Multi-space random mapping for speaker identification

This paper presents a work of utilizing multi-space random mapping (MRM) to formulate a dual-factor identification system, which combines speaker biometric and personal token. Personal token will be assigned to the client to constitute a unique random subspace during enrollment and the speaker template will be generated within the random subspace. Test features will be mapped to the random subspace that is described by the personal token presented during identification. Our work has shown that MRM-system exhibits stronger discriminative ability when comparing test features to its counterfeit templates, which lied in other different random subspaces. This advantage thus contributes to better F-ratio and greater accuracy recognition. Experiments on YOHO corpus demonstrate a remarkable result where the system achieves the perfect identification rate.