Online System Identification on Learned Acoustic Manifolds Using an Extended Kalman Filter

Many communication devices require real-time tracking of acoustic paths based on noisy measurements. For this task, solutions such as the Kalman filter are known, which require little or no information about the system to be identified. When the acoustic path is time-varying and the measurements are noisy, these algorithms reach their limits. In this paper, we combine the Kalman filter with neural autoencoders to overcome these limitations. Our approach restricts the set of possible systems to a lower-dimensional manifold, while still exploiting the sophisticated step-size control of the Kalman filter. In simulations, the algorithm shows state-of-the-art performance with reduced computational complexity.