Adaptive IIR filters offer a simple method for in-situ tracking of loudspeaker parameters. The convergence properties of standard adaptive IIR algorithms can, however, be sub-optimal, when applied to loudspeaker system identification using electrical impedance measurement. The normalized-LMS IIR algorithm described by Widrow and Stearns (1985) [1] either converges slowly or creates excess residual error. This paper shows that, by varying the convergence parameter with the error signal, and by using the square-root of the gradient estimate, convergence time and residual error can be kept within useful limits. Examples based on music signals measured from actual loudspeakers are shown.