Audio‑visual speech enhancement (AVSE) improves intelligibility in noisy settings but typically relies on computationally heavy models unsuitable for resource‑limited edge devices. This paper introduces a reinforcement‑learning‑driven neural architecture search (RL‑NAS) framework to automatically discover compact and efficient AVSE networks. The search space preserves temporal alignment between audio and visual streams, and the reward function jointly optimises SI‑SNR improvement and model compactness. Using proximal policy optimisation (PPO) with constraint‑aware action masking, only 35 candidate architectures are evaluated to identify a model with 2.9 M parameters, compared to 22 M in the baseline. The resulting model achieves a 14.5 dB SI‑SNR gain, just 0.7 dB below the baseline, while reducing model size by roughly nine‑fold. Real‑time factor measurements on commercial mobile CPUs show a 4.3‑fold reduction, confirming practical on‑device deployment. Compared with random search, RL‑NAS improves search efficiency by 92%. Architectural analysis reveals that shallower yet wider temporal convolutional blocks and aggressive pruning of the visual pathway can retain enhancement performance while markedly lowering computational demands.
Review: The study demonstrates the promise of reinforcement‑learning‑guided NAS for creating lightweight, deployable multimodal speech enhancement systems.