NVIDIA-NemotronLabs-RE-USE: Universal voice enhancement up to 48kHz with 9.6M parameters

NVIDIA NemotronLabs RE-USE is a multilingual universal voice enhancement model composed of 9.6 million parameters. It simultaneously handles various real-world degradation factors, including added noise, reverberation, clipping, bandwidth limitations, codec artifacts, packet loss, and low-quality microphones. It is not limited to specific languages, uniformly restoring voice quality across various languages.

It supports seven input sampling rates from 8kHz to 48kHz, accepting mono-channel audio in .wav file format and outputting enhanced audio in the same format. The model architecture consists of a convolutional encoder and decoder, along with 30 layers of bidirectional Mamba for time-frequency modeling. For long audio exceeding 30 seconds or situations with insufficient GPU memory, it can be processed using a chunk processing script.

Based on the URGENT 2025 non-blind test set, it outperforms existing open-source models such as Sidon with 250M parameters and UniPASE with 500M parameters on key metrics like PESQ, STOI, and speaker similarity. In particular, it recorded higher scores than Resemble Enhance and ClearVoice in the 44.1kHz and 48kHz subsets, demonstrating performance in the high-fidelity range. It is available for commercial use and is optimized to run on GPUs based on the NVIDIA Ampere architecture.

Content creators can use it to improve the quality of recorded voice, while researchers can utilize it to refine datasets for training text-to-speech (TTS) or voice LLMs. If more laughter preservation is needed, use the separately provided RE-USE-laughter model, which is trained to preserve the primary speaker and suppress other speakers in overlapping speech.