#Sourcing Datasets for Multi-Stem Separation Standard vocal/instrumental separation is simple to set up because the input data contains only two stems. For a multi-stem model (Vocals, Backing Vocals, Drums, Bass, Guitars, Synths), you need a dataset structured with parallel instrument files.
1. MoisesDB (The Ultimate Dataset)
- What it is: Released by Moises.ai, MoisesDB is a massive open dataset designed specifically for advanced, multi-stem music demixing.
- Why it's perfect: It features 12 detailed stems:
- lead_vocals & backing_vocals
- drums & bass
- acoustic_guitar & electric_guitar
- piano & synthesizer
- strings & wind
- percussion & other
- How to get it: You can access the data and registration information through the official MoisesDB Github Page.
2. How to Organize Your Stems on your 2TB Drive
Your training folders should be organized as follows so the new MultiStemDataset in multi_stem_pipeline.py can read them automatically:
/Volumes/Your_External_Drive/Multi_Stem_Training/
├── song_1/
│ ├── vocals.wav
│ ├── backing_vocals.wav
│ ├── drums.wav
│ ├── bass.wav
│ ├── guitar.wav
│ └── synth.wav
├── song_2/
│ ├── vocals.wav
│ ├── backing_vocals.wav
│ ├── drums.wav
│ ├── bass.wav
│ ├── guitar.wav
│ └── synth.wav
Pro-Tip for Mixing Custom Stems:
If a song in your dataset contains multiple guitars (e.g. acoustic_guitar.wav and electric_guitar.wav), simply mix them together to create a single guitar.wav file before feeding it into your training setup!