Nonlinear Speech Coding Using Backward Adaptive Variable-Length Quadratic Filters

The ADTCM coding technique with nonlinear prediction based on quadratic Volterra filters is examined using backward prediction schemes based on LMS and RLS algorithms. Utilizing backward adaptive quadratic filters in ADTCM based speech coding, by itself, does not result in an overall improvement in the quality of reconstructed signal in comparison with a linear scheme using the same bit rate. However, it is shown that a scheme can be developed in which, for each frame of constant length, a set of quadratic filters with different memory sizes is examined and the nonlinear filter resulting in best improved quality is decided on. The identifying code of the selected filter is sent to the decoder along with the quantized residual signals. The simulation results show that the proposed scheme results in a good improvement (up to 2 dB) in the overall quality of the reconstructed speech signal. This improvement is achieved at the cost of a slight increase in the bit rate and a small delay.