Challenges & benchmarks
I founded and lead a sustained international challenge series on face–voice association. It directs a ten-person organising team across six institutions and five countries, and covers dataset curation, evaluation-protocol design, baselines and community infrastructure.
Editions
FLAG 2027 — Face-voice Association across Languages and Gender
Introduces gender-controlled evaluation, designed to test whether models learn identity-specific structure rather than demographic shortcuts. Competitively selected for IEEE Signal Processing Society Challenge Program support.
FAME 2026
Second IEEE edition of the multilingual face–voice association benchmark, built on the MAV-Celeb data line.
FAME 2024 — Associating Faces with Voices in Multilingual Environments
The inaugural edition. Summarised in A Synopsis of FAME 2024 Challenge, ACM MM 2024.
Why gender-controlled evaluation
Face–voice association is easy to fake. A model that has merely learned to sort speakers by apparent gender scores well on an uncontrolled benchmark while learning nothing about identity. FLAG 2027 controls for gender within each evaluation trial, so a shortcut-driven model and an identity-driven model can be told apart. The same logic drives the multilingual protocol: a model tested only in the language it was trained on is never asked whether its representation is about the speaker or about the language.