Information-Theoretic Limits on the Performance of Auditory Attention Decoders

Speaker-specific attention decoding from neural recordings to suppress the acoustic background and extract a target speaker in an in-the-wild multi-speaker conversation scenario poses a cornerstone challenge for advanced hearing devices. Despite several recent advances in auditory attention decoding, most existing approaches fail to reach the real-time performance and attention decoding accuracy required by hearing aid devices. In this work, we aim to quantify fundamental limits on the performance of auditory attention decoding by establishing and computing the trade-off between accuracy and decision window length. We demonstrate the utility of our theoretical bounds in benchmarking the performance of existing widely-used attention decoding algorithms using both simulated and experimentally recorded magnetoencephalography data.