Prototype adaptation method handles audio class additions and removals
New FCIAC method outperforms prior approaches on three public datasets.
Traditional few-shot class-incremental audio classification assumes classes only increase over time, but in real-world scenarios the number of classes can also decrease. To address this, Yanxiong Li and colleagues from an unnamed institution introduce Few-shot Class-variable Incremental Audio Classification (FCIAC). Their method consists of an encoder and a classifier, where the classifier is initialized by a class-variable prototype adaptation network that dynamically changes its structure as classes are added or removed. Additionally, they design a pseudo class-variable training strategy that exposes the model to both increasing and decreasing class configurations during training, enhancing its adaptability to real-world changes.
The experiments on three public datasets show that the FCIAC method outperforms previous few-shot incremental audio classification methods in average accuracy. The paper has been accepted for publication at Interspeech 2026 and includes code on GitHub. This work is significant because it moves beyond the unrealistic assumption of monotonically increasing class counts, enabling audio AI systems that can robustly adapt to dynamic environments where categories may appear or disappear, such as in smart home assistants monitoring evolving soundscapes or industrial audio monitoring with changing equipment states.
- FCIAC method handles both class additions and removals, unlike prior incremental learning approaches.
- Uses a class-variable prototype adaptation network that dynamically restructures the classifier.
- Pseudo class-variable training improves model adaptability, leading to higher accuracy on three datasets.
Why It Matters
Enables audio AI to adapt to real-world class changes with minimal examples.