Non-linear predictor for speech enhancement

Addresses the application of nonlinear prediction to speech enhancement, considering 3 common cases of speech degraded by distinctly nonlinear system (CELP coder), the addition of (Gaussian) noise, and convolution by a linear system. A time-domain nonlinear predictor, in the form of an MLP neural net structure, is applied as an enhancer and the performance examined in all three cases, with respect to the influence of nonlinearity and net topologies. Experimental results show that, in the case of low bit-rate CELP coder degradation, nets with multiple outputs give significant improvement over single-output structures. It is also clear that in this case, i.e. nonlinear degradation, the matching nonlinear enhancer is consistently better than the equivalent linear structures. In contrast, when the degradation is from additive noise, the (matching) linear enhancer is superior. Again, the multiple output cases give the best results. There is less consistence in the case of linear system degradation. It is observed that the most significant factor here is the number of input nodes which in turn reflects the chosen linear system characteristics