Output Recursively Adaptive (ORA) Tree Coding of Speech with VAD/CNG

We study a new speech coding architecture based on the concept of tree coding along with recursive least squares lattice short term prediction and gradient/autocorrelation based long term prediction algorithms that adapt on the reconstructed output values of the codec, thus substantially reducing the number of parameters to be quantized, coded, and transmitted. The codec uses a perceptually weighted distortion measure similar to the popular analysis-by-synthesis codecs widely deployed today, but the codec differs from current codecs in that the codec is a sliding block analysis-by-synthesis scheme that is nearer waveform coding, which should allow smoother tracking of transitions and the reduction of block based artifacts. Voice activity detection (VAD) and comfort noise generation (CNG) are added to reduce the bit rate and the number of computations required. We present performance comparisons with G.711 using the same VAD/CNG and investigate the performance of a standard long (pitch)/short term predictor structure with a 10th order autoregressive predictor. It is shown that a bit rate reduction of 70 per cent can be achieved while maintaining nearly perceptually equivalent speech quality.