Survey Tracks Audio Super-Resolution Shift from Discriminative to Generative Models

A comprehensive survey from researchers at multiple institutions (Yang et al.) published on arXiv (2605.16681) provides a structured taxonomy of audio super-resolution (SR) and bandwidth extension (BWE) techniques. The paper traces the evolution from early discriminative deep neural networks, which treat BWE as a deterministic mapping and often suffer from regression-to-the-mean and spectral over-smoothing, to modern generative models. Specifically, it systematically reviews autoregressive models, VAEs, GANs, diffusion and score-based methods, flow-based approaches, and Schrödinger bridges. The survey also examines key design dimensions: representation domain, architecture, conditioning mechanisms, and the trade-offs between reconstruction fidelity, perceptual quality, robustness, and computational efficiency.

Looking forward, the authors highlight emerging directions involving large language models (LLMs) and multimodal foundation models, which offer new conditioning signals for audio generation. They also identify persistent open challenges, including phase modeling, perceptual evaluation metrics that align with human hearing, and generalization to real-world low-resolution or band-limited inputs. By providing a unified perspective and practical roadmap, this survey aims to guide future research toward distribution-aware generative modeling rather than deterministic point estimation, potentially enabling higher-fidelity audio reconstruction for applications like teleconferencing, hearing aids, and audio restoration.