Most speech-cleanup models are big. NVIDIA’s new NemotronLabs RE-USE is not: it has 9.6 million parameters, yet its model card shows it matching or beating open models that are 26 to 50 times larger on several measures. NVIDIA researchers promoted the open model on October 30, 2026, according to GIGAZINE.
What it does
According to NVIDIA’s Hugging Face model card, RE-USE is a multilingual universal speech enhancement model. It restores speech damaged by noise, reverberation, clipping, bandwidth limits, codec artifacts, packet loss and low-quality microphones. It accepts audio from 8 kHz to 48 kHz, is language-agnostic, and, when several people talk at once, preserves the main speaker while suppressing others.
Small model, strong scores
The model card compares RE-USE with other open models on the URGENT 2025 non-blind test set, measured against clean reference speech. On the full set, RE-USE (9.6M parameters) scored PESQ 2.57 against 2.17 for UniPASE (500M) and 1.73 for Sidon (250M); speech-recognition character accuracy on the cleaned audio reached 89.17 percent versus 87.28 and 84.68, and speaker similarity was 0.86 versus 0.81 and 0.74. Sidon edged it on DNSMOS and UTMOS, so it is not a clean sweep. At 44.1 kHz it led every listed metric against Resemble Enhance. These are NVIDIA’s own numbers, so independent testing will matter.
Under the hood and who can use it
The architecture is a convolutional encoder and decoder with a 30-layer bi-directional Mamba block. Training used roughly 2,500 hours of speech from public corpora such as LibriTTS, VCTK, Multilingual LibriSpeech and Common Voice, plus an internal dataset, mixed with simulated noise and room responses.
NVIDIA lists two use cases: content creators improving recordings, and researchers cleaning datasets for text-to-speech and speech language models. It is released under the OpenMDW-1.1 license, marked ready for commercial use, with a browser demo on Hugging Face and a laughter-preserving variant. Limits: it is tuned for clips under 10 seconds (a chunked script handles long files), targets NVIDIA Ampere GPUs on Linux, and GIGAZINE noted a user observing that it struggles with extremely noisy audio.
Practical takeaways
- Podcasters and video makers can try the free demo before installing anything.
- Voice-AI teams can clean training data cheaply with a model this small.
- Test on your own audio; benchmark gains don’t guarantee results on extreme noise.
Bottom line
RE-USE shows that high-quality speech cleanup no longer needs a giant model. If the numbers hold up independently, small open models could make clear audio a default.