New audio AI learns new classes from few examples without forgetting old ones

Few-shot Class-incremental Audio Classification (FCAC) tackles the challenge of progressively recognizing new audio classes from only a handful of labeled samples while retaining knowledge of previously learned classes—a classic stability-plasticity dilemma. To address this, researchers from the team of Yanxiong Li introduce a decoupled model consisting of two independent modules: an embedding learner built on a residual convolutional network backbone, and a stochastic classifier composed of distributions (each with a mean and variance vector per class). After base session training, the embedding learner is frozen to preserve base class knowledge, while the stochastic classifier is continually updated in each incremental session to adapt to new classes. A key innovation is pseudo-incremental training: during the base session, diverse data augmentation is used to simulate incremental classes, giving the embedding learner stronger representation ability for unseen categories without ever forgetting known ones.

Experiments on three datasets—FSC-89, NSynth-100, and LS-100—demonstrate that the proposed method consistently outperforms existing FCAC techniques in accuracy, while also showing lower computational complexity than most comparison methods. The code is publicly available on GitHub. This work, accepted for publication in IEEE TASLP, offers a practical and efficient solution for real-world audio systems that must continuously adapt to new sounds—such as smart assistants learning new wake words or environmental monitoring systems identifying novel acoustic events—without catastrophic forgetting.