Oct 22, 2026
NVIDIA releases a 9.6-million-parameter multilingual model for speech enhancement, with downloadable weights and an interactive demo. It targets noise, reverberation, clipping and other recording distortions.
The model card dates release to October 22; NVIDIA promoted it on October 30. OpenMDW-1.1 permits commercial use subject to its terms. The model supports multiple sample rates; with overlapping speech, it is trained to preserve the main speaker while suppressing others.