2025 Automatic Music Transcription Challenge

(Details of this competition are subject to change)

Summary

A competition of Automatic Music Transcription (AMT) will be held in April 2025. In this competition, participants submit computer programs that can convert audio recordings of classical styles to Musical Instrument Digital Interface (MIDI). A submitted computer program has up to 10 minutes to convert 100 recordings; each recording may be up to 20 seconds long and may contain at most three musical instruments. The rank of the submissions will consider multiple factors, including correct instruments, pitches, onset, offset, and dynamics. Most of the audio will be produced by using synthesized sound; a small percentage will be recorded from human performance.

More details will come soon.

Organizers:

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

Technical Details

Contestants will register on the website of https://ai4musicians.org/. Sample music (including the scores and the audio recordings) will be provided on the website. Contestants can use any data for training their programs. Submissions will be open in March 2025. The programs submitted to the website will run on a computer equipped with GPU (Graphics Processing Unit) in Purdue's Rosen Center for Advanced Computing (https://www.rcac.purdue.edu/). Each team can submit at most once every 24 hours. A leaderboard will show the rank of the teams running the sample data. An additional set of holdout data will be used to determine the final rank and awards. A sample (open-source) solution will be available for contestants to study the input and output formats. If a submitted solution is inferior to the sample solution, the submitted solution is disqualified for consideration of an award. Winners will be invited to present their solutions at a conference (TBD).

For Contributing Composers

For the 2025 challenge, the composers are by invitation by the organizers only. The reason is that the organizers wish to manage the difficulty of the music so that computer programs can provide meaningful results. If a composer is interested in contributing, please contact the organizers. The composers will retain their copyright. Composers will agree to provide their music royalty-free and non-exclusive rights to the challenge organizers for the purpose of this automatic music transcription challenge, including (but not limited to) redistributing the music as samples and analyzing the music by computer programs. The organizers may modify the music from the composers to meet the restrictions of the scoring program. The names, profile photos, and links to the composers' websites will be listed on the https://ai4musicians.org/ website if a composer is selected by the organizers. 

A composer will provide at least 10 and at most 30 pieces of music. Each piece is approximately 20 seconds long. These numbers aim to give each composer a reasonable representation without causing computer programs to "overfit" into a specific composer's style. Each piece is between 10 and 20 seconds. The music should be divided into 3 levels of difficulties: easy, medium, and difficult. 

• Timing: 4/4 
• Beats per minute: 90
• Fastest note: sixteenth (4 sixteenth = 1 beat)
• Dotted notes: allowed in the difficult group (only one dot, not doubly dotted)
• Each piece is approximately 20 seconds.
• Style: "classical music" or the subdivision of the beats must be equal (no swing or Jazz interpretations)
• No human voice.
• The number of notes played at any moment by all instruments combined: easy at most two; medium: at most four; difficult: at most 5.
• The notes will be between C2 and C7.

Instruments: At most three instruments in any recording.

piano violin clarinet cello saxophone

6. trumpet 7. bassoon 8. piano and cello 9. saxophone 10. flute and violin

11. flute and trumpet 12. clarinet and violin 13. violin and bassoon 14. piano, violin, and trumpet

Please notice that the instrument groups do not include instruments that can be misclassified easily even by human experts (e.g., violin and viola). 

Cash Awards (tentative)

1. The top award is up to $1,500 US Dollars.
2. The second award is up to $1,000 US Dollars.
3. The third award is $500 US Dollars.

The winners must open source their solutions before receiving the cash awards (required in the registration agreement). The source code should have proper documentation for reproducibility. The organizers will examine the source code before announcing the winners. The cash awards can be given to only the participants that are not in the embargoed and sanctioned countries declared by the United States.

The award may be in the form of cash or travel grants that may include conference registration + hotel + airplane ticket.

Timeline (tentative)

• 2024/10/01 five pieces of music from each composer (will be available to contestants).
• 2024/10/15 initial results from state-of-the-art transcription programs (such as MT3 and Anthem Score). The purpose is to evaluate existing technologies.
• 2024/11/01 five pieces of music from each composer (will be available to contestants).
• 2024/12/01 announce the competition.
• 2024/12/31 ten pieces of music from each composer (will not be available to contestants).
• 2025/01/01 registration open.
• 2025/02/01 release sample solution.
• 2025/03/01 submission open.
• 2025/03/31 submission close.
• 2025/04/30 announce winners.
• 2025/06 winners present their solutions in a conference (TBD)

Frequently Asked Questions

Q: Will a leaderboard be provided?

A: Yes.

Q: If my team is No. 1 in the leaderboard, does that mean my team will be No. 1 in the final winner announcement?

A: Maybe yes, maybe not. The final winners will be determined by a set of holdout data that is similar to, but different from, the public sample data.

Q: How can it be possible that No.1 in the leaderboard is not No.1 in the final announcement?

A: It is possible if the solution overfits the public data but does not generalize well for the holdout data.

Q: Will the organizers provide training data?

A: The competition will provide sample data, not training data. Contestants can use any data (including public and proprietary data).

Q: Can industry participate?

A: Yes.

Q: Can industry partner with academia?

A: Yes.

Q: Is open-source required?

A: Open-source is required for winners before receiving the awards.

Q: Can I participate and receive the ranking, without open-source?

A: You can participate without open-source if you are not a winner. If you are a winner and do not open-source, you will not receive an award.

Q: What is the required license for the winners solutions?

A: The winners decide their licenses.

Q: Can winners submit papers about their solutions?

A: Yes, certainly. The organizers are also seeking the opportunities of a special issue in a journal.