On robust Kalman filtering with forgetting factor for sequential speech analysis

Abstract We propose a robust Kalman filter with forgetting factor to estimate the time-varying parameters of speech signals. The proposed robust Kalman filter is based on a modified least-squares criterion with forgetting factor. The input signal is assumed to have a heavy-tailed non-Gaussian nature with outliers due to spiky excitation. To alleviate the effects of outliers, this algorithm extends the concept of Huber's min-max approach, named M-estimation , to the Kalman filtering. The introduction of forgetting factor enables the time-varying speech parameters to be estimated, giving more weight on the most recent portion of the data. Experimental results show that the proposed algorithm achieves more accurate estimation and provides improved tracking performance.