2025 Automatic Music Transcription Challenge

Submission Details

Participants who have registered for the competition must follow these guidelines for submitting their models:

Repository Access

During registration, participants are required to provide a link to their code repository along with a fine-grained access token. This token is necessary for the competition backend to pull the model for execution.

Submission Branch

To submit a model, participants must create a branch titled submission in their repository. The competition backend will automatically pull from this branch to run the model.

Submissions are validated if the following conditions are met:

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

Environment Configuration

Participants must include an environment.yml file in the root directory of their GitHub repository. This file will be used to create a conda environment during the model execution. Please ensure that the environment file correctly specifies all dependencies and versions required to run the model.

Model Execution Requirements

To run the model, the repository must contain a file named main.py located in the root directory. This script must accept the following command-line arguments:

• -i: Path to the input audio file (in .mp3 format).
• -o: Path to save the output MIDI file.

Example usage:

python main.py -i input.mp3 -o output.midi

All paths and directories referenced in the main.py file should be relative to its own location. This ensures compatibility with the backend environment.

Input File Naming Convention

The input audio file names will include the MIDI instrument codes to specify the instruments used in the recording. The MIDI instrument codes follow the General MIDI standard as described here. The file name format will be as follows:

Example:

1._0_40_70.mp3

In this example, the numbers 0, 40, and 70 correspond to the MIDI instrument codes of the instruments present in the audio file. Participants may choose to use this instrument information or ignore it. However, if the model incorrectly identifies the instruments, there will be a scoring penalty. The penalty is reduced if the mistake involves instruments from similar families.

Submission Process

As long as the submission branch exists and new commits are detected, the system will automatically run the model. The backend will handle the following:

• Pulling the latest code from the submission branch.
• Creating a conda environment using the provided environment.yml file.
• Executing the model using the specified command.
• Sending email notifications to the participant, including:
• Status updates on the model run.
• Input and output of the model.
• Performance statistics of the model.
• The SLURM job output, allowing the participant to verify that the backend successfully processed their model.

Model Weights

Participants may use Git LFS (Large File Storage) for managing model weights within their repository. The backend supports Git LFS, so there are no restrictions on using large files.

Schedule

The following schedule outlines the key dates for the 2025 Automatic Music Transcription Challenge. Please refer to these dates to stay on track throughout the competition.

Cash Awards

The top-performing teams will be awarded cash prizes as follows:

• First Place: Up to $1,500 USD
• Second Place: Up to $1,000 USD
• Third Place: $500 USD

To receive the cash award, winners must open-source their solutions as specified in the registration agreement. The submitted code must include adequate documentation to ensure reproducibility. The organizers will conduct a thorough examination of the source code before officially announcing the winners.

Please note that cash awards will only be provided to participants from countries not subject to United States embargoes or sanctions. In some cases, the award may take the form of travel grants, covering conference registration, hotel accommodations, and airfare.

Sample Compositions

To provide participants with a clear understanding of the type of music their models will be evaluated on, we are releasing 20 sample compositions. These compositions feature a diverse range of instrumental arrangements, allowing participants to fine-tune their models for optimal performance.

The following instruments are included in the sample compositions:

1. Piano
2. Violin
3. Cello
4. Flute
5. Bassoon
6. Trombone
7. Oboe
8. Viola

Each sample composition is provided as an MP3 audio file along with its corresponding sheet music, available in both MIDI and PDF formats. These files serve as a representative dataset to assist participants in developing and testing their models.

You can access the sample compositions here: Sample Compositions Google Drive Link

Sample Solution

To assist participants in developing their models, we provide a reference implementation of the MT3 (Multi-Task Multitrack Music Transcription) model. MT3, developed by Google's Magenta team, is an advanced model designed to transcribe music involving multiple instruments simultaneously. It utilizes a Transformer-based architecture to process audio inputs and generate accurate musical notations.

To further support participants, we have created a dedicated repository that adapts MT3 for the competition environment. This repository demonstrates how to set up and run the model within the submission framework.

You can access the MT3 competition-ready implementation here:

• MT3 Competition-Ready Implementation: github.com/ojas-chaturvedi/mt3-pytorch
• Original MT3 Repository by Google Magenta: github.com/magenta/mt3
• MT3 Research Paper: MT3: Multi-Task Multitrack Music Transcription

Additional Sample Implementations

In addition to MT3, we provide implementations of two other models to offer a broader perspective on music transcription approaches.

• Basic Pitch by Spotify: A model designed for robust pitch tracking and transcription.
  • Implementation: github.com/KayshavBhardwaj/basic-pitch-test-1
  • Research Paper: Basic Pitch: Spotify's Pitch Tracking Model
• ReconVAT: A model designed for semi-supervised music transcription using variational autoencoders.
  • Implementation: github.com/ojas-chaturvedi/ReconVAT
  • Research Paper: ReconVAT: Semi-Supervised Learning for Music Transcription

Participants are encouraged to study these implementations as foundational models to gain insights into effective music transcription architectures. They serve as valuable starting points for developing and refining your own transcription models.

Note: To qualify as a winning submission, your model must achieve a transcription accuracy score higher than MT3. This criterion ensures that new models demonstrate meaningful progress in the field of music transcription.

Organizers

The 2025 Automatic Music Transcription Challenge is organized by leading experts from multiple institutions, dedicated to advancing research in music transcription and computational musicology.

Photo	Name	Email	Organization
	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 K. Thiruvathukal	gthiruvathukal@luc.edu	Department of Computer Science, Loyola University Chicago
	Tae Hong Park	thp@purdue.edu	Department of Music, Purdue University
	Harry Bulow	hbulow@purdue.edu	Department of Music, Purdue University
	Ojas Chaturvedi	ochaturv@purdue.edu	Department of Computer Science, Purdue University
	Kayshav Bhardwaj	bhardw43@purdue.edu	Department of Liberal Arts, Purdue University

Contributing Composers

The following composers have contributed music to the challenge, helping to establish a diverse and realistic dataset for model training and evaluation.

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

For Contributing Composers

The 2025 challenge features invited composers to ensure the controlled complexity of the musical data. The objective is to produce high-quality compositions that allow meaningful evaluation of automatic transcription models.

Composers retain the copyright of their works while granting royalty-free, non-exclusive rights to the challenge organizers for redistribution and analysis. The organizers may modify compositions to meet scoring requirements.

Each invited composer is expected to contribute between 10 to 30 pieces, each approximately 20 seconds long, and divided into three difficulty levels: easy, medium, and difficult.

Guidelines for Composition:

• 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, avoiding combinations like violin and cello due to similar timbre
• Submit files in PDF (score), MusicXML, and MIDI formats

The selected instrument set is as follows:

1. Piano
2. Violin
3. Cello
4. Flute
5. Bassoon
6. Trombone
7. Oboe
8. Viola

Frequently Asked Questions