Feature Request: Apple MLX Support for ZipVoice
Hi k2-fsa team,
First of all, thank you for open-sourcing this excellent project! ZipVoice looks very impressive, especially its fast inference speed, compact model size, and high-quality zero-shot voice cloning. I’ve really enjoyed trying it out.
I would like to request support for Apple MLX.
Apple Silicon devices (M1, M2, M3, M4) have become increasingly popular among developers and researchers. MLX has also become one of the most widely used inference frameworks for running AI models efficiently on macOS, thanks to its excellent performance and low memory overhead.
At the moment, ZipVoice provides PyTorch and ONNX support, but there doesn’t seem to be an official MLX version. Having one would make ZipVoice much more accessible for the Apple Silicon community.
Some potential benefits include:
- Native acceleration on Apple Silicon GPUs
- Lower memory usage compared to PyTorch
- Easier local deployment for macOS users
- Better integration with the growing MLX ecosystem
I believe many developers in the local AI community would love to use ZipVoice with MLX.
Is there any plan to officially support MLX in the future? Even an experimental implementation would be greatly appreciated.
Thank you again for your amazing work, and I look forward to future updates!
If there is community interest, I’d also be happy to help test an MLX implementation on my Apple Silicon machine.