Release Date: December 27, 2025
AudioMuse v0.8.5 continues the work started in #225, with further improvements to the stability of CLAP analysis.
We introduced a new environment variable, PER_SONG_MODEL_RELOAD, which is enabled by default. When active, both the Musicnn and CLAP models are reloaded before each song inference.
The main benefit of this approach is improved memory stability: it prevents RAM usage from continuously increasing, at the cost of slightly slower analysis. We believe prioritizing stability over speed is the right choice.
A big thank you to everyone who helped test this release and shared their results, your feedback is invaluable in making AudioMuse-AI better every day.
What's Changed
- Add standalone student CLAP distillation framework by @Copilot in #252
- Devel -> Main - CLAP mini batch by @NeptuneHub in #253
Full Changelog: v0.8.4...v0.8.5