Bridging conventional signal processing and edge AI: Speech enhancement for hearing devices under perceptual and computational constraints

Abstract

Prescriptive amplification restores audibility, yet many hearing-aid users continue to struggle in noise because amplification does not resolve target-to-masker competition. Conventional noise reduction improves comfort and effort more reliably than intelligibility, whereas spectral and formant enhancement approaches strengthen selected speech cues. Deep learning can produce larger laboratory gains, but translation to wearable devices requires causal, real-time processing within tight memory, computation, and battery budgets. This review connects evidence from conventional signal processing, cue-directed enhancement, human-listener studies, and hardware implementations. It argues that hearing-device algorithms should be judged by perceptual benefit within a stated device budget. Targeted formant enhancement illustrates interpretable auditory-informed processing, while compact neural models may support target identification, scene analysis, and control. A likely near-term direction is a hybrid architecture in which compact AI guides efficient, controlled signal processing.

Keywords: Speech enhancement; conventional signal processing; artificial intelligence; hearing aids.