🟢1. Dataset Creation

To create a model, it all starts with the dataset.

To create a dataset, the first thing you need is good acapellas. Using official studio acapellas would be the best option. Currently, the best AI model to use is UVRv5 with the Kim_Vocal and UVR-Karaoke models:

Vocal and Music Separation

Then, take your acapellas and drag them into Audacity. Highlight a section of good vocals, then press CTRL + B to create a new label. The label does not need to contain any text. Continue to do this for all the good vocal parts you can extract from your acapellas.

Once you have finished labeling, go to File -> Export -> Export Multiple... Select these options along with the location for your dataset. Then, click "Export," and all your files should be exported automatically.

Your data should be WAV files, placed in a folder, and then this folder should be compressed into a ZIP file.

The vocals you use with the AI should be as raw (unprocessed) as possible. This will ensure consistent results.