MuseTimbre weights
Trained control modules for MuseTimbre: Zero-Shot Timbre Transfer by Controlling a Frozen Music Generator (Yuan-Chiao Cheng, Music Informatics Group, Georgia Institute of Technology; Zhiyao Duan, University of Rochester).
musetimbre_v1.pt holds only the trainable parts of the model: the pitch encoder, the pitch
cross-attention layers, the timbre projection and AdaLN layers, and the fine-tuned LAION-CLAP audio
encoder (about 420 M parameters, 1.7 GB). The frozen Stable Audio 3 Medium backbone and the
LAION-CLAP music checkpoint are not included; see the code repository for how to obtain them.
- Code and usage: https://github.com/barry-mir/muse-timbre
- Audio examples: https://barry-mir.github.io/muse-timbre-demo/
The inference script downloads this file automatically on first use.
License
The weights in this repository are released under the MIT license. Running the model also requires the Stable Audio 3 Medium backbone from Stability AI, which carries its own license.