The VoxCeleb Speaker Recognition Challenge 2019
(VoxSRC-19)

Welcome to the first VoxCeleb speaker recognition challenge! The goal of this challenge is to probe how well current methods can recognize speakers from speech obtained 'in the wild'. The dataset is obtained from YouTube videos of celebrity interviews, consisting of audio from both professionally edited and red carpet interviews as well as more casual conversational audio in which background noise, laughter, and other artefacts are observed in a range of recording environments.

The 2019 challenge is now over. Congratulations to all the winners! Details about the 2020 challenge can be found here.

A report summarising the challenge baselines, results and discussions can be found below. If you would like to reference the challenge mechanics or your performance in the challenge, please cite the report below. If you would like to cite the methods of the winners, please see if the authors have uploaded any papers describing their methods and be sure to give them the appropriate credit. Slides of the winners can be found at the workshop website.

            @article{chung2019voxsrc,
              title={VoxSRC 2019: The first VoxCeleb Speaker Recognition Challenge},
              author={Chung, Joon Son and Nagrani, Arsha and Coto, Ernesto and Xie, Weidi and McLaren, Mitchell and Reynolds, Douglas A and Zisserman, Andrew},
              journal={ISCA Challenges},
              year={2019}
            }
        

Tasks

VoxSRC 2019 consists of the following two tasks:

1. Audio only speaker verification - Fixed training data: This task requires that participants train only on the VoxCeleb2 dev dataset for which we have already released speaker verification labels. The dev dataset contains 1,092,009 utterances from 5,994 speakers.
2. Audio only speaker verification - Open training data: For the open training condition, participants can use the VoxCeleb datasets and any other data (including that which is not publicly released) except the challenge's test data

Fixed and open training conditions allow cross-system comparisons and an understanding of the effect of additional and unconstrained amounts of training data on system performance.

Evaluation

We have an evaluation server to allow teams to benchmark their performance on a blind test set over the course of (and even after) the evaluation period. In both training conditions, i.e. fixed and open, the test data can be used strictly for reporting of results alone - it cannot be used in any way to train or tune systems. Teams will only be allowed a limited number of submissions to prevent overfitting to the test set.

Results will be announced at the challenge workshop, where we will invite presentations from the most exciting and novel submissions, as well as from the challenge winners. The challenge workshop will be held on the 14th of September, 2019 in conjunction with Interspeech 2019 in Austria.

Further details about the evaluation can be found at the competition site of each tasks. See the registration section or the use the direct links below:

1. Evaluation - Audio only speaker verification - Fixed training data
2. Evaluation - Audio only speaker verification - Open training data