The Spheres Dataset

Garcia-Martinez, Jaime; Diaz-Guerra, David; Anderson, John; Falcón Pérez, Ricardo; Cabañas-Molero, Pablo; Virtanen, Tuomas; Carabias Orti, Julio José; Vera-Candeas, Pedro

doi:10.5281/zenodo.17347681

Published October 14, 2025 | Version v1

Dataset Open

The Spheres Dataset

1. Universidad de Jaén
2. Tampere University
3. Odratek BV

The Spheres dataset is a novel multitrack orchestral recording resource designed to advance machine learning research in music source separation and related MIR tasks within the classical music domain. The dataset is composed of musical pieces performed by the Colibrì Ensemble at The Spheres recording studio, capturing two canonical works—Tchaikovsky’s Romeo and Juliet and Mozart’s Symphony No. 40—along with chromatic scales and solo excerpts for each instrument. The recording setup employed 23 microphones, including close spot, main, and ambient microphones, enabling the creation of realistic stereo mixes with controlled bleeding and providing isolated stems for supervised training. In addition, room impulse responses were estimated for each instrument position, offering valuable acoustic characterization of the recording space.

The signals are provided as lossless FLAC files at 48 kHz and structured as in the EnsembleSet and SynthSOD datasets: a folder per music piece, inside them and folder per microphone, and, inside them, an audio file per instrument line. Metadata files in JSON format, similar to the ones included in SynthSOD, are also provided.

Technical info

This repository contains the following files:

TheSpheresDataset-Multichannel: full version of the dataset, with the signals of all the instruments (up to 18 different instruments with up to 33 separated lines) in all the microphones (23) and the stereo mix for the Mozart and Tchaikovsky pieces, and some extracts with scales and solos from every instrument.
TheSpheresDataset-StereoMix: signals of all the instruments (up to 18 different instruments with up to 33 separated lines) in the stereo mixture for the Mozart and Tchaikovsky pieces, and some extracts with scales and solos from every instrument. Extracted from the full version of the dataset and provided separately for those users who might be only interested in the stereo mix but not in the microphone signals.
TheSpheresDataset-RIRs: room impulse responses (RIRs) between the different source positions and microphones computed according to the procedure described in the paper. The RIRs are provided as numpy arrays and plots are also included in PDF format.
TheSpheresDataset-ClapsAndSweeps: raw recordings at every microphone of claps and sweeps generated at every source position. The Python scripts employed to compute the RIRs from these recordings are also included.

Files