A computationally scalable speaker recognition system

Computationally scalable speaker recognition systems are highly desirable in practice. To achieve this objective, we use a two-stage architecture for text-prompted speaker recognition. In this system, the input speech is first segmented on subword boundaries using a Viterbi alignment. The second stage applies a polynomial classifier to each subword for verification. Through a simple approximation, the scoring criterion for the polynomial classifier is made highly scalable. The resulting combination of speaker independent segmentation and a scalable recognition system results in a system which can perform speaker recognition on a large population with minimal computation.