ECHO Challenge

The ECHO (Ethical Communication for Human Outreach) Challenge 2026: Multilingual Multimodal Hate Speech Detection

2026 IEEE International Conference on Multimedia and Expo (ICME 2026), 5-9 July 2026, Bangkok.

Training and Validation Dataset released!!!!

Introduction

Automated hate speech detection has become a critical component of digital safety infrastructure, yet the overwhelming majority of existing research and deployed systems are built for English and text-only environments. The reality is that large segments of the world communicate online not through typed text but through speech: voice messages, livestream interactions, podcasts, short-form video, community radio, and spoken social platforms. In these contexts, hate speech is not only present, but often amplified by prosodic cues, tone, emphasis, sarcasm, and code-mixed language elements that cannot be captured by text-only analysis. Despite the scale and urgency of this problem, there exists no established multilingual speech benchmark for hate speech detection that spans low-resource and high-resource languages. As a result, the speech community lacks a shared foundation on which to build safe, fair, and inclusive speech moderation systems. The ECHO 2026 Challenge is designed to fill this gap. It introduces the first cross-lingual hate speech detection benchmark built from raw speech, aligned ASR transcripts, and carefully validated annotations across Tamil, Bengali, Hindi, Malayalam, Telugu, French, Spanish, English, Chinese, and Japanese, representing more than 3 billion speakers worldwide. This challenge offers not only a technically ambitious task, but a socially urgent benchmark that directly connects speech technology with public safety, platform accountability, and linguistic equity.

Timeline

January 10, 2026 - Registration Starts
January 27, 2026 – Training and development sets released
February 10, 2026 – Test data released
February 25, 2026 – System submissions due
March 5, 2026 – Leaderboard opens

Registration Link

Awards for ECHO Challenge

1)The Top 5 teams will be invited to contribute to the competition summary paper

2) Prize money TBD

Competition Outline

Challenge Structure and Rules

Participants may choose to compete in one or more of the following three tasks:

Task 1: Audio-Based Classification
- Detects hate speech directly from raw speech audio.
- Models must rely solely on acoustic and prosodic features; textual transcripts are not permitted.
Task 2: Text-Based Classification
- Detect hate speech from ASR-generated transcripts only.
- No audio data may be used in this task.
Task 3: Transcript-Based (Multimodal) Classification
- Use both speech and text features jointly to classify hate and non-hate content.
- Participants are encouraged to explore multimodal fusion and cross-representation learning approaches.

Data and Submission Guidelines

Only the test data (without labels) will be released publicly.
Participants must train and validate models using the provided training and validation splits.
For each task, participants must submit a CSV file containing model predictions for the test set.

Submissions that include extra or missing samples, or deviate from the prescribed format, will be disqualified.
Each team may submit up to three entries per task; only the best-performing submission per team will appear on the leaderboard.

General Rules

Each team must consist of at most 5 team members.
Each team may submit up to three entries per task; only the best-performing submission will be shown on the leaderboard.
All models must be accompanied by a brief model card describing:
- Model architecture and features used
- Training configuration
- Ethical considerations and bias mitigation steps
Any form of data leakage, use of pretrained or external models, or augmentation from outside sources will result in immediate disqualification.
Use of data or models for surveillance, censorship, or discriminatory applications is strictly prohibited.

For queries and suggestions, contact us at: team.echo2026@gmail.com