Switching source online segmented recursive least squares for real-time stimuli tracking in auditory attention decoding

The stimulus reconstruction-based auditory attention decoding (AAD) model reconstructs the attended auditory envelope from an electroencephalogram (EEG) recording with a finite impulse response (FIR) filter of an arbitrary order. Nonetheless, by nature, the filter coefficients, as well as the source of interest, are free to independently vary over time. To model such a phenomenon, the existing AAD works utilize piece-wise constant linear predictor of fixed observation length to track the switching source of interest as well as the dynamic linear predictive coefficients. However, such a method requires an appropriate selection of the fixed partitioning size, a priori. A suboptimal selection can exacerbate the overall performance, especially given sequences where the underlying structure does not conform to a constant partition size. Our proposed algorithm, the switching source online segmented recursive least squares (SS-OSRLS), based on the recursive Bellman equation identifies a set of partition points and the corresponding piece-wise constant linear predictors for each detected sub-segments, as well as the detected speaker of interest, all in real-time. We aim to demonstrate the real-time performance of our algorithm using the real EEG/audio dataset (i) without and (ii) with the pretrained decoder weight.