New AI System Can Restore Old Music, Making It Sound New Again
This could bring your grandparents' scratchy records back to life.
This paper presents the NPU-ASLP music restoration system developed for Track 1 of the China Computer Federation (CCF) Advanced Audio Technology Competition (AATC) 2026 Challenge. The task aims to recover clean music from mobile-device-captured recordings affected by coupled degradations from playback devices, acoustic propagation, environmental noise and recording terminals. System development is further constrained by limited and imperfectly aligned real data pairs and a strict model-complexity budget.
In our system, we adopt TF-MoE as the backbone for its dual-axis modeling of cross-band and temporal dependencies, whose Top-1 routing activates a single expert per token to bound computation under the strict budget. We extend it with two task-specific adaptations: a scale-aware encoder and a complex mapping decoder. The former encodes normalized complex spectra jointly with per-band log-RMS to preserve the band-energy cues discarded by band-wise normalization.
The latter directly estimates the clean complex spectrum rather than masking the input, thereby repairing coupled degradations. To address the limited, imperfectly aligned real recordings, we further build a challenge-calibrated playback--propagation--recording simulator that generates 1,572.12 h of sample-aligned data across 202 acoustic conditions. The final system contains 2.951M parameters and requires 82.659 GMACs for a 10-s input. On the official blind test, it achieved -49.067 dB SI-SNR, 1.273 LSD, 2.540 FAD, and 3.495 ViSQOL. It ranked third in LSD, FAD, and ViSQOL and received the third prize in the challenge.
- AI can now clean up old recordings, removing noise and clicks.
- It could restore vinyl records, cassette tapes, and historical audio.
- The technology is still in research, not available for public use yet.
Why It Matters
Your old music and family recordings could soon sound brand new, preserving memories for future generations.