Analysis of the Second Phase of the 2013-2014 i-Vector Machine Learning Challenge
Desire Banse, G R. Doddington, Craig S. Greenberg, John M. Howard, Alvin F. Martin, Daniel Garcia-Romero, John J. Godfrey, Jaime Hernandez-Cordero, Lisa Mason, Alan McCree, Douglas A. Reynolds
In late 2013 and 2014, the National Institute of Standards and Technology (NIST) coordinated an i-vector challenge utilizing data from previous NIST Speaker Recognition Evaluations. Following the evaluation period, a second phase of the challenge was held, where speaker labels were made available for system development. The second phase included system submissions from 23 participants representing 13 different countries, of which 18 also participated in the first phase of the challenge. The top 10 systems participating in both of the challenge phases demonstrated an average relative improvement of approximately 26% between the first and second phases, which represents the value of having access to the speaker labels. The top five participants submitted a system that outperformed the oracle system from the first phase on the evaluation data.