Transformer-based audio classifiers excel at capturing long-range dependencies but pay for it with quadratic complexity. Mamba, a state-space model, offers linear scaling, yet its audio classification performance has lagged behind. The new AD-Mamba framework addresses this by using multi-stage knowledge distillation: a CNN teacher guides the student on local spectro-temporal patterns, then a Transformer teacher refines global context and semantic discrimination.
On AudioSet, AD-Mamba reaches 50.3% mean average precision with only 30 million parameters, outperforming representative Transformer- and Mamba-based baselines at lower computational cost. The authors also tested it under long-form audio, sparse information, and noise interference, where it consistently beat CNN and Transformer baselines.
Transferring the model to FSD50K, ESC-50, and DCASE16 produced superior results across all three, suggesting the distilled Mamba generalizes well beyond AudioSet. The paper is available on arXiv as 2610.27713.