metadata
license: other
license_name: nvidia-one-way-noncommercial-and-mit-apache
license_link: LICENSE
base_model: nvidia/Real-time_RE-USE
pipeline_tag: audio-to-audio
library_name: hush-ai
tags:
- speech-enhancement
- noise-suppression
- noise-cancellation
- real-time
- streaming
- mamba
- deepfilternet
- rust
- cpu
- on-device
Hush AI · voice models
The model files that the Hush AI app downloads: free, open-source, on-device noise cancellation for every app.
Download the app · Website · GitHub
What is here
| File | Contents | License |
|---|---|---|
models/DeepFilterNet3_onnx-v0.5.6.tar.gz |
DeepFilterNet 3 by Hendrik Schröter, from release v0.5.6, unchanged (ONNX encoder and two decoders, 8 MB) |
MIT or Apache-2.0 |
models/realtime-reuse-4c0c186.safetensors |
NVIDIA Real-time RE-USE weights, revision 4c0c186, unchanged (FP32, 13.5 M parameters, 54 MB) |
NVIDIA One-Way Noncommercial |
models/SHA256SUMS |
SHA-256 of both model files | |
samples/noisy.mp3, samples/hush.mp3 |
NVIDIA's sample recording, before and after Hush AI | NVIDIA One-Way Noncommercial |
The app checks the SHA-256 before it uses a file. If this mirror is unreachable, it downloads the same file from its original source at the same version.
How Hush AI runs the models
Hush AI has two engines, and you can switch between them in the app. Both run on the CPU, with no Python, GPU or network:
- DeepFilterNet 3: the official Rust inference (libDF with tract), 48 kHz, 10 ms frames. Light on the CPU, and its license allows commercial use.
- NVIDIA Real-time RE-USE: a native Rust port of NVIDIA's Mamba model, 20 ms frames.
RE-USE quality settings in the app:
| Quality | Model setting | Apple M1, 2 threads (p99 per 20 ms frame) |
|---|---|---|
| Light (app default) | 16 kHz, 3 layers | 11.7 ms |
| Balanced | 24 kHz, 4 layers | 21.3 ms |
| Studio | 48 kHz, 3 layers | 36.8 ms |
- Delay: about 60–80 ms plus device buffers.
- Accuracy: NVIDIA's seven GPU regression cases pass in CI on Windows, macOS and Linux. The CPU output matches NVIDIA's GPU reference at 113–118 dB waveform SNR on short speech.
- Comparison: measured against DeepFilterNet 3 and RNNoise on VoiceBank-DEMAND in COMPARISON.md.
Use the RE-USE weights from Rust with the hush-core crate:
let mut denoiser = hush_core::Denoiser::with_config(
"models/realtime-reuse-4c0c186.safetensors",
hush_core::Config { layers: 3, sample_rate: 16_000, look_ahead: 0 },
)?;
let mut clean = Vec::new();
denoiser.process_into(&noisy_chunk, &mut clean)?;
Licenses
- DeepFilterNet 3 (
models/DeepFilterNet3_onnx-v0.5.6.tar.gz): © Hendrik Schröter, MIT (LICENSE-DeepFilterNet-MIT) or Apache-2.0 (LICENSE-DeepFilterNet-APACHE), at your option. Commercial use is allowed. Paper: Schröter et al., DeepFilterNet: Perceptually Motivated Real-Time Speech Enhancement, Interspeech 2023. - Real-time RE-USE (
models/realtime-reuse-4c0c186.safetensors) and the samples: NVIDIA's work, redistributed unchanged under the NVIDIA One-Way Noncommercial License (LICENSE): non-commercial research and educational use only. NVIDIA Corporation and its affiliates may use them commercially. See the original NVIDIA model card for the model, its training and its papers.
The Hush AI app and engine are Apache-2.0. Hush AI is an independent project and is not affiliated with or endorsed by NVIDIA or the DeepFilterNet authors.