2025 AMT Challenge | Purdue SERIS
♪  AMT Challenge  ·  2025

2025 Automatic Music
Transcription Challenge

Develop advanced models capable of transcribing synthesized classical music into MIDI files.

Register Now → Competition Leaderboard
Submission window: April 1, 2025 – May 1, 2025

Summary

The 2025 Automatic Music Transcription (AMT) Challenge invites participants to develop computer programs capable of accurately transcribing synthesized audio recordings of classical music into MIDI files. Each submission will process 100 recordings, each up to 20 seconds long, within a maximum time limit of 4 hours.

The audio data has been synthesized to sound as realistic as possible, closely resembling natural instrumental performances. Unlike previous challenges, participants will be informed of the specific instruments present in each recording. Incorrect instrument identification will incur a penalty, with smaller penalties applied if the mistake involves similar instrument families.

Evaluation criteria include the accuracy of instrument identification, pitch, onset, offset, and dynamics.

An Online Competition

Technical Details

Participants will register on ai4musicians.org, where sample music files including scores and audio recordings will be provided to assist in model development. Contestants may use any public or proprietary data for training.

Submissions will be open in April 2025. Each team's program will be executed on a GPU-equipped system at Purdue's Rosen Center for Advanced Computing. Teams may submit models once every 24 hours, and a live leaderboard will display performance results. Final rankings will be determined using holdout data. Winning teams will be invited to present their solutions at the 2025 IEEE ICME Conference.

For questions and updates, join the AMT Slack Workspace →

Submission Details

Repository Access

During registration, participants must provide a link to their code repository along with a fine-grained access token for the competition backend to pull the model.

Submission Branch

Create a branch titled submission in your repository. The backend will automatically pull from this branch. Submissions are valid when:

  • The submission branch exists
  • New commits have been made since the last successful run

Environment Configuration

Include an environment.yml file in the root of your repository. This will be used to create a conda environment during execution.

Model Execution Requirements

The repository must contain a main.py file in the root directory accepting these arguments:

  • -i: Path to the input audio file (.mp3 format)
  • -o: Path to save the output MIDI file
python main.py -i input.mp3 -o output.midi

Input File Naming Convention

Audio file names include MIDI instrument codes following the General MIDI standard. Example:

1._0_40_70.mp3

Here 0, 40, and 70 are the MIDI instrument codes present in the recording. Incorrect instrument identification incurs a scoring penalty.

Model Weights

Participants may use Git LFS (Large File Storage) for managing model weights. The backend fully supports Git LFS.

Schedule

Nov 20, 2024
Release of ten sample compositions from each contributing composer (available to contestants)
Dec 15, 2024
Competition announcement
Dec 31, 2024
Release of ten additional compositions from each composer (not available to contestants)
Jan 31, 2025
Registration opens
Feb 1, 2025
Sample solution release
Apr 1, 2025
Submission window opens
May 1, 2025
Submission window closes
May 15, 2025
Winner announcement
Jun 2025
Presentation of winning solutions at the 2025 IEEE ICME conference

Cash Awards

First Place
$1,500
Second Place
$1,000
Third Place
$500

Winners must open-source their solutions as specified in the registration agreement. Cash awards will only be provided to participants from countries not subject to U.S. embargoes or sanctions. In some cases, awards may take the form of travel grants covering conference registration, hotel, and airfare.

Sample Compositions

We are releasing 20 sample compositions featuring a diverse range of instrumental arrangements. Each composition is provided as an MP3 audio file along with its corresponding sheet music in both MIDI and PDF formats.

Instruments included:

Piano
Violin
Cello
Flute
Bassoon
Trombone
Oboe
Viola
Access Sample Compositions →

Sample Solution

We provide a reference implementation of the MT3 (Multi-Task Multitrack Music Transcription) model developed by Google's Magenta team. MT3 uses a Transformer-based architecture to process audio inputs and generate accurate musical notations.

Note: To qualify as a winning submission, your model must achieve a transcription accuracy score higher than MT3.

MT3 Resources

Additional Sample Implementations

Basic Pitch by Spotify — Robust pitch tracking and transcription:

ReconVAT — Semi-supervised music transcription using variational autoencoders:

Organizers

Kristen Yeon-Ji Yun
Department of Music, Purdue University
yun98@purdue.edu
Yung-Hsiang Lu
School of Electrical and Computer Engineering, Purdue University
yunglu@purdue.edu
George K. Thiruvathukal
Department of Computer Science, Loyola University Chicago
gthiruvathukal@luc.edu
Tae Hong Park
Department of Music, Purdue University
thp@purdue.edu
Harry Bulow
Department of Music, Purdue University
hbulow@purdue.edu
Ojas Chaturvedi
Department of Computer Science, Purdue University
ochaturv@purdue.edu
Kayshav Bhardwaj
Department of Liberal Arts, Purdue University
bhardw43@purdue.edu

Contributing Composers

Harry Bulow
Department of Music, Purdue University
hbulow@purdue.edu
Tae Hong Park
Department of Music, Purdue University
thp@purdue.edu
Allen McCullough
Department of Music, Purdue University
mccullo3@purdue.edu
Hubert Howe
Juilliard School of Music (retired)
hubert.howe@gmail.com
Ka-Wai Yu
Department of Music, Utah Tech University
ka-wai.yu@utahtech.edu

For Contributing Composers

Composers retain the copyright of their works while granting royalty-free, non-exclusive rights to the challenge organizers for redistribution and analysis. Each invited composer is expected to contribute 10–30 pieces, each approximately 20 seconds long, divided into three difficulty levels: easy, medium, and difficult.

Composition Guidelines

  • Tempo: 60–90 bpm
  • Pitch Range: C2 to C7
  • Smallest rhythmic duration: sixteenth-notes/rests
  • No swing rhythms; use precise notation
  • No doubly-dotted notes; no trills or mordents
  • Meters: 3/4, 4/4, 6/8
  • Dynamic range: pp to ff
  • Up to three distinct instruments per composition
  • Submit files in PDF (score), MusicXML, and MIDI formats

Frequently Asked Questions

Yes, a live leaderboard will display the ranking of participating teams based on the performance of their models.
Not necessarily. The final ranking will be determined using holdout data that differs from the public sample data, ensuring the evaluation considers the model's generalization capabilities.
This can happen if a model overfits the public sample data and fails to perform well on the holdout data. The final ranking reflects the model's ability to generalize beyond the public samples.
The competition will provide sample data to understand the expected input/output formats. However, contestants are free to use any data (including public or proprietary data) for training their models.
Yes, industry participation is welcome. Partnerships between industry and academic teams are also encouraged.
Yes, open-sourcing the solution is mandatory for winners. Cash awards will only be provided after the winning models are made publicly available.
Yes, non-winning participants are not required to open-source their solutions. However, winners who choose not to open-source will not receive a cash award.
Yes, winners are encouraged to publish their findings. The organizers are also exploring opportunities for a special journal issue dedicated to the competition results.
No, presentation at the 2025 ICME is not mandatory, but winners are invited to share their work at the conference.