Recap¶

In this Chapter we covered:

Introduction ¶

What are the inputs/outputs of a source separation model
How mixtures are created
How data are used to train a source separation model (high level)
Why data are key for music source separation, and why they are hard to obtain

Datasets ¶

Overview of existing datasets
Important dataset characteristics to look out for
Why it’s important to always listen to your data!

MUSDB18 ¶

Overview of the MUSDB18 dataset we will use in this tutorial
How to download the dataset preview clips using nussl
How to inspect our data via audition and visualization

Generating mixtures with Scaper ¶

Why use Scaper rather than write ad-hoc mixing code
Overview of Scaper
How to prepare your source material (stems) for processing with Scaper
How to generate randomly augmented mixtures
The difference between incoherent and coherent mixtures and how to generate both
How to plug our Scaper mixing code into nussl

Generating mixtures with Scaper Introduction