NVIDIA NemotronLabs RE-USE-laughter

NVIDIA NemotronLabs RE-USE-laughter

Cleans noisy speech recordings in many languages while keeping laughter. A 9.6M parameter enhancement model for creators who fix audio and researchers who prepare speech datasets.

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: This is a small speech enhancement model from NVIDIA with 9.6 million parameters. It restores the quality of speech recordings across many languages, and it keeps laughter in the result. It copes with noise, echo, clipping, codec damage, packet loss and low-quality microphones.

Who it suits: Content creators who want cleaner recordings, and researchers who clean speech datasets before training speech models.

Industry | Who uses it | How they use it | Example
Media | Content creators | Improve the sound of spoken recordings | Clean a noisy interview before editing
Speech research | Researchers | Enhance datasets used to train text-to-speech models | Prepare noisy clips for a TTS training set
Speech research | Researchers | Enhance data for speech language models | Clean recordings before training a speech LLM

If your audio is already clean, you can skip it.

Where to use it:

• Your own computer, from a command line, on Windows, macOS or Linux with Python
• A Docker setup for Mamba, which the card suggests to simplify installation
• The Gradio demo the card points to, where you upload noisy audio or video

Use case | Who | What they get back
Clean up a recording | Creators | A cleaned version of each speech file
Prepare a dataset | Researchers | Enhanced clips ready for training
Process long files | Anyone | Cleaned audio from long recordings, handled in chunks

How to set it up:

• Python and the Hugging Face download tool
• Mamba support, which the card says is easier with a pre-built Docker environment
• The resampy package, only if you want bandwidth extension

1. Set up Mamba, for example with the Docker environment the card points to.
2. Optionally install resampy for bandwidth extension.
3. Download the repository and move into the new folder.

pip install resampy

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

How to use it: Drop the noisy recordings into noisy_audio, run the inference script, and pick up the cleaned versions from enhanced_audio. Bandwidth extension is switched on through the BWE argument in the script.

sh inference.sh

For files longer than about 30 seconds, or if the GPU runs out of memory, use the chunked script. Put long files in long_noisy_audio, and the results land in Long_enhanced_audio.

sh inference_chunk.sh

Tool | What it does
inference.sh | Enhances the files in noisy_audio
inference_chunk.sh | Enhances long files in chunks
BWE argument | Sets the target bandwidth for extension

Example: You copy a noisy phone recording into noisy_audio and run the inference script. A cleaned version of the file appears in enhanced_audio, with laughter kept in.

For newcomers and experts: If you are new, try the Gradio demo the card points to first. Upload a noisy clip and listen to the cleaned result before you set anything up on your machine.

For experts, it accepts input at seven sampling rates, from 8 kHz up to 48 kHz, including 22.05 and 44.1. It was trained on clips under 10 seconds, which is why long files go through the chunked script. Where speech overlaps, it keeps the main speaker.

You place noisy files in noisy_audio
The script runs the enhancement model
Bandwidth extension is optional
Cleaned files land in enhanced_audio

Good to know: On the card's comparison, this model scored 89.21 on character accuracy of speech recognition, higher than the two larger models listed.

Title: Character accuracy of ASR on the card's test set
Noisy input: 81.29 points
Sidon 250M: 84.68 points
UniPASE 500M: 87.28 points
RE-USE-laughter 9.6M: 89.21 points

• License: OpenMDW 1.1, as the card states it.
• Cost to run: no API key or paid service is named. A GPU helps, as long files can run out of memory.
• Data it reaches: it works on the audio files you place in its folders. The card names no web access.
• Not for: the card says the project is not accepting contributions.

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: 2,845
• Likes: 17
• License: openmdw-1.1
• Quantized versions: 3
• Tags: mamba-ssm, safetensors, speech-enhancement, language-agnostic, audio-to-audio

Links

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

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