Estimation of handset nonlinearity with application to speaker recognition

A method is described for estimating telephone handset nonlinearity by matc hing the spectral magnitude of the distorted signal to the output of a nonl inear channel model, driven by an undistorted reference. This "magnitude-on ly" representation allows the model to directly match unwanted speech forma nts that arise over nonlinear channels and that are a potential source of d egradation in speaker and speech recognition algorithms. As such, the metho d is particularly suited to algorithms that use only spectral magnitude inf ormation. The distortion model consists of a memoryless nonlinearity sandwi ched between two finite-length linear filters. Nonlinearities considered in clude arbitrary finite-order polynomials and parametric sigmoidal functiona ls derived from a carbon-button handset model. Minimization of a mean-squar ed spectral magnitude distance with respect to model parameters relies on i terative estimation via a gradient descent technique. Initial work has demo nstrated the importance of addressing handset nonlinearity, in addition to linear distortion, in speaker recognition over telephone channels. A nonlin ear handset "mapping," applied to training or testing data to reduce mismat ch between different types of handset microphone outputs, improves speaker verification performance relative to linear compensation only. Finally, a m ethod is proposed to merge the mapper strategy with a method of likelihood score normalization (hnorm) for further mismatch reduction and speaker veri fication performance improvement.