Drama Studio: native macOS app + Metal inference engine for Drama 3
Hi everyone, I'm Peter, the developer of Drama Studio. I've built a native macOS app that runs Fish Audio's Drama 3 locally on Apple silicon.
Repository and installation instructions: https://github.com/maralski/drama-studio
The 35B mixture-of-experts backbone and codebook tower run on optimised Metal compute kernels that read the released safetensors weights in place. The DAC audio codec uses Core ML. There is no MLX or PyTorch dependency at runtime.
The app includes:
- Multi-speaker scripts, voice cloning from recordings you have permission to use, and a local voice library.
- PDF, EPUB and PowerPoint narration, with audio and caption exports.
- A local HTTP API and an MCP server so scripts and agents can use the same engine and voices.
Hardware requirements are substantial. Generation has been tested on an M3 Max with 128 GB of unified memory. The engine maps about 41 GB of model weights, and setup needs about 49 GB of free disk space. 64 GB or more of unified memory is recommended, but 32 GB and 64 GB Macs remain untested. Speech generation runs locally after the initial downloads.
The app source is Apache-2.0 licensed; the model weights remain subject to Fish Audio's license terms.
I'd welcome feedback from other Drama 3 users, particularly hardware reports with your chip, memory and macOS version, or questions about the Metal implementation. If you'd like to follow development, you can star the repository.
Thanks to the Fish Audio team for releasing Drama 3!