Feature reduction based transfer structural subspace learning for small-footprint cross-domain keyword spotting via linear discriminant analysis

Abstract

Small-footprint keyword spotting has received considerable attention in recent years, which is often conducted on the assumption that the predefined keywords in training and testing data are obtained under the same condition. However, in practical situations, this assumption does not hold due to the widespread existence of various speech scenarios or voice-embedded devices. To tackle this problem, in this paper, we propose a new transfer subspace learning method called feature reduction based transfer structural subspace learning (FRTSSL) for small-footprint cross-domain keyword spotting. FRTSSL aims to learn a domain-invariant and discriminative subspace by which (1) feature reduction is used to high dimensional features to avoid unnecessary computation; and (2) transfer structural subspace learning jointly exploits the statistical properties and geometric structure to reduce the distribution discrepancy of the source and target domains based on a joint linear discriminant analysis (LDA) framework; and (3) the feedback term is constructed for improving the discrimination of the subspace based on source labels and pseudo-target labels. To preserve the intrinsic geometric structure of samples in the projection subspace, we first preserve the global subspace structure by imposing the reconstruction constraints on the reconstruction coefficient matrix, and then we preserve the space relationship of samples using a graph regularization method. Furthermore, we formulate a minimization problem that integrates marginal and conditional distribution alignment, reconstruction constraints, and graph regularization into the joint LDA framework, giving an effective optimization algorithm. Experimental results on four cross-domain keyword datasets show that our method outperforms some state-of-the-art conventional transfer learning methods and no transfer learning methods.