The face and voice of a person have distinctive characteristics and are widely used as biometric cues for person authentication, either independently or jointly in multimodal systems. The strong perceptual correspondence humans establish between faces and voices has motivated a large body of work on automated face-voice association. Most of these approaches, however, are evaluated under controlled experimental settings that neglect the varying conditions of real-world scenarios. Two important sources of this variation are the language spoken by the speakers and the distribution of demographic traits across speakers. Since a substantial proportion of the global population is bilingual or multilingual, it matters whether face-voice association models remain reliable when the same speaker communicates in different languages. At the same time, gender-controlled evaluation is necessary to determine whether models rely on speaker-specific cross-modal cues rather than simply exploiting demographic differences between the speakers forming the positive and negative pairs.

The goal of the Face-voice Association across LAnguages and Gender (FLAG) 2027 Challenge is therefore two-fold: to evaluate face-voice association methods in multilingual conditions while preventing models from using gender as a proxy for identity, and to foster the development of multilingual and gender-aware methods that learn identity-specific associations. FLAG 2027 builds on the prior FAME 2026 and FAME 2024 Grand Challenges hosted at IEEE ICASSP and ACM Multimedia, and has been accepted for support under the IEEE Signal Processing Society Challenge Program.

EVALUATION TRACKS

Language impact

Face-voice association is established through a cross-modal verification task: given a face and a voice, the system decides whether both belong to the same identity. The dataset is split following the unseen-unheard configuration, in which train and test splits consist of disjoint speakers. Each video shows a speaker using one language only, while every speaker appears in videos of at least two distinct languages. Evaluation is carried out on both a heard language, available during training, and an unheard language, not available during training. This track measures whether a voice recorded in a different language can still be associated with the corresponding speaker's face.

Gender impact

This track evaluates gender-controlled verification. Negative samples are restricted to speakers of the same gender as the speaker of the positive sample, so a model can no longer infer identity from demographic differences alone. This yields a more rigorous assessment of whether models learn genuinely identity-specific face-voice associations rather than using gender as a shortcut.

DATASET

Building on the prior FAME 2024 and 2026 Grand Challenges, we extended the MAV-Celeb dataset with one additional language, curating a new split of 100 English-Bengali speakers (MAV-Celeb v4). The new split provides both language and gender annotations, which is what makes the two FLAG tracks possible. Alongside it, the earlier splits remain available: MAV-Celeb v1 covers English and Urdu speakers, and v3 covers English and German speakers, each with distinct (non-overlapping) speaker identities.

As in previous releases, the audio-visual samples are obtained from YouTube videos of celebrities appearing in interviews, talk shows, and television debates. The visual data spans a vast range of setups, including different poses, motion blur, background clutter, video quality, occlusions and lighting conditions. Because the videos originate from real-world situations, they reproduce the same difficulties encountered when deploying face-voice association tools in practice: noise, background chatter or music, overlapping voices, and compression artifacts.

The training and validation sets are shared with participants on the first day of the progress phase. The input data for the final evaluation phase is released on the first day of the evaluation phase, without ground-truth labels, so that the final evaluation tests the generalization capability of models developed during the progress phase.

The dataset is available at the following links:

• MAV-Celeb v4 (English-Bengali, FLAG 2027)
  • Raw files link — released on the first day of the progress phase
  • Train/Test Split — released on the first day of the progress phase

BASELINE MODEL & STARTER KIT

To allow participants to benchmark their results, we release a pretrained instance of a competitive multimodal method for the face-voice association task, together with a leaderboard on CodaBench where teams can compare their performance during the progress phase. The model is a two-branch network taking face and voice embeddings as input (see Fig. 3). Face embeddings are obtained with a popular convolutional network pretrained on a large-scale facial recognition dataset; voice embeddings are obtained with an audio encoding network for speaker recognition, trained on the language available in the training set (i.e. the heard language). The multimodal model then combines the face and voice embeddings and is optimized with a loss function that imposes orthogonality constraints on the multimodal embeddings of different speakers.

