Multimodal Biometric Fusion - Joint Typist (Keystroke) and Speaker Verification

Identity verification through fusion of features from keystroke dynamics and speech is addressed in this paper. We experimentally compare performances from identity verification algorithms based on fusion of median pitch (prosodic level information), mel-frequency cepstral coefficients (short-time spectral level information) and keystroke dynamics (down-down intervals). These experimental results are also compared to the corresponding unimodal performances. All experiments are carried out on a small but true multimodal public database. Three fusion strategies are experimentally compared. A nonnegligeable performance improvement is then observed, mainly with a simple linear fusion strategy based on fusion of estimates.