This paper presents an approach to compensate the effects of noise with an Interacting Multiple Model algorithm using Unscented Kalman Filters (IMM-UKF) in log-spectral domain. The performance of this approach is studied experimentally on a continuous digits recognition task with additive noise conditions and compared with results previously obtained by the implementation of the Interacting Multiple Model algorithm using Extended Kalman Filters (IMM-EKF) in log-spectral domain. Simulation results show that a better performance in terms of word recognition rates can be obtained with the suggested approach.