Estimation of Diagonal Volterra Kernels of an Audio System During Normal Operation with Multiple Least Mean Squares Adaptive Filters

The usage of Complete Volterra Kernels for emulating the nonlinear behavior of sound systems has been investigated for decades. Due to the computational load, the real-time implementation is typically limited to second order distortion and not feasible for higher orders. This is usually unsatisfactory for audio systems in which the disturbing distortions occur mostly at orders three and five. The same authors of this work already solved the problem with the Diagonal Volterra Kernels technique, which allowed to model arbitrarily high distortion orders. The estimation of the coefficients was obtained by exciting the system with an Exponential Sine Sweep signal. However, the result was often suboptimal since the signal reproduced by the sound system is usually different from a sinusoid. In this paper, a new method for estimating the Diagonal Volterra Kernels coefficients is proposed, by employing any music, noise or speech signal being played by a sound system in real-time. Multiple Least Mean Square algorithms are used to estimate the coefficients up to the 5
^th
distortion order, thus allowing to emulate the nonlinearities of a typical audio system.