Call for Participation

The Radar Acoustic Speech Enhancement (RASE) Challenge 2026 is an official ICASSP Grand Challenge that invites researchers and practitioners in speech processing, machine learning, and signal enhancement to tackle a novel task: reconstructing intelligible, full-bandwidth speech from degraded signals captured using millimeter-wave (mmWave) radar — even through glass walls.

Conventional microphones struggle in noisy or occluded environments. mmWave FMCW radar offers a non-contact alternative by capturing surface vibrations induced by speech. However, radar-captured signals are often band-limited and noisy. This challenge aims to bridge the gap between radar sensing and high-quality speech reconstruction.

To know more about the problem statement, go to this page: Problem Statement

Dataset

Participants will receive a curated dataset comprising paired radar-captured and microphone-recorded speech. The data is collected using a TI AWR2243BOOST mmWave FMCW radar through a glass-wall.

Data Scenarios

Direct diaphragm vibration (Simple case)
Radar captures strong vibrations directly from a loudspeaker diaphragm.
Secondary surface vibration (Challenging case)
Vibration captured from an aluminium foil surface placed near the loudspeaker.

Notes

All recordings are provided in .wav format.
Raw radar signals are not shared — no radar signal processing expertise is needed.
The dataset is split into training, development, and test sets.
Baseline code in Python will be provided.
Registered participants will be granted access.

📦 Release date: September 20, 2025 ~~September 07, 2025~~

Evaluation

Submissions will be evaluated using four standard metrics:

PESQ – Perceptual Evaluation of Speech Quality
ESTOI – Extended Short-Time Objective Intelligibility
DNSMOS – Deep Noise Suppression Mean Opinion Score
MFCC Cosine Similarity

Difficulty-weighted scoring:

Easy: 40%
Hard: 60%

Top‑5 teams will be invited to present their work at ICASSP 2026.

Baseline

Baseline deep learning models implemented in PyTorch will be provided to help participants kickstart their solutions. These include:

Spectral mapping networks
Phase-aware enhancement techniques
Feature extraction and post-processing tools

📦 Release date: September 20, 2025 GitHub Repository

Participation Guidelines

To ensure fairness and reproducibility:

Use only the provided dataset — external data or augmentation is not allowed.
Train models from scratch (no pre-trained or fine-tuned models).
A maximum of 300 training epochs is allowed.
Submit code, trained weights, and environment (Docker or requirements file).
Participants retain all intellectual property rights.

Paper Submission Guidelines

Papers must follow the official ICASSP paper template (including ICASSP LaTeX style files or Word template).
Maximum length: 2 pages of technical content + 1 additional page for references only (total maximum 3 pages).
The reference page should contain only references (no technical content, figures, or tables allowed on the third page).

Timeline

📅 Registration Starts: ~~September 07, 2025~~ September 20, 2025
📂 Data & Baseline Release: ~~September 07, 2025~~ September 20, 2025
🔎 Validation Submission Phase Starts: ~~September 27, 2025~~ October 5, 2025
🚀 Test Submission Phase Starts: November 14, 2025
🧪 Submission Deadline: ~~November 20, 2025~~ November 22, 2025, 23:59 UTC+0
🏆 Ranking Announcement: ~~November 24, 2025~~ November 26, 2025 VIEW RESULTS
✍️ 2-page Papers (by invite): ~~December 07, 2025~~ January 07, 2026
✅ Acceptance Notification: ~~January 11, 2026~~ January 21, 2026
📌 Camera-ready Deadline: ~~January 18, 2026~~ January 28, 2026