jspr/symbolic-jazz-standards

Symbolic Jazz Standards

A symbolic-domain music dataset of jazz standards, transcribed stem by stem from the audio domain into the symbolic domain. The dataset contains the equivalent of 10,000 minutes of audio from ~200 public-domain well-known songs.

Methodology

To create this dataset, recordings of public-domain jazz standards were downloaded and separated into their component stems using the venerable Demucs source separation library in 4-stem mode. The resulting stems are:

vocals
bass
drums
other

The resulting audio-domain stems are then fed through a proprietary polyphonic music transcription model to obtain the stems' corresponding symbolic-domain representations — that is, the notes that are being played or sung in the music.

The transcriptions are polyphonic for 'other' stems, percussive for drum stems, and monophonic for vocal and bass stems.

Finally, the raw symbolic-domain data is tokenized via the following strategies:

• REMI
• MIDI-like
• TSD

This is performed with the excellent MidiTok library.

Dataset Structure

The dataset has the following columns:

• song_title: the title of the transcribed song, possibly including information about the performing artist.
• instrument_type: one of vocals, bass, drums, or other
• remi.tokens: a list of strings containing human-readable music tokens, in REMI format
• remi.ids: a list of integers representing machine-readable music tokens, in REMI format
• midilike.tokens: a list of strings containing human-readable music tokens, in MIDI-like format
• midilike.ids: a list of integers representing machine-readable music tokens, in MIDI-like format
• tsd.tokens: a list of strings containing human-readable music tokens, in TSD format
• tsd.ids: a list of integers representing machine-readable music tokens, in TSD format

Uses

This dataset is intended for fine-tuning or pre-training generative symbolic-domain music models, or for jointly conditioning audio-domain music models on the underlying symbolic-domain data.

Contact

For more info on this dataset, or to inquire about building similar datasets for your audio-domain data, please reach out to [email protected], or visit https://atonaldata.com

Legal Disclaimer

The user of this dataset ("User") assumes all responsibility and risk for the use of this dataset. The User agrees to indemnify, defend, and hold harmless Atonal Data, its affiliates, officers, directors, employees, consultants, agents, and representatives from any and all third party claims, losses, liability, damages, and/or costs (including reasonable attorney fees and costs) arising from the User's access to or use of the dataset, violation of this Agreement, or infringement of any intellectual property or other right of any person or entity.

Atonal Data provides this dataset on an "as is" basis without any express or implied warranties, including, but not limited to, warranties of merchantability or fitness for a particular purpose. In no event shall Atonal Data be liable for any direct, indirect, incidental, punitive, or consequential damages of any kind whatsoever with respect to the dataset.

This dataset is compiled under the doctrine of fair use, and it is the User's responsibility to ensure that their use of the dataset does not infrive upon any copyright laws. All songs contained in this dataset are believed to be in the public domain. However, Atonal Data does not warrant or represent that use of the dataset will not infringe rights of third parties. The User is responsible for ensuring that their use of this dataset complies with all applicable laws and regulations.

Citation

@misc{symbolicjazzstandards,
      title={Symbolic Jazz Standards}, 
      author={Atonal Data},
      year={2024},
}