Task: Given a speech recording, classify it as real (bonafide) or fake (synthetic) under unknown media transformations.

RADAR Challenge 2026 Logo

Robust Audio Deepfake Recognition under Media Transformations

The RADAR Challenge 2026 is an APSIPA Grand Challenge on robust audio deepfake detection under realistic media conditions. Participants must determine whether a speech recording is bonafide (real) or spoofed (synthetic) after it has undergone unknown media transformations such as codec compression, resampling, background noise, music mixing, and room effects.

While recent advances in speech synthesis and voice conversion have made synthetic speech increasingly realistic, most existing detection benchmarks evaluate systems using clean audio. In practice, however, audio shared through social media, messaging platforms, and online media rarely remains pristine. Instead, it is typically edited, compressed, resampled, or mixed with other sounds before reaching detection systems.

The RADAR Challenge addresses this gap by evaluating detection models under realistic media processing pipelines. By introducing diverse and partially unseen transformations, the challenge emphasizes robustness and generalization, which are critical for real-world deployment.

RADAR aims to establish a benchmark for media-robust audio deepfake detection and encourage the development of detection systems that remain reliable beyond controlled laboratory conditions.

Venue

RADAR Challenge 2026 is organized as an APSIPA Grand Challenge and will be presented at APSIPA ASC 2026 in Hanoi, Vietnam.

Quick Updates

Competition Results

These are preliminary results. Official results will be announced at APSIPA ASC 2026 in Hanoi, Vietnam.

Team ID Dev EER% Eval EER% Team ID Dev EER% Eval EER% Team ID Dev EER% Eval EER%
A 1.27 5.67 N 29.77 14.92 27 37.75 —
B 4.46 9.05 O 30.20 — 28 38.82 —
C 4.63 5.10 P 31.62 20.23 29 40.96 36.92
D 5.26 17.14 Q 32.03 53.58 30 41.36 —
E 8.22 32.76 R 33.59 — 31 41.99 —
F 9.87 — S 34.44 17.34 32 47.37 —
G 11.04 10.14 T 35.16 27.01 33 71.72 —
H 11.12 18.12 U 35.37 32.39      
I 14.90 22.50 V 35.50 —      
J 16.03 8.04 W 35.81 —      
K 16.35 20.63 X 36.07 —      
L 19.58 23.43 Y 36.19 42.02      
M 20.27 16.47 Z (baseline) 37.71 42.60      

Highlighted cells are the five lowest Eval EER% values among teams with an evaluation score (lower is better). Values match the Codabench leaderboard format.

Challenge Papers

Paper Submission

Teams that had access to the RADAR2026 development set and its labels could submit a system description to APSIPA 2026. The baseline system includes evaluation and analysis scripts for reproducibility and for drafting challenge papers.

Paper submission deadline: May 15, 2026 (closed).

When submitting, authors selected the “Grand Challenge: RADAR - Robust Audio Deepfake Recognition under Media Transformations” track, which places more emphasis on technical correctness than on novelty.

Competition Portal

We use Codabench for submissions and the leaderboard:

Please register on Codabench as early as possible, as there is a daily limit on the number of submissions.

Note that the submission deadlines and daily submission limits reset according to UTC time.

Only one member per team will be approved on Codabench.

Registration (Closed)

Please register your team using the registration form:

Registration deadline: April 10, 2026. By registering, you agree to the Terms & Conditions.

We welcome both academic and industry teams. Individual researchers are also encouraged to participate.

Timeline

Contact

For questions please use GitHub Discussion or via email (radarchallenge2026 (at) gmail.com)

We strongly encourage participants to use GitHub Discussions so answers benefit all teams.

How to Get Started

  1. Read the rules of the challenge.
  2. Prepare your data pipeline and models using publicly available training datasets (respecting the training data policy).
  3. Download the development set (sent to participants via email) once released and evaluate your systems under the provided transformations.
  4. Use the baseline system and inference script as a reference for data format and scoring.
  5. Submit your scores and system description according to the challenge instructions and deadlines.

Task Summary

Item Description
Task Binary classification (bonafide vs spoof)
Input Speech waveform
Twist Speech is passed through unknown media processing chains
Output Spoof detection score
Metric Equal Error Rate (EER)
Training data Open (public datasets only)
Evaluation Blind test set
Goal Robust detection systems for realistic media processing

Dataset & Protocol

Phases

Phase 1 - Development

The development phase is intended to help participants validate their systems and familiarize themselves with the evaluation protocol. Results from Phase 1 (Development) will not affect the final ranking. The top 5 teams in the Development Phase will receive honorary mentions and certificates.

Phase 2 - Evaluation

To be eligible for Phase 2, teams must have at least one Phase 1 submission with an EER below 37%.

Results from Phase 2 (Evaluation) will determine the final ranking. Top-performing teams will be recognized at APSIPA ASC 2026 in Hanoi, Vietnam.

Media Transformations

Examples of media transformation

  Original Transformed
Bonafide
Spoofed

Each utterance in the development and evaluation sets may undergo one or more of the following while maintained their original label (bonafide/spoofed):

Some transformations used in the evaluation set will not appear in the development set in order to explicitly evaluate model generalization. The number and combination of transformations applied to each utterance are randomly sampled to emulate diverse real-world media processing pipelines. Additional undisclosed transformations with similar characteristics may also be included in the evaluation set to further assess robustness to unseen conditions.

Evaluation

Primary metric: Equal Error Rate (EER) on the evaluation set.

Participants must submit one detection score per evaluation utterance (higher scores indicate higher confidence that the sample is spoofed). Leaderboard rankings are based on EER, with additional metrics possibly reported for further analysis.

EER is selected due to its simplicity and widespread use in spoofing detection research. While participants are free to interpret their results based on the leaderboard, the organizers do not endorse or take responsibility for any performance claims made based on these results.

Baseline systems

We released a baseline system to help participants get started and to illustrate the expected data pipeline and submission format.

Organizers could also submit scores and description papers for guidance; those organizer entries were excluded from the final ranking.

01. SSL AASIST Antispoofing

This baseline demonstrates the expected submission format and includes offline evaluation and analysis scripts.

Awards

This is an academic challenge without monetary prizes.

Top-performing teams will receive certificates, and outstanding submissions will be recognized at APSIPA 2026 in Hanoi, Vietnam.

Organizers

Dr. Hieu-Thi Luong

Dr. Hieu-Thi Luong
Fortemedia, Singapore
Challenge Lead Organizer

Asst. Prof. Xuechen Liu

Asst. Prof. Xuechen Liu
Xi'an Jiaotong-Liverpool University, China
Co-organizer

Dr. Ivan Kukanov

Dr. Ivan Kukanov
KLASS Engineering & Solutions, Singapore
Co-organizer

Assoc. Prof. Kong-Aik Lee

Assoc. Prof. Kong-Aik Lee
The Hong Kong Polytechnic University, Hong Kong SAR, China
Advisor

FAQs

Yes. Publicly available pretrained models (Wave2Vec, WavLM, XLSR, etc.) are allowed, even if they were trained on large public speech datasets.

However, participants must not explicitly train or fine-tune models using LibriTTS dev/test splits or any derivatives of the RADAR development or evaluation data for the task of audio deepfake detection.

Q: Can I use [this dataset] for training?

We only restrict the use of LlamaPartialSpoof, LibriTTS dev/test sets, and their derivatives.

We recognize that some datasets (especially those derived from LibriSpeech/LibriTTS) may contain overlapping speakers or recordings. To avoid unnecessary restrictions, we explicitly allow the use of the following datasets for training, even if partial overlap may exist:

Participants may use these datasets with confidence for training their systems.