Rapid signer adaptation for continuous sign language recognition using a combined approach of eigenvoices, MLLR, and MAP

Current sign language recognition systems are still designed for signer-dependent operation only and thus suffer from the problem of interpersonal variability in production. Applied to signer-independent tasks, they show poor performance even when increasing the number of training signers. Better results can be achieved with dedicated adaptation methods. In this paper, we describe a vision-based recognition system that quickly adapts to new signers. For rapid signer adaptation it employs a combined approach of eigenvoices, maximum likelihood linear regression, and maximum a posteriori estimation. An extensive evaluation was performed on a large sign language corpus, that contains continuous articulations of 25 native signers. The proposed adaptation approach significantly increases accuracy even with a small amount of adaptation data. Supervised adaptation with only 10 adaptation utterances yields a recognition accuracy of 75.8%, which is a relative error rate reduction of 30.2% compared to the signer-independent baseline.