Selective transfer subspace learning for small-footprint end-to-end cross-domain keyword spotting

Fei Ma, Chengliang Wang, Xusheng Li, Zhuo Zeng

Abstract: Highlights•STSL is proposed for improving the accuracy of CDKWS in various speech scenarios.•STSL avoids negative transfer by actively selecting source samples for TL.•STSL aligns TL and active selection to learn a domain-invariant projection subspace.•STSL outperforms some transfer learning methods and no transfer learning methods.