Hybrid of PSO and SOM neural network for immittance spectral frequency quantization in AMR-WB speech codecs

Abstract

Adaptive multi-rate wideband (AMR-WB) speech coding algorithm is standardized by 3 rd Generation Partnership Project (3GPP) and third generation/wideband code division multiple access (WCDMA 3G) system and has several applications in mobile communication systems, audio and video teleconferencing, and digital radio broadcasting. To reduce the bit rate in AMR-WB, vector quantization (VQ) technique is used. In this paper, a self-organizing map (SOM) neural network is used instead of traditional VQ method proposed in ITU-T G.722.2 recommendation for AMR-WB speech coding. As another contribution of this study, the SOM is trained using a modified supervised training algorithm whose training parameters are optimized by particle swarm optimization (PSO) algorithm. The total number of operations in the proposed method is reduced as compared to split vector quantizer (S-SVQ) and split multistage vector quantizer (S-MSVQ) with no significant spectral distortion.