Zum Inhalt springen
ModelAufgelistetExtern
Herunterladen

Re Use Semamba Mlx

nvidia-source-code-license-nc

Quellehuggingface.co/appautomaton/re-use-semamba-mlx/tree/c993ab948a50631e37a4cd7c6bbe39b5ab0aca3f
ArtOtherBasismodellnvidia/RE-USEVersionModelCBewertung

Modellquelle

Art
Other
Dateigröße
36,8 MB
Basismodell
nvidia/RE-USE
Version
vModel
Aktualisiert
22. Okt. 2026
Quelle
Hugging Face

Quellenbeschreibung

RE-USE SEMamba Speech Enhancement (MLX)

GitHub App Automaton DramaBox TTS

Pure-MLX conversion of NVIDIA RE-USE, a ~9.6M-parameter SEMamba universal speech-enhancement model. In mlx-speech it cleans a voice reference before VAE conditioning when DramaBox TTS runs with denoise_ref=True, giving the cloning model a clean speaker anchor.

Non-commercial weights. These weights derive from NVIDIA RE-USE, licensed under the NVIDIA Source Code License (non-commercial). See the License section.

Model Details

  • Developed by: App Automaton
  • Upstream model: nvidia/RE-USE (SEMamba, bidirectional Mamba over STFT magnitude + phase)
  • Role: input-side voice-reference denoiser for DramaBox denoise_ref=True. Optional, off by default.
  • Conversion: format-only port of the fp32 weights to MLX .safetensors (1416 keys, ~9.6M params). No quantization, no architecture change.
  • Runtime: pure MLX on Apple Silicon. The selective scan mirrors the mamba_ssm selective_scan_ref reference math, so no CUDA kernels (mamba-ssm / causal-conv1d) are required.
  • Parity: the MLX port matches the torch reference at amplitude-weighted complex correlation 0.9998 (model) and 0.9997 (end-to-end waveform on real speech).

Contents

FileComponentFormatSize
model.safetensorsSEMamba enhancerfp32~38 MB
config.jsonModel + STFT configJSONn/a

How to Get Started

Used automatically by DramaBox when you opt in:

import mlx_speech

tts = mlx_speech.tts.load("dramabox")
result = tts.generate(
    "Voice cloning from a noisy reference.",
    reference_audio="noisy_speaker.wav",
    denoise_ref=True,   # cleans the reference with this model first
)

tts.load("dramabox") resolves these weights automatically. To run the enhancer directly:

hf download appautomaton/re-use-semamba-mlx --local-dir models/reuse/mlx

from pathlib import Path
from mlx_speech.generation.reuse import REUSEEnhancer

enhancer = REUSEEnhancer.from_dir(Path("models/reuse/mlx"))
clean = enhancer.enhance(noisy_waveform, in_sr=16000)  # mono in, mono out

Intended Use

Denoising a short voice-reference clip before voice cloning, so the model conditions on a clean speaker/style anchor rather than the recording's noise. The enhancer runs on the reference input, never on generated output, so the TTS model's paralinguistic events (breaths, laughs) are preserved.

Links

License

NVIDIA Source Code License (non-commercial). These weights are a format conversion of nvidia/RE-USE and remain governed by NVIDIA's license terms; by downloading or using them you agree to those terms. They may not be used commercially. Set denoise_ref=False (the default) to run DramaBox voice cloning without this model. The mlx-speech runtime code is MIT.

Quellen

1 Quelle
Hugging FacePrimärVersion c993ab948a50631e37a4cd7c6bbe39b5ab0aca3f
huggingface.co/appautomaton/re-use-semamba-mlx/tree/c993ab948a50631e37a4cd7c6bbe39b5ab0aca3f

Verifiziert 7. Aug.

Modellartefakte

1 Artefakt

model.safetensors

safetensors · 36,8 MB · SHA-256 0c50ae8207f3…bb76 · Hugging Face

Herunterladen

Quellenauszüge

2 Auszüge
Readme7. Aug.huggingface.co/appautomaton/re-use-semamba-mlx

--- library_name: mlx pipeline_tag: audio-to-audio language: - en license: other license_name: nvidia-source-code-license-nc license_link: https://huggingface.co/nvidia/RE-USE base_model: - nvidia/RE-USE tags: - mlx - semamba - speech-enhancement - denoising - mamba - ssm - dramabox - apple-silicon - non-commercial --- # RE-USE SEMamba Speech Enhancement (MLX) [![GitHub](https://img.shields.io/badge/GitHub-mlx--speech-181717?logo=github&logoColor=white)](https://github.com/appautomaton/mlx-speech) [![App Automaton](https://img.shields.io/badge/App%20Automaton-project-1f6feb)](https://appautomaton.github.io) [![DramaBox TTS](https://img.shields.io/badge/%F0%9F%A4%97%20model-DramaBox%20TTS-yellow)](https://huggingface.co/appautomaton/dramabox-tts-3.3b-bf16-mlx) Pure-MLX conversion of [NVIDIA RE-USE](https://huggingface.co/nvidia/RE-USE), a ~9.6M-parameter SEMamba universal speech-enhancement model. In [mlx-speech](https://github.com/appautomaton/mlx-speech) it cleans a voice reference before VAE conditioning when [DramaBox TTS](https://huggingface.co/appautomaton/dramabox-tts-3.3b-bf16-mlx) runs with `denoise_ref=True`, giving the cloning model a clean speaker anchor. > **Non-commercial weights.** These weights derive from NVIDIA RE-USE, licensed > under the NVIDIA Source Code License (non-commercial). See the License section. ## Model Details - Developed by: [App Automaton](https://appautomaton.github.io) - Upstream model: [`nvidia/RE-USE`](https://huggingface.co/nvidia/RE-USE) (SEMamba, bidirectional Mamba over STFT magnitude + phase) - Role: input-side voice-reference denoiser for DramaBox `denoise_ref=True`. Optional, off by default. - Conversion: format-only port of the fp32 weights to MLX `.safetensors` (1416 keys, ~9.6M params). No quantization, no architecture change. - Runtime: pure MLX on Apple Silicon. The selective scan mirrors the `mamba_ssm` `selective_scan_ref` reference math, so no CUDA kernels (`mamba-ssm` / `causal-conv1d`) are required. - Parity: the MLX port matches the torch reference at amplitude-weighted complex correlation 0.9998 (model) and 0.9997 (end-to-end waveform on real speech). ## Contents | File | Component | Format | Size | | --- | --- | --- | --- | | `model.safetensors` | SEMamba enhancer | fp32 | ~38 MB | | `config.json` | Model + STFT config | JSON | n/a | ## How to Get Started Used automatically by DramaBox wh...

Note7. Aug.huggingface.co/appautomaton/re-use-semamba-mlx

Source context: 20 downloads · 3 likes · Pipeline audio-to-audio · Library mlx · Repo appautomaton/re-use-semamba-mlx