Consistent stem format - drums, bass, vocals, other?

[sevagh]

If you've used the current versions of the OnAir dataset, you'll notice the stems are in an inconsistent raw format (that we receive from the artists).

For example, the stems for the track "Ronin" (https://open.spotify.com/track/7jULsqywM6aWSZfaB7BAtY?si=a84eb0880f644267) are:

├── Ronin - OnAir Music
│   ├── Ronin ft. KXNE Stems - 808.wav
│   ├── Ronin ft. KXNE Stems - Acappella.wav
│   ├── Ronin ft. KXNE Stems - Drums.wav
│   ├── Ronin ft. KXNE Stems -  Guest Hook.wav
│   ├── Ronin ft. KXNE Stems -  Guest Verse.wav
│   ├── Ronin ft. KXNE Stems - Hook Lead.wav
│   ├── Ronin ft. KXNE Stems - Instrumental.wav
│   ├── Ronin ft. KXNE Stems - Main Loop.wav
│   ├── Ronin ft. KXNE Stems - Plucks.wav
│   ├── Ronin ft. KXNE Stems - Verse.wav
│   ├── Ronin ft. KXNE Stems - Vox FX.wav
│   └── Ronin - OnAir Music.wav

Should I continue exposing the raw stems? Will researchers be happy to listen to each track and decide for themselves which waveform they want to use in different parts of their algorithm or neural network?

Or should we follow the example of MUSDB18-HQ with 4 stems:

1. drums
2. bass
3. vocals
4. other

This is open for the community to discuss.

[sevagh]

I created a musdb-like loader: https://github.com/OnAir-Music/onair-py
It defines mappings from the custom stems of OnAir, into the 4 musdb targets. E.g. "drums = brushes + kick + snare + ..."