Closing the Loop on the Cocktail Party Effect: From Attention Decoding to Neuro-Steered Selective Hearing.

Selective hearing in multi-talker environments remains a central challenge for assistive listening technologies: conventional devices improve audibility but cannot infer which talker the listener intends to follow. This Perspective examines the emerging transition from auditory attention decoding to neuro-steered selective hearing. We argue that a recent real-time intracranial proof-of-principle study marks an important inflection point in this transition, because it shows, under high-fidelity recordings from neurosurgical patients and in controlled two-talker conditions, that decoded neural attention can be coupled online to relative-gain control, thereby improving speech perception and reducing listening effort within that experimental setting. Building on this proof of principle, we argue that the field should move beyond offline decoding accuracy as its primary benchmark and instead treat neuro-steered hearing as a closed-loop translational problem linking neural inference, acoustic scene analysis, and control policy. From this perspective, future benchmarks should integrate device-level control metrics, including latency, effective switch time, false-switch rate and recovery after erroneous updates, with listener-centered outcomes such as speech intelligibility, listening effort, user preference and everyday benefit. Accordingly, we outline a staged roadmap for electroencephalography-based systems, from controlled two-speaker paradigms to scene-aware control and ultimately wearable real-world implementations. We also discuss the complementary roles of magnetoencephalography, functional near-infrared spectroscopy, and functional magnetic resonance imaging in defining the neural mechanisms and design constraints of future systems. Overall, this Perspective frames neuro-steered selective hearing as a systems-level research pathway for evaluating whether neural attention signals can be translated into practical assistive listening technologies.