Sequential State-Space Filters for Speech Enhancement

Particle filters have recently been proposed as a new form of state-space filtering for speech enhancement applications. Despite theoretical foundations that suggest superior performance, robust performance in practical applications has been elusive. This paper presents a comparative analysis of three popular recursive filtering algorithms that share a state-space formulation: Kalman filters, unscented Kalman filters, and particle filters. We present a general formulation of these state-space models and then introduce applications of these to time series prediction based on autoregressive models. Results indicate that for signals produced from linear systems, as expected, particle filters and unscented Kalman filters do not perform significantly better than a Kalman filter. For typical speech signals, the traditional Kalman filter provides more robust performance at the same level of computational and representational complexity. At extremely low signal-to-noise ratios, unscented Kalman filter gave the best results