NVIDIA Releases RE-USE Audio Enhancement AI

NVIDIA released NemotronLabs RE-USE on October 22, 2026, to correct noise, echo, clipping, and other flaws in recorded audio. The compact model has about 9.6 million parameters and is designed for multilingual audio processing. It can improve recorded material and prepare training data for audio-generation AI, and commercial use is permitted.

RE-USE is a Universal Speech Enhancement model, a technology for correcting a wide range of audio degradation while preserving the original sound. It addresses background noise, room echo, clipping from excessive input, limited frequency range, compression damage, packet loss during communication, and reduced quality from low-grade microphones. It supports input sampling rates of 8, 16, 22.05, 24, 32, 44.1, and 48 kHz. Input and output use mono WAV format.

For audio with overlapping voices, the model is trained to retain the main speaker and suppress other speakers; it is not designed to preserve everyone in a conversation equally. NVIDIA released NemotronLabs RE-USE-laughter on the same day, a version that preserves laughter more during enhancement. It also has about 9.6 million parameters.

Both models are available under the OpenMDW-1.1 license, which NVIDIA says permits commercial use. Intended uses include improving recorded audio for content creators and improving datasets for training speech synthesis and Speech LLMs. The earlier RE-USE was released on March 18, 2026, under a noncommercial license limited to research and development. Its release page also points to the NemotronLabs version for commercial use.

Model weights and inference code are available on Hugging Face, along with a browser demo where users can upload or record audio. The demo lets users compare audio before and after enhancement and inspect changes in a spectrogram, a visual display of sound frequencies.

NVIDIA released NemotronLabs RE-USE on October 22, 2026, to correct noise, echo, clipping, and other flaws in recorded audio. The compact model has about 9.6 million parameters and is designed for multilingual audio processing. It can improve recorded material and prepare training data for audio-generation AI, and commercial use is permitted.

RE-USE is a Universal Speech Enhancement model, a technology for correcting a wide range of audio degradation while preserving the original sound. It addresses background noise, room echo, clipping from excessive input, limited frequency range, compression damage, packet loss during communication, and reduced quality from low-grade microphones. It supports input sampling rates of 8, 16, 22.05, 24, 32, 44.1, and 48 kHz. Input and output use mono WAV format.

For audio with overlapping voices, the model is trained to retain the main speaker and suppress other speakers; it is not designed to preserve everyone in a conversation equally. NVIDIA released NemotronLabs RE-USE-laughter on the same day, a version that preserves laughter more during enhancement. It also has about 9.6 million parameters.

Both models are available under the OpenMDW-1.1 license, which NVIDIA says permits commercial use. Intended uses include improving recorded audio for content creators and improving datasets for training speech synthesis and Speech LLMs. The earlier RE-USE was released on March 18, 2026, under a noncommercial license limited to research and development. Its release page also points to the NemotronLabs version for commercial use.

Model weights and inference code are available on Hugging Face, along with a browser demo where users can upload or record audio. The demo lets users compare audio before and after enhancement and inspect changes in a spectrogram, a visual display of sound frequencies.