A Kalman filtering algorithm with joint metrics-based tuning for single-channel speech enhancement

Abstract

In this paper, we present an iterative Kalman filtering algorithm that exhibits better speech enhancement by jointly utilising robustness and sensitivity metrics. Typically, poor model parameter estimates lead to a biased Kalman filter gain, which results in innovation noise ‘leaking’ into the output. In the proposed algorithm, the Kalman filter gain is dynamically tuned based on a varying operating point of balanced robustness and sensitivity. Speech enhancement experiments showed the proposed Kalman filtering algorithm to produce higher quality speech than conventional methods using objective and subjective measures.