2025 Music Transcription Challenge

Automatic Music Transcription (AMT) is the task of converting raw audio into a symbolic representation such as MIDI. Recent advances in neural architectures and pretraining have driven major improvements on piano-only recordings, but performance on multi-instrument classical music remains uneven. This competition challenges teams to push the state of the art on classical music transcription using a shared dataset and evaluation pipeline.

Hosted by ai4musicians.org in collaboration with the IEEE Technical Community on Multimedia Computing (TCMC) ↗.

Repository access

Each team receives access to a private fork of the competition repository. All submissions are pushed to a team-specific branch.

Submission branch

Push your solution to your team's branch. The CI pipeline runs your model against the held-out test set on the competition runner.

Environment configuration

Provide a self-contained environment (Docker image, conda env, or pinned requirements). The runner has no network access at evaluation time.

Model execution requirements

Models must run end-to-end on a single GPU under a fixed wall-clock budget. Wall-clock time is reported as “Runtime” on the leaderboard.

Input file naming convention

Test audio is provided as a sequence of WAV files; outputs must be MIDI files with matching basenames.

Submission process

Push to your team branch and tag the commit. Submissions are evaluated and the leaderboard is refreshed on a rolling basis.

Model weights

You may include checkpoints in your branch (via Git LFS) or download them at build time from a public host that is reachable at image-build time.

• Registration opens — January 1, 2025
• Submissions close — April 30, 2025
• Winners announced — at the ICME 2025 workshop (June 30, 2025)

Top-ranked teams receive cash prizes and an invited slot at the ICME 2025 AI for Music workshop. Final standings are computed from the F-measure (precision/recall on note onsets and offsets), with runtime as a tiebreaker.

Reference audio + MIDI pairs for participating teams to develop and validate their pipelines:

Sample Compositions Google Drive Folder ↗

See also a GM instrument numbering reference ↗ for output MIDI conventions.

The competition ships four reference baselines:

MT3 — Google Research

• github.com/ojas-chaturvedi/mt3-pytorch ↗ (PyTorch port for the competition runner)
• github.com/magenta/mt3 ↗ (original Magenta implementation)
• Paper: MT3: Multi-Task Multitrack Music Transcription ↗

Basic Pitch — Spotify

• github.com/KayshavBhardwaj/basic-pitch-test-1 ↗
• Paper: Basic Pitch: Spotify's Pitch Tracking Model ↗

ReconVAT

• github.com/ojas-chaturvedi/ReconVAT ↗
• Paper: ReconVAT: Semi-Supervised Learning for Music Transcription ↗

Composers interested in contributing to future iterations of the competition can reach out to the organizers using the contact information below. We welcome new pieces with associated reference MIDI for evaluation.