Lip-Movement Based Speaker Recognition Focused on the Distributed Structure of Lip-Movement Data

A speaker recognition method is presented that uses the distributed structure of lip-movement data. It overcomes the degradation in accuracy of a previous method based on the kernel mutual subspace method due to the genuine samples being close to those of other persons. The degradation in accuracy results from the strong nonlinearity of the distribution. This degradation is reduced by increasing the weights of the samples near the cluster centers. Evaluation of the proposed method demonstrated that it outperforms other person authentication methods based on lip movement.