Deep Learning Based Source Separation Applied To Choir Ensembles

Darius Petermann, Pritish Chandna, Helena Cuesta, Jordi Bonada, and Emilia Gomez

Music Technology Group, Universitat Pompeu Fabra, Barcelona



Two of our pre-trained models - Conditioned-U-Net Local & Global - can be found at the following:
https://drive.google.com/drive/folders/1zqdSLCGJ7cqw7oCP6iEhh3t2uIY9sC31?usp=sharing



Audio Separation Examples:

Use-Case#1 - 4-Singer Mixture:

Mixture:

Model Domain Soprano Alto Tenor Bass
Reference -
Wave-U-Net Domain-Agnostic
U-Net Domain-Agnostic
Open-Unmix Domain-Agnostic
Conditioned-U-Net (Original) Domain-Agnostic
Conditioned-U-Net (Local) Domain-Specific
Conditioned-U-Net (Global) Domain-Specific
Conditioned-U-Net (Encoder) Domain-Specific
Conditioned-U-Net (Global, x2) Domain-Specific

Use-Case#2 - 16-Singer Mixture:

Mixture:

Model Domain Soprano Alto Tenor Bass
Reference -
Wave-U-Net Domain-Agnostic
U-Net Domain-Agnostic
Open-Unmix Domain-Agnostic
Conditioned-U-Net (Original) Domain-Agnostic
Conditioned-U-Net (Local) Domain-Specific
Conditioned-U-Net (Global) Domain-Specific
Conditioned-U-Net (Encoder) Domain-Specific
Conditioned-U-Net (Global, x2) Domain-Specific


Metrics and Boxplots

Use-Case#1 - 4-Singer Mixture:






Use-Case#2 - 16-Singer Mixture:



















[1] Stoller, Daniel, Sebastian Ewert, and Simon Dixon. "Wave-u-net: A multi-scale neural network for end-to-end audio source separation." arXiv preprint arXiv:1806.03185 (2018).

[2] Jansson, Andreas, et al. "Singing voice separation with deep U-Net convolutional networks." (2017).

[3] Meseguer-Brocal, Gabriel & Peeters, Geoffroy. "Conditioned-U-Net: Introducing a Control Mechanism in the U-Net for Multiple Source Separations" (2019).

[4] Helena Cuesta, Emilia Gómez, Agustín Martorell, & Felipe Loáiciga. Choral Singing Dataset. Zenodo. (2018).