Machine learning improves hearing aids

Auditory Neuroscience A key challenge in improving assistive hearing devices is isolating and separating different voices in noisy or crowded settings. This task is especially difficult without prior exposure to the voices of interest and without knowing which voice to amplify. Han et al. used deep attractor networks—a powerful machine-learning technique—to project unfamiliar mixed audio signals into a high-dimensional space to separate signals from independent speakers. Comparing these separated sources with auditory cortical responses of a user allowed the method to determine and amplify the attended voice. The approach was as accurate as methods trained with clean audio sources and may thus enable hearing-impaired users to communicate more naturally in complex social environments. Sci. Adv. 10.1126/sciadv.aav6134 (2019).