Few-Shot Keyword-Incremental Learning with Total Calibration

Abstract

Keyword spotting (KWS) models need to continuously recognize new keywords for user demand. However, two significant challenges exist in satisfying this requirement: catastrophic forgetting, where the model loses its ability to classify previously learned keywords, and insufficient data for new classes. To address these challenges, we propose a Few-shot keyword-Incremental Learning with total caLibration (FILL), a novel few-shot class-incremental learning (FSCIL) approach for KWS. FSCIL trains a model with sufficient data in an initial session, followed by incremental sessions where it learns new classes with limited data. FILL employs prototype calibration throughout total sessions to enhance class separation and mitigate misclassification. Notably, it utilizes manifold mixup in the initial session to generate new classes for prototype calibration. Experimental results on two KWS datasets demonstrate that FILL outperforms three baselines in terms of average accuracy.