Nvidia