What is NVIDIA NemotronLabs RE-USE?
NVIDIA NemotronLabs RE-USE is a universal speech enhancement model: it takes a damaged speech recording and returns a cleaner one. NVIDIA's card lists noise, reverberation, clipping, limited bandwidth, codec artefacts, packet loss and poor microphones among the problems it repairs. It accepts mono audio at seven sample rates from 8 kHz to 48 kHz and can optionally extend narrowband audio to a higher bandwidth. The card gives a Hugging Face release date of 22 October 2026.
It has 9.6 million parameters: convolutional encoder and decoder around a 30-layer bidirectional Mamba network. It was trained on about 2,500 hours of speech in English, German, Spanish, French and Chinese mixed with synthetic noise and room echo, and NVIDIA describes it as language-agnostic. For overlapping talkers it keeps the main speaker and suppresses the others; a separate laughter variant is offered for better laughter preservation.
The weights are under OpenMDW-1.1, a permissive licence that allows commercial use and places no restrictions on outputs. The research behind it is in two 2026 papers, at NeurIPS and IEEE SLT.
Is NVIDIA NemotronLabs RE-USE worth it?
RE-USE's strength is how much it fixes for its size. On the URGENT 2025 non-blind test set, NVIDIA reports a PESQ score of 2.57 against 1.73 for the 250-million-parameter Sidon and 2.17 for the 500-million-parameter UniPASE, and an ASR character accuracy of 89.17% against 84.68% and 87.28%. Higher speaker similarity, 0.86 against 0.74 and 0.81, means it changes the voice less while cleaning it. Sidon still edges it on two perceived-quality scores, DNSMOS and UTMOS. All of these are NVIDIA's runs.
The licence is the other draw. OpenMDW-1.1 grants use without restriction, asks only that you keep the notices when you redistribute, and says nothing about what you do with the cleaned audio. NVIDIA calls it ready for commercial use. The licence also makes you responsible for clearing any third-party rights, and rights end if you sue claiming the model infringes a patent or copyright.
The catch: despite its size, it is not a CPU model as shipped. The official inference script stops with "CPU mode is not supported" when no CUDA device is found, the card lists Linux and an A100 as the tested hardware, and the Mamba layers rely on the mamba-ssm CUDA package. It was trained on clips under 10 seconds, so files longer than about 30 seconds go through a separate chunking script.
Our take: for a team cleaning a speech dataset before training a TTS or speech model, or restoring call-centre and archive audio in bulk on GPUs it already has, RE-USE is a small, permissively licensed default to try first. For a one-off podcast clean-up on a laptop, a hosted tool is less work.
Where it earns its place
An ML engineer preparing a 1,000-hour TTS training set from scraped audio
Remove noise, echo and codec damage without changing who is speaking
Batch the files through inference.sh on an A100 or similar; NVIDIA reports speaker similarity of 0.86 on URGENT 2025, above larger open models. NVIDIA's own use case names exactly this: cleaning data for TTS and speech LLMs.
A contact-centre analytics team at a telecoms firm
Improve transcription of 8 kHz phone calls with packet loss
Run calls through RE-USE before ASR, with bandwidth extension on; on NVIDIA's full test set, ASR character accuracy rose from 81.29% on the noisy input to 89.17%. Long calls go through inference_chunk.sh.
An archivist at a regional broadcaster
Restore old interview tapes in several languages
The model was trained on five languages and is described as language-agnostic. Use the laughter variant for conversational material, and keep the originals, since enhancement can suppress background speakers.
Before you commit
- Setup
- Install a Mamba environment (NVIDIA points to SEMamba's Docker images), download the repository, put WAV files in noisy_audio/ and run sh inference.sh; add pip install resampy for bandwidth extension.
- Integrations
- Plain PyTorch scripts (inference.py, inference_chunk.py) with a YAML config; no NeMo, Riva or API wrapper. A Gradio demo runs on a Hugging Face ZeroGPU Space.
- Data and privacy
- Runs on your own hardware, so audio stays local. OpenMDW-1.1 allows commercial use and modification, keeps outputs unrestricted, and puts clearing third-party rights on you; training data includes an internal NVIDIA speaker set.
- Cost at scale
- Weights are free and the model is 9.6 million parameters, so GPU memory is not the constraint; a single Ampere-class GPU runs it. No hosted price exists.
- Limits
- CUDA and Linux only in the official code; mono input; trained on clips under 10 seconds; suppresses overlapping speakers; the project does not accept contributions.
What’s good
- 9.6 million parameters, against 250M and 500M for open rivals
- Fixes noise, echo, clipping, codec and packet-loss damage in one model
- Seven input sample rates from 8 kHz to 48 kHz
- OpenMDW-1.1: commercial use allowed, outputs unrestricted
What’s not
- Official script needs a CUDA GPU; no CPU mode
- Benchmarks are NVIDIA's own
- Laughter preservation needs a separate variant
- Long files need the chunked script
Best for
Speech and voice-AI teams with NVIDIA GPUs who need to clean large amounts of real-world recordings under a permissive licence.
NVIDIA NemotronLabs RE-USE: common questions
- Is NVIDIA NemotronLabs RE-USE worth it?
- RE-USE repairs damaged speech better than open models 25 to 50 times its size on NVIDIA's own tests, with a licence that lets you use it anywhere, if you have a CUDA GPU. We rate it 3.9 out of 5.
- How much does NVIDIA NemotronLabs RE-USE cost?
- Free under OpenMDW-1.1, which permits commercial use. No hosted API; you run it on your own NVIDIA GPU.
- What are the downsides of NVIDIA NemotronLabs RE-USE?
- Official script needs a CUDA GPU; no CPU mode; benchmarks are NVIDIA's own; laughter preservation needs a separate variant; long files need the chunked script.
- Who is NVIDIA NemotronLabs RE-USE best for?
- Speech and voice-AI teams with NVIDIA GPUs who need to clean large amounts of real-world recordings under a permissive licence.
- How much does NVIDIA NemotronLabs RE-USE cost for a team or in production?
- Weights are free and the model is 9.6 million parameters, so GPU memory is not the constraint; a single Ampere-class GPU runs it. No hosted price exists.
- Is NVIDIA NemotronLabs RE-USE safe for company data?
- Runs on your own hardware, so audio stays local. OpenMDW-1.1 allows commercial use and modification, keeps outputs unrestricted, and puts clearing third-party rights on you; training data includes an internal NVIDIA speaker set.
Answers checked Nov 11, 2026.