BiEAR: Bio-Inspired AI Localizes Multiple Speakers with Adaptive Hearing

Traditional machine hearing systems struggle in noisy, multi-speaker environments because they rely on fixed binaural front-ends that cannot adapt to changing acoustic conditions. Now, a team led by researchers from several institutions (including Hanyu Meng, Eliathamby Ambikairajah, and Haizhou Li) has introduced BiEAR—a biologically inspired system that mimics the human ear's medial olivocochlear (MOC) reflex. This reflex automatically adjusts the ear's sensitivity to focus on important sounds while filtering out noise. BiEAR implements this adaptation via a neural controller that dynamically adjusts the frequency selectivity of a binaural auditory filterbank in real time during inference. The result is a time-frequency adaptive representation that changes as sound conditions shift, much like how our ears tune into a conversation in a loud room.

The system was evaluated on multi-speaker localization (pinpointing speaker positions) and distance estimation tasks, tested in both anechoic chambers and realistic room environments with reverberation. BiEAR consistently outperformed commonly used fixed binaural front-ends, showing higher localization accuracy and greater robustness when confronted with unseen speakers and unfamiliar room acoustics. Visualizations of the learned filter adaptations reveal that BiEAR emphasizes informative frequency bands over time, effectively prioritizing speech-relevant signals. Accepted at INTERSPEECH 2026, this work demonstrates that adaptive, biologically inspired front-ends can substantially improve machine hearing in complex acoustic scenes—with potential applications in hearing aids, smart speakers, robotic navigation, and surveillance systems.