A Probabilistic Generative Model for Spectral Speech Enhancement

Speech enhancement in hearing aids remains difficult in nonstationary acoustic environments. This paper introduces AIDA-2, a unified modular framework that formulates signal processing, learning, and personalization as Bayesian inference with explicit uncertainty tracking. The proposed framework replaces ad hoc algorithm design with a single probabilistic generative model that continuously adapts to changing acoustic conditions and user preferences. It extends spectral subtraction with principled mechanisms for in situ personalization and adaptation to acoustic context. The system is implemented as an interconnected probabilistic state-space model, and inference is performed via variational message passing in the RxInfer.jl probabilistic programming environment, enabling real-time Bayesian processing under hearing-aid constraints. We focus the evaluation on the warped-frequency filter bank (WFB) and speech enhancement model (SEM), evaluated on the public VoiceBank+DEMAND corpus. Additionally, an acoustic context model (ACM) and end-user model (EUM) are described architecturally. The evaluated system requires 85 parameters, two orders of magnitude fewer than the tens of thousands to millions of trainable parameters used by contemporary deep-learning baselines. This proof of concept establishes an interpretable, data-efficient foundation for uncertainty-aware hearing-aid processing, pointing toward devices that learn continuously through probabilistic inference.