Real-world audio often contains overlapping sound events, yet standard self-supervised methods treat each clip as an isolated unit. The paper introduces MiCLA, a two-stage pre-training framework that uses synthetic log-mel mixing to explicitly align mixture representations with their constituent clips.
Stage 1 structures single-clip representations through masked self-distillation and contrastive regularization. Stage 2 then performs bidirectional alignment: for a synthetic mixture, MiCLA suppresses the component aligned with one constituent and aligns the residual with the other, using projective residualization on patch tokens and sample-adaptive soft residualization on the global class token.
The authors report 43.2% mean average precision on AS-20K and 50.3% on AS-2M under fine-tuning, and the highest mAP among evaluated models on both AudioSet settings under frozen Convex Gated Probing. A Joint Semantic Retrieval diagnostic also shows improved simultaneous retrieval of both constituent classes over target-averaging baselines. As the sole source, this summary reflects the paper's claims without independent comparison.