September 16, 2021
By Pavel Jiřík in Blog
This summer, for the first time ever, our Phonexia researchers participated in the VoxCeleb Speaker Recognition Challenge 2021 (VoxSRC-21) to test their latest approaches for developing cutting-edge speaker verification systems.
The VoxCeleb Speaker Recognition Challenge 2021 contained four separate tracks, each focusing on a different speaker recognition challenge.
Our Phonexia team of top voice biometrics researchers—Josef Slavíček, Albert Swart, Michal Klčo, and Niko Brümmer—decided to tackle the VoxSRC-21’s toughest speaker verification challenge outlined in Track 3: Self-supervised speaker verification.
Unlike in tracks 1 and 2, this challenge was to create the most accurate speaker verification system based on the VoxCeleb2 dev dataset without using its speaker labels.
In other words, the creation of a speaker verification system was made much harder due to missing information about the speakers (who is who) in the training dataset, forcing such an unsupervised system to rely only on its ability to distinguish between the speakers by itself.
The Phonexia researchers explored various approaches and ended up creating a Deep Neural Network-based speaker verification solution.
The Phonexia Team's Results
Out of the 42 teams participating in this particular VoxSRC-21 challenge, only 12 teams submitted their final results by September 4. Out of these teams, the Phonexia research team achieved the second-best results, winning themselves second place in the Track 3 challenge with an Equal Error Rate (EER) of 6.486%.
Interestingly, the solution of the team that won first place in Track 3 was based on the combination of both the audio and video data available in the training dataset.
Therefore, the Phonexia researchers not only won second place in the Track 3 challenge, but they also became the best among all the teams in Track 3 that attempted to develop a self-supervised speaker verification system based purely on audio data.
As Phonexia’s prime focus is voice biometrics and speech recognition technologies, the above result is an exceptional acheivement for our researchers.
If you are interested in seeing a further explanation of the approach that our researchers used during the VoxSRC-21 challenge, watch the following video where Niko Brümmer explains it in greater detail:
You can also read a full report about the Phonexia VoxCeleb Speaker Recognition Challenge 2021 solution.
Once again, this is an outstanding achievement for our researchers, and we at Phonexia are looking forward to leveraging their latest know-how to push our cutting-edge voice recognition solutions to new limits.