NVIDIA NemotronLabs RE-USE

NVIDIA NemotronLabs RE-USE

Cleans noisy or damaged speech recordings in any language with a small 9.6M-parameter model from NVIDIA. For audio and speech developers who need clearer voices.

Made by nvidia on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: NVIDIA NemotronLabs RE-USE is a speech enhancement model with 9.6 million parameters. It repairs recordings hurt by background noise, echo, clipping, narrow bandwidth, codec artifacts, packet loss or poor microphones, and it tries to keep the voice faithful. It works across languages and takes audio at several sampling rates.

Who it suits: It suits audio engineers, speech developers and researchers who need cleaner voice recordings from messy sources.

Industry | Who uses it | How they use it | Example
Media | Podcast and video editors | Clean speech recorded on weak microphones | An interview recorded in an echoing room
Telecom | Call platform teams | Repair speech damaged by codecs and packet loss | A call recording that cuts out and sounds thin
Speech AI | Recognition developers | Clean audio before it reaches a transcriber | Reported character accuracy rises from 81.29 for noisy audio to 89.17
Research | Audio researchers | Compare enhancement models | Test against Sidon and UniPASE on the same set

If your audio is already clean, or you need to separate several speakers, you can skip it.

Where to use it:

• Python projects that use the mamba-ssm library
• A Gradio demo on Hugging Face where you upload noisy audio or video
• Pre-built Docker environments for the Mamba setup
• Windows, macOS or Linux machines with Python

Use case | Who | What they get back
Cleaning recordings | Editors | An enhanced speech file for each noisy file you provide
Bandwidth extension | Audio engineers | Speech rebuilt up to a higher target bandwidth
Preparing audio for recognition | Speech developers | Cleaner input for transcription tools

How to set it up: You need these first:

• Python and a working Mamba setup, where the project points to pre-built Docker environments
• The resampy package, only if you want bandwidth extension
• A little disk space, since the F32 weights are about 38 MB

1. Set up Mamba, using one of the pre-built Docker environments if you like.
2. Install resampy when you need bandwidth extension.
3. Download the repository and move into its folder.

pip install resampy

huggingface-cli download nvidia/NVIDIA-NemotronLabs-RE-USE --local-dir ./NemotronLabs_REUSE --local-dir-use-symlinks False
cd ./NemotronLabs_REUSE

How to use it: Put your noisy speech files in the noisy_audio folder, then run the inference script. The enhanced files appear in the enhanced_audio folder.

sh inference.sh

Tool | What it does
noisy_audio/ | Holds the speech files you want cleaned
inference.sh | Runs the enhancement on those files
enhanced_audio/ | Receives the enhanced speech files
BWE argument | Sets the target bandwidth in the script to turn on bandwidth extension

Example: You drop a recording made on a laptop microphone into noisy_audio and run the script. The folder enhanced_audio then holds a cleaner version of the same speech.

For newcomers and experts: This model removes noise and damage from speech. A first small thing to try is the Gradio demo, where you upload a noisy clip and listen to the result before installing anything.

Under the hood, the model is small, at 9.6M parameters. Input can be sampled at 8 kHz, 16 kHz, 22.05 kHz, 24 kHz, 32 kHz, 44.1 kHz or 48 kHz. With overlapping speakers it keeps the main voice and suppresses the others. The project tests it against larger models on parts of the non-blind URGENT 2025 test data.

Noisy speech files go in noisy_audio
The script runs the enhancement model
Optional bandwidth extension is applied
Enhanced files land in enhanced_audio

Good to know: Scores are mixed across measures, and the model does not win every one. On the full set, Sidon beats it on DNSMOS and UTMOS, and UniPASE beats it on SBERT.

Title: Character accuracy of ASR on the full set
Noisy audio: 81.29
Sidon: 84.68
UniPASE: 87.28
NemotronLabs RE-USE: 89.17

• Laughter: the project suggests a separate RE-USE-laughter model when you want laughter preserved better
• Commercial use: the project states the model is ready for commercial use
• License: OpenMDW 1.1
• Cost to run: none named
• Data it reaches: the audio files in noisy_audio, the enhanced files it writes, and any clips you upload to the hosted demo

On Hugging Face, 13 November 2026

• Task: audio to audio
• Library: mamba-ssm
• Parameters: 9.6M
• Weights: about 38 MB in F32; about 5 MB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 43,030
• Likes: 223
• License: openmdw-1.1
• Quantized versions: 8
• Tags: mamba-ssm, safetensors, speech-enhancement, language-agnostic, audio-to-audio

Links

• Model: https://huggingface.co/nvidia/NVIDIA-NemotronLabs-RE-USE
• Files: https://huggingface.co/nvidia/NVIDIA-NemotronLabs-RE-USE/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download nvidia/NVIDIA-NemotronLabs-RE-USE

Then follow the usage in its model card. Nothing in this kit runs by itself.