Scalability analysis of audio-visual person identity verification

In this work, we present a multimodal identity verification system based on the fusion of the face image and the text independent speech data of a person. The system conciliates the monomodal face and speaker verification algorithms by fusing their respective scores. In order to assess the authentication system at different scales, the performance is evaluated at various sizes of the face and speech user template. The user template size is a key parameter when the storage space is limited like in a smart card. Our experimental results show that the multimodal fusion allows to reduce significantly the user template size while keeping a satisfactory level of performance. Experiments are performed on the newly recorded multimodal database BANCA