A neural network based coding mode selection scheme of hybrid audio coder

New generation mobile communication transfers multiple contents, thus, codec no longer processes specific audio signal content (speech or music) only, but has to process mixed audio signal content. As different audio signal types have different characteristics, they have to be processed by using proper coding core. Hybrid audio coder architecture can be adopted to improve the coding efficiency. The coder must determine signal type correctly in order to maintain the quality of decoded audio signal. The computational complexity of hybrid audio codec can be reduced if signal classification can be performed quickly. In this paper, a low complexity coding mode selection technique based on neural network has been investigated. The experimental results demonstrate that the proposed scheme enables a reduction of about 75 percent of the computational load of 3GPP AMR-WB+ coder. The degradation of SNR is about 0.4dB. The informal listening test shows that the quality degradation is almost negligible. The proposed scheme is more effective than the fast algorithm recommended in standard.