Using a predefined passphrase to evaluate a speaker verification system

摘要

This article presents a standardized and repeatable process used to evaluate the performance of a speaker verification system. Through the use of a common passphrase and a subset of extracted feature vectors that outperforms other combinations, the study limits the exposure to potential experimental flaws, while measuring true biometric performance more effectively than existing evaluation methodologies. After collecting a dataset of 33 participants, the researchers achieved a performance rate of 99.8% for the 22 users who contributed at least 20 text-dependent samples. The primary focus of the research, however, was to illustrate a variety of testing techniques that can be used to efficiently analyze the performance of a speaker verification system and advocate the use of a common passphrase in this process.