NVIDIA has published NemotronLabs RE-USE on Hugging Face, an open 9.6-million-parameter model that removes background noise, reverberation, distortion and compression artifacts from voice recordings. The model is ready for commercial use and is distributed under the OpenMDW 1.1 license.
The task is universal speech enhancement: a single network that restores a voice degraded by several causes at once. RE-USE covers additive noise, reverberation, clipping, bandwidth limitation, codec artifacts, packet loss and low-quality microphones. It accepts sampling rates from 8 to 48 kHz and, according to the model card, is language-agnostic. In overlapping speech it aims to preserve the main speaker and suppress the others.
Image: namiru.ai — https://namiru.ai/blog/nvidia-releases-massive-collection-of-open-models-data-and-tools-to-accelerate-ai-development
The architecture combines convolutional encoder and decoder layers with 30 layers of bidirectional Mamba for time-frequency modeling. Training data is roughly 2,500 hours of synthetic, single-channel, single-speaker audio at 8 kHz. A separate variant, RE-USE-laughter, is meant to better preserve laughter.
Comparison with other open models
In NVIDIA's own table, on the full URGENT 2025 non-blind test set against anechoic clean speech, RE-USE scores 2.57 on PESQ, 0.78 on STOI, 0.85 on LPS, 0.86 on speaker similarity and 89.17% character accuracy, against 3.21 and 3.34 on PESQ for UniPASE and Sidon, which have 500 and 250 million parameters. On UTMOS, RE-USE reaches 2.94, below Sidon's 3.21. On NISQA it scores 4.33, against 4.03 for UniPASE.
Results vary with sampling rate. On the 16 kHz subset, RE-USE reaches 87.09% character accuracy against 54.82% for UniSE, while UniSE leads on UTMOS (2.77 against 2.64). At 44.1 kHz it beats Resemble Enhance on nearly every metric, with 91.54% character accuracy. At 48 kHz it scores 3.63 on UTMOS, against 3.01 for ClearVoice.
The figures come from NVIDIA's tests. A commentary on alphasignal.ai summarizes the same picture: Sidon scores higher on some perceptual predictors such as UTMOS and DNSMOS, while RE-USE leads on PESQ, STOI, speaker similarity and recognition results.
Why it matters
The model card names two main uses: improving recordings for content creators, and cleaning speech datasets used to train text-to-speech models and speech large language models. NVIDIA released RE-USE in a series of speech-related launches: on 28 October it published Nemotron 3.5 Transcribe, an open 30-billion-parameter model that transcribes in 40 languages and distinguishes up to eight speakers, alongside version 2 of the Granary dataset.
Image: LinkedIn — https://www.linkedin.com/pulse/nvidia-unveils-nemotron-3-new-era-open-enterprise-ready-fxjqc
For developers, the concrete point is scale: a model this size can run on modest hardware. The repository lists NVIDIA Ampere GPUs and Linux as supported environments. For clips longer than 30 seconds, or when GPU memory runs out, it provides a script that splits the audio into chunks, with configurable chunk size and overlap.
How to use it
The model can be tried immediately in the Gradio demo on Hugging Face by uploading an audio or video file. For local use, the repository provides inference scripts: noisy files go into noisy_audio/ and results are saved to enhanced_audio/. Bandwidth extension, which is optional, requires the resampy package.
No comments yet. Be the first!