LOUDAR's diffusion-based AI restores music from unknown distortions

Music restoration typically assumes you know the exact degradation—reverb, clipping, or effect pedals—and trains models on paired clean/degraded recordings. LOUDAR (Latent-space Optimization of Unknown Distortion for Audio Restoration) flips this assumption. Developed by Michal Švento, Eloi Moliner, and colleagues, it works in the latent space of a pretrained audio autoencoder, where it learns a parameterized operator that approximates the unknown distortion at inference time. The algorithm alternates between estimating the clean latent representation and updating the operator parameters, while an unconditional latent diffusion model acts as a prior, steering the estimate back onto the manifold of clean audio. This per-input adaptation means LOUDAR needs no paired training data and no prior knowledge of the forward degradation process.

In experiments on singing voice effect removal and guitar distortion removal, LOUDAR consistently improved over degraded inputs and proved competitive with existing supervised and unsupervised baselines in both waveform and latent domains. The paper, accepted at ISMIR 2026, positions LOUDAR as a versatile tool for scenarios where the corruption is uncharacterized—such as old recordings, unknown plugin chains, or amateur capture. The code and data are likely to follow on arXiv, offering researchers a new baseline for blind audio restoration. It also highlights the growing use of latent diffusion models not just for generation, but as learned priors for inverse problems in audio processing.