A Speaker Verification Biometric In 40 Bytes

This paper describes a novel Speaker Verification (SV) architecture that minimises many of the challenging problems faced by SV implementors in the Card Security arena. The architecture is unusual in two respects. First, the verification engine is not fed with conventional frequency-domain descriptors. Instead the output of a Time-Encoded Signal Processing and Recognition (TESPAR) coder is used. The TESPAR coder is a low-complexity, vector quantiser capable of converting the speech waveform into a highly informative set of time-domain descriptors. A simple 29-element matrix representation of TESPAR descriptors is described which, when used in conjunction with a Fast Artificial Neural Network (FANN), provides a wide degree of cross-speaker separation and strong same-speaker consistency [1]. The paper focuses on a novel SV engine that secures a robust verification performance by encoding the classification power of 100 orthogonal artificial neural networks into a biometric of only 40 bytes, offering a high level of protection to low-tech Dumb Cards as well as to state-of-the-art Smart-Card embodiments. The architectural rationale that enables this is presented, together with indications that a final verification decision can be delivered in under 1 second. The key features of the TESPAR/FANN combination and a system implementation in silicon for Smart-Card and Dumb-Card based systems are also discussed.