Restoring Music Integrity via Predictive Modeling and Spectral Inpainting

C. Aironi, L. Gabrielli, S. Cornell, S. Squartini (2025)

We present bin2bin-v2, a hybrid deep learning framework for Packet Loss Concealment (PLC) in Networked Music Performance (NMP). The system combines a Linear Predictive Coding (LPC) stage with a GAN-based spectrogram inpainting network, restoring missing audio segments in real time. When a packet is lost, LPC provides a coarse time-domain estimate, while bin2bin-v2 refines the result in the time–frequency domain, recovering spectral details and improving perceptual continuity. The proposed method was submitted to the IEEE-IS² Music PLC Challenge, where it achieved first place, demonstrating superior quality and efficiency compared to state-of-the-art approaches. The system runs efficiently on standard CPUs and delivers smoother, more natural reconstructions of monophonic and harmonic musical content under realistic network conditions.