7kHz) tosuper-wideband (0~14kHz) was implemented by incorporating the nonlinear theoryand neural network model. Furthermore, the proposed technology was applied into thereal codec.Based on the verification of the nonlinear characteristics of the audio signals,RBF neural network and BP neural network were utilized to predict the fine structureof high-frequency in the reconstructed phase space. In order to adapt to the actualcircumstance of fine structure prediction, fuzzy theory was introduced to optimize theRBF neural network. In addition, the technologies of additional momentum term andadaptive learning rate were used to improve the learning rate.In order to estimate the sub-band envelope of high-frequency components, RBFneural networks were utilized to fit the relationship between low-frequency featuresand sub-band envelope of high-frequency components. To accurately recover thesub-band envelope of high-frequency components, all the RBF neural networks weretrained by the super-wideband audio signals offline and used to estimate the sub-bandenvelope of high-frequency components online. Combining with the predicted finestructure, a blind audio bandwidth extension algorithm was implemented.In order to evaluate the performance of the proposed nonlinear bandwidthextension algorithm, it has been applied into ITU-T G.722.1wideband audio codec,and the performance is compared with ITU-T G.722.1C super-wideband audio codec.Both the objective and subjective quality tests show that, in most cases, theperformance of extended G.722.1is comparable with G.722.1C super-wideband codecat the bit rate of24kbps.