resemble-enhance-mlx

Apple MLX port of resemble-enhance (speech denoising + enhancement). Runs entirely on the MLX runtime — no PyTorch dependency at inference time.

The weights are converted from the original resemble-enhance checkpoint and stored as a single fp16 safetensors (710 MB). Every component (mel frontend, spectral denoiser UNet, latent conditional-flow-matching enhancer, UnivNet/LVCNet vocoder) was ported op-by-op and validated against the PyTorch reference.

Install

pip install mlx numpy scipy soundfile huggingface_hub

Usage

import soundfile as sf
from huggingface_hub import hf_hub_download
import resemble_enhance_mlx as re

ckpt = hf_hub_download("multimodalart/resemble-enhance-mlx", "resemble-enhance-mlx.safetensors")
model = re.load(ckpt, lambd=0.5)          # lambd in [0,1]: denoiser strength

wav, sr = sf.read("noisy.wav", dtype="float32")
out, out_sr = re.enhance(model, wav, sr)  # resamples to 44.1 kHz, chunks long audio
sf.write("clean.wav", out, out_sr)

Place resemble_enhance_mlx/ (the package) next to your script, or add this repo to PYTHONPATH.

Parity & speed

Validated against the PyTorch reference (max abs diff, random-weight and full-model):

Component	vs PyTorch
Mel frontend	2.8e-5
UnivNet / LVCNet vocoder	7.2e-7
Denoiser (STFT + 2D UNet + ISTFT)	1.6e-4
Latent CFM enhancer (IRMAE + WN + ODE)	9.2e-6
End-to-end waveform	corr 0.99999

fp16 storage vs fp32 compute: corr 1.000000, mean diff 4.5e-5 (inaudible).

On an M4 Max, a 6.5 s clip enhances in ~1.7 s (nfe=64), ~7.5× faster than the PyTorch reference on CPU.

Notes

• Input audio is resampled to 44.1 kHz with a polyphase (kaiser) filter; this is a close approximation of torchaudio's sinc_interp_kaiser, so non-44.1 kHz input is not bit-identical to the reference.
• lambd=0 skips the denoiser (enhancement only); lambd=0.5 is the default.

Original model and method: Resemble AI. This repository only re-packages the weights for MLX.