A neural network based perceptual audio coder

The implementations and performance results of a neural network based perceptual audio coder is reported. The coder uses the configuration of the ISOIMPEG audio layer II coder with the perceptual analysis block replaced by a 2 layer network trained to estimate the masking thresholds required for the bit allocation. The 2 layer network is trained by a back propagation algorithm using the energies in the subbands as inputs and the masking thresholds of obtained from psychoacoustic model II of the ISOIMPEG audio coder as the reference outputs. The result is a coder which performs favourably in quality against the ISOIMPEG audio layer 2 coder at bit rates of 256 kbit/s and 192 kbit/s stereo. Performance at a bit rate of 128 kbit/s stereo was however, found to be poorer