New adaptive nonlinear FIR filters for sub-band processing in speech enhancement systems

A new class of general ANN based adaptive nonlinear FIR type filters was presented. Simulated data were used to show that the use of these nonlinear filters within a multi-band noise cancellation system can significantly enhance its performance compared to the conventional linear filtering based multi-band and full-band noise cancellers. The superior performance of the MBNLFIR system is due to the use of nonlinear ANN based processing within the subbands, which enables a more efficient modeling of the nonlinear noise transfer function. Furthermore, the complexity of the MBNLFIR system is also comparable to that of the MBLFIR system, and it can be further reduced by employing for example, a self-structuring LMS type algorithm. Currently, experiments are ongoing using real anechoic speech signals.