IBM model-based and frame-by-frame speaker-recognition

Abstract

Amidst the recent crazes for emerging technologies like speech recognition and biometrics, speaker recognition is slowly reaching the maturity to be deemed practical in many different applications. This paper presents new approaches for text-independent speaker recognition. The performances of the model-based algorithm presented concurrently at the ICASSP'98 conference and the frame-based algorithm presented in this paper are compared here. The engine, described here, provides multiple functionalities including those of identißcation, verification and classification. The modes of operation and design choices allow for tight integration of the speech recognition and speaker recognition engines in a broad sense. This new architecture as well as the results obtained for very specific tasks undoubtedly announce myriads of new applications where both technologies complement each other and can no longer be clearly distinguished as illustrated by the concept of speech biometrics. Hands-free and eyes-free human/machine transactions are moving a step further toward easier and more efficient interfaces and speech transactions are becoming more ubiquitous.