DOMAIN-INCREMENTAL AUDIO CLASSIFICATION USING DOMAIN-SPECIFIC EXPERTS AND PROTOTYPE CLASSIFIER

In domain-incremental learning (DIL), models must learn to classify audio across sequential domains without accessing past data, which typically leads to catastrophic forgetting. This paper addresses the DCASE 2026 Challenge Task 7, which requires maintaining high classification accuracy on ten target classes as new, distinct audio domains arrive, without the ability to revisit raw data from previous stages.

The authors propose a frozen-feature replay architecture. At each stage, they train one or two compact expert models on the current domain and then freeze them. To prevent catastrophic forgetting, they use DeepInversion-based generative replay to approximate past data. Because later-stage experts do not exist when processing earlier-stage data, the authors train a cross-stage regression MLP to impute the missing feature slots. Finally, they concatenate the penultimate features from all frozen experts and train a lightweight, per-class cosine prototype classifier on these cached features.

The proposed system significantly outperforms the official DCASE 2026 baseline, achieving approximately 78% micro-accuracy on the development set. The authors demonstrate that the diversity of the expert pool is the primary driver of performance; adding specialized experts trained from scratch on individual domains provides complementary representations that the prototype classifier can effectively leverage. The cross-stack ensemble of different expert backbones further improves robustness, mitigating the performance variance observed in individual models.

This work provides a practical, modular framework for continual learning in audio classification. By decoupling feature extraction (via frozen, domain-specific experts) from classification (via a prototype head), the system avoids the need to update shared parameters, effectively eliminating catastrophic forgetting while maintaining a manageable memory footprint.