Link to the paper: Fusion and Orthogonal Projection for Improved Face-Voice Association
Link to the Paper's code: https://github.com/msaadsaeed/FOP
Link to the Baseline code: https://github.com/mavceleb/mavceleb_baseline

EVALUATION METRICS

As commonly done for face-voice association tasks, submissions are evaluated with the Equal Error Rate (EER): the value at which the false acceptance rate (FAR) and the false rejection rate (FRR) are equal. As for FAR and FRR, a low EER indicates good performance. Participants submit a .txt file containing output scores for every pair in the test set, indicating the system's confidence that the face and voice belong to the same person. Within a submission file, a pair with a higher score is interpreted as one the model is more confident about.

This setup is consistent with the evaluation of the earlier FAME 2024 and FAME 2026 Grand Challenges. Technically, EER does not require confidence scores of different models to span the same range, and it does not depend on a fixed threshold, unlike metrics such as precision or accuracy. In real deployments a system developer would tune that threshold to their own needs — a high threshold gives a low FAR and a high FRR — which is precisely why a threshold-independent metric is the fairer basis for comparing systems.

Since models are evaluated on different tasks, each model yields several EERs. The overall score used to decide the challenge winners is computed as:

Overall Score = (Sum of all EERs) / 2

SUBMISSION

Alongside the audio (.wav) and image (.jpg) files, the dataset includes .txt files of the face-voice test pairs, where each line has the following format:

• ysuvkz41 voices/English/00000.wav faces/English/00000.jpg
• tog3zj45 voices/English/00001.wav faces/English/00001.jpg
• ky5xfj1d voices/English/00002.wav faces/English/00002.jpg
• yx4nfa35 voices/English/01062.wav faces/English/01062.jpg
• bowsaf5e voices/English/01063.wav faces/English/01063.jpg

The first entry of each line is the ID of the pair; the remaining two are the local paths of the corresponding audio and of a visual frame extracted from the video. The ground-truth files contain the ID of the pair and a 1 or a 0, depending on whether the face and voice correspond to the same speaker. Ground truth is withheld for fair evaluation during the challenge.

Participants are expected to compute and submit text files including the ID and the confidence score, in the following format:

• ysuvkz41 0.9988
• tog3zj45 0.1146
• ky5xfj1d 0.6514
• yx4nfa35 1.5321
• bowsaf5e 1.6578

Within the directory containing the submission files, use zip archive.zip *.txt and do not zip the folder. The exact file names expected for each evaluation condition are announced on the CodaBench page at the start of each phase.

In the progress phase, each team may make a maximum of 150 submissions, with at most 15 per day. In the evaluation phase, the total number of submissions is limited to 15.

Link to CodaBench: to be announced — the FLAG 2027 competition page will be linked here before the progress phase opens.

RULES FOR SYSTEM DEVELOPMENT

Since FLAG 2027 aims at analyzing whether face-voice association capabilities translate across language and gender constraints, the following rules apply:

• A pretrained encoder for faces or voices is allowed.
• Participants are required to submit a 2-page system description in the ICASSP template to the challenge organizers. Teams without a system description will be disqualified from the challenge. Teams describing a setup that violates one of these rules will be disqualified.
• Participants are required to submit a link to a working version of their setup, e.g. on a platform for open-source development such as GitHub. Teams without a code submission, or with a setup that violates one of these rules, will be disqualified.

REGISTRATION

We welcome participants to apply for the “FLAG Challenge 2027” by expressing their interest via google forms at to be announced — the registration form will be linked here at the start of the registration period.

For any queries please contact us at our email mavceleb@gmail.com.

TIMELINE

Tentative — dates may be updated.

• Registration Period: 24 August – 7 September 2026
• Progress Phase: 1 September – 15 October 2026
• Evaluation Phase: 15 October – 21 October 2026
• Challenge Results: 27 October 2026
• Submission of System Descriptions: 30 October 2026
• Challenge Paper Submission: 07 December 2026

ORGANIZERS