Hybrid Transformer for Music Source Separation

We present here samples for Hybrid Transformer Demucs, both the sparse and non sparse version, along with the classic Hybrid Demucs retrained on the same dataset. Go to our repository for more information. Note that given the strong SDR of the base Hybrid Demucs model retrained on the dataset (8.3 dB), some differences can be subtle.

The Easton Ellises - Falcon 69 (CC BY-NC SA)

Mixture:

Model	Drums	Bass	Other	Vocals
Reference
HDemucs
HT Demucs
HT Demucs (f.t.)
Sparse HT Demucs (f.t.)

Cy Curnin - Comfy Couches (Proprietary test set)

Mixture:

Model	Drums	Bass	Other	Vocals
Reference
HDemucs
HT Demucs
HT Demucs (f.t.)
Sparse HT Demucs (f.t.)