RE-USE Apple

Offline speech enhancement on Apple silicon with Swift, MLX, and Metal.

RE-USE Apple brings the pretrained 9.61M-parameter NVIDIA NemotronLabs RE-USE model to Apple silicon. It uses Swift, MLX, and Metal for offline speech enhancement, with a Mac command-line tool and an iPhone/iPad recording and import app sharing the same inference engine.

Recorded audio passes through a shared Swift, MLX, and Metal RE-USE engine to produce enhanced audio, with Mac and iPhone/iPad interfaces.
A conceptual view of the shared inference pipeline and platform interfaces.

The port includes a fused Metal Mamba scan, verified model downloads, audio conversion, and overlapping inference windows. After setup, audio processing stays on the device and produces 16 kHz mono WAV files.

Validation

Actual GPU inference is verified on an M4 Pro. The packaged tool produced byte-for-byte identical output to the validated Swift/Metal prototype and processed a four-second diagnostic clip in approximately 2.3 seconds, excluding model loading and file I/O. Checks also cover short inputs, silence, stereo resampling, and recordings spanning multiple inference windows.

The unsigned ARM64 iPhone/iPad build passes. Physical-device installation, performance, memory, and thermal behavior remain untested; the Mac timing is not a phone benchmark or a live-streaming claim.

Model and attribution

This is an independent port of NVIDIA NemotronLabs RE-USE. NVIDIA developed and trained the model; this project adapts inference and application packaging for Apple platforms. The pinned checkpoint permits commercial use under OpenMDW-1.1, with its license and attribution notices retained.

Code and quickstart · Validation details · RE-USE paper