A new regularized TVAR-based algorithm for recursive detection of nonstationarity and its application to speech signals

This paper develops a new recursive nonstationarity detection method based on time-varying autoregressive (TVAR) modeling. A local likelihood estimation approach is introduced which gives more weights to observations near the current time instant but less to those distance apart. It thus allows the Wald test to be computed based on RLS-type algorithms with low computational cost. A reliable and efficient state regularized variable forgetting factor (VFF) QR decomposition (QRD)-based RLS (SR-VFF-QRRLS) algorithm is adopted for estimation for its asymptotically unbiased property and immunity to lacking of excitation. Advantages of the proposed approach over conventional approaches are 1) it provides continuous parameter estimates and the corresponding stationary intervals with low complexity, 2) it mitigates low excitation problems using state regularization, and 3) stationarity at different scales can be detected by appropriately choosing a certain window size. The effectiveness of the proposed algorithm is evaluated by testing vocal tract changes in real speech signals.