STSR: High-Fidelity Speech Super-Resolution via Spectral-Transient Context Modeling

Abstract

Speech super-resolution (SR) requires a delicate balance between preserving global harmonic coherence and sharp local transients. While diffusion-based models achieve impressive fidelity, their computational demands limit practical use; conversely, efficient time-domain architectures often struggle to capture the long-range spectral dependencies essential for precise harmonic alignment. We introduce STSR, an end-to-end MDCT-domain framework designed to bridge this gap. STSR leverages a Spectral-Contextual Attention mechanism with hierarchical windowing to aggregate non-local context, ensuring consistent harmonic restoration up to 48 kHz. To prevent the loss of transient clarity in compressed spectral representations, we incorporate a sparse-aware regularization strategy. Experimental results demonstrate that STSR surpasses state-of-the-art baselines in both perceptual quality and zero-shot generalization, establishing a robust, real-time paradigm for high-fidelity speech restoration.