The Inaugural Music Source Restoration Challenge

ICASSP 2026 Grand Challenge • Pioneering the Future of Audio Technology

Call for Participation

ICASSP 2026 Grand Challenge: We invite researchers, engineers, and audio enthusiasts worldwide to participate in the first-ever Music Source Restoration (MSR) Challenge. We target on restoring individual instruments to their original, unprocessed state from mixtures that have undergone various audio transformations.

This challenging scenario differs from traditional source separation where mixtures are assumed to be simple sums of unprocessed sources. MSR's requirement for generative solutions is crucial for real-world audio production including:

  • Stem-level re-production and remixing of existing recordings
  • Recovery of degraded historical recordings
  • Professional restoration of live recordings affected by venue acoustics
  • High-quality remixing of recordings made with limited equipment

Timeline

August 15, 2025 Registration Opens
September 1, 2025 Validation Set Release
September 15, 2025 Baseline System Release
October 28, 2025 Data Submission Deadline
November 18, 2025 Test Set Release
November 20, 2025 Final Submission Deadline
November 27, 2025 Results Announcement
December 7, 2025 2-page Paper Due (By Invite)

Organizers

Industry

  • Yongyi Zang (Smule, USA)
  • Yuki Mitsufuji (Sony AI, Japan)

Academia

  • Jiarui Hai & Helin Wang (Johns Hopkins University, USA)
  • Wanying Ge (National Institute of Informatics, Japan)
  • Zheqi Dai & Qiuqiang Kong (CUHK, Hong Kong SAR)
  • Mark Plumbley (University of Surrey, UK)

Challenge Overview

Dataset

Participants are encouraged to innovatively use open-source data or propose new data sources, including synthetic data generation methods. Reference pipeline scripts will be provided. To ensure fairness, all newly created data and pipelines must be submitted and shared by October 28.

Evaluation Protocol

1

Reconstruction-Dominant

Emphasizes accurate recovery of original stem waveforms with focus on signal restoration

2

Semantic Alignment

Prioritizes perceptual similarity and musical coherence

3

Subjective Quality

Based on blind subjective ratings from professional engineers

Target Stems

Eight target stems will be evaluated: vocals, guitars, keyboards, bass, synthesizers, drums, percussion, and orchestral elements.

Real-World Scenarios

The blind test set includes 500 clips from four challenging scenarios:

  1. Historical Recordings from UCSD Cylinder Audio Archive
  2. Live Recordings from YouTube with venue acoustics
  3. FM Radio Broadcasts with analog transmission artifacts
  4. Lossy Streaming with codec degradation

Guidelines for Participants

Submission Requirements

Baseline Systems

Complete training and evaluation code for U-Net and BSRNN baseline systems will be provided.

Join Us in Pioneering Audio Technology!

Be part of this groundbreaking ICASSP 2026 Grand Challenge that bridges academic research with professional audio engineering.

Register Now