FSC-Net: New AI Model Upgrades 4kHz Audio to Studio Quality with 1.5M Parameters

FSC-Net (Full-Spectrum Context Network), developed by Xinan Chen and colleagues, tackles the long-standing problem of speech bandwidth extension (BWE) – converting low-quality narrowband audio (e.g., 4kHz) into high-fidelity wideband sound (e.g., 48kHz). Previous methods often introduced perceptual artifacts by failing to reconstruct realistic high-frequency phase and harmonic structures. FSC-Net integrates Fast Fourier Convolutions (FFCs) into a complex spectral mapping framework, giving the model an effective receptive field across the entire frequency spectrum. This allows it to explicitly model cross-band harmonic dependencies. A novel frequency-progressive learning curriculum then guides the network from coarse to fine spectral detail reconstruction, addressing the inherently ill-posed nature of generating high-frequency content.

Experimental results on the VCTK dataset (4kHz-to-48kHz) and the unseen EARS dataset demonstrate consistently strong reconstruction quality and generalization. FSC-Net achieves leading Log Spectral Distortion (LSD) and Perceptual Evaluation of Speech Quality (PESQ) scores, outperforming scaled-up baseline models while maintaining a highly compact parameter count of just 1.54 million. This efficiency makes FSC-Net practical for deployment in real-time voice applications such as teleconferencing, hearing aids, and voice assistants, where bandwidth or device compute is limited but audio quality is critical.