Defining an Enhanced Feature Subset to Assess the True Biometric Performance of Speaker Verification Systems
Jonathan Leet · DigitalCommons@Pace (Pace University) · 2015
The speaker verification system industry requires increased standardization in testing processes and evaluation methodology, specifically those that select the best combination of feature vectors (or enhanced feature subset) to use as input for analysis in these systems. There are numerous methods currently used to “benchmark” these offerings, which translates to a marketplace of solutions that cannot be compared against each other on equal terms. Organizations like NIST have also failed to standardize these evaluation methodologies, which is necessary to create an environment that fosters reliable performance metrics. These inconsistencies indicate that the current research is flawed and unreliable. This dissertation will demonstrate a reusable methodology to identify the best subset of feature information to isolate in the measurement of the actual biometric performance of a speaker verification system with the expressed intent of standardizing the manner that testing is conducted.