NVIDIA's open RE-USE model cleans noise and echo from speech with just 9.6M parameters
NVIDIA's RE-USE, cleared for commercial use, repairs noise, echo and damaged recordings. In its model card's test it is far smaller than rivals, leads on five of nine measures, trails on three and ties on one.
NVIDIA has released NemotronLabs RE-USE, a model that cleans up speech recordings, on Hugging Face under the OpenMDW 1.1 license. The model card gives October 22, 2026 as the release date, and the company announced it on X on October 30. The card says the model is ready for commercial use. A demo page on Hugging Face lets anyone upload noisy audio or video and listen to the result. As an open model, it can also run on an organization's own hardware.
One model aims at many kinds of damage
RE-USE goes beyond background noise. According to the model card, it also aims to fix reverberation, clipping, limited bandwidth, codec artifacts, packet loss and low-quality microphones. It accepts sampling rates of 8, 16, 22.05, 24, 32, 44.1 and 48 kHz, and can optionally extend bandwidth. Input and output are mono .wav files. When several people talk at once, the model is trained to keep the main speaker and suppress the others, so background conversations in meeting recordings may be removed too. For better laughter preservation, NVIDIA offers a separate laughter version.
The table shows a small model's strengths and limits
The model card compares RE-USE on the URGENT 2025 test set with the 250M-parameter Sidon and the 500M-parameter UniPASE. On the full set, RE-USE posts the top scores on NISQA at 4.33, PESQ at 2.57, STOI at 0.78, speaker similarity at 0.86 and speech recognition character accuracy at 89.17. The damaged recordings score 81.29 on that accuracy measure. Sidon leads on DNSMOS, 3.34 to 3.29, and on UTMOS, 3.21 to 2.94. UniPASE leads on SBERT, 0.89 to 0.88, and the two tie on LPS at 0.85. So despite being far smaller, the model leads on many measures, though it is not the best on all of them.
Training data and method
The model pairs a 30-layer bidirectional Mamba with a convolutional encoder and decoder. Training used about 2,500 hours of speech, 180 hours of noise and about 60,000 simulated room responses. The speech comes from sets such as LibriTTS, VCTK, Multilingual LibriSpeech and CommonVoice, plus NVIDIA's internal data. The languages named on the card are German, English, Spanish, French and Chinese, and the card says the model aims to work independently of language.
The paper, accepted at NeurIPS 2026, highlights three points. For removing reverberation, time-shifted anechoic clean speech works better as a target than speech with early reflections. A two-stage design aims to minimize distortion at a given level of perceived quality. And large uncurated datasets put a ceiling on how well a model can remove subtle artifacts.
Who it suits
NVIDIA is aiming at content creators and researchers who clean speech data. The paper says the method is effective for improving training data for speech synthesis. Old recordings, phone calls, in-car footage and podcasts are natural uses too. For live calls there is a separate Real-time RE-USE model, whose paper describes fitting one model to different latency budgets. For files over 30 seconds, the card suggests a chunked script, because the model was trained on clips under 10 seconds.
Setup is aimed at developers for now. The card lists Linux and NVIDIA Ampere hardware, names the A100 as a reference and suggests Docker for the Mamba environment. It gives no figures for running on a CPU. The comparisons come from NVIDIA's own table. For a fuller picture, follow reviews from independent researchers too.
Sources: Hugging Face (model card), Demo, arXiv paper, Real-time RE-USE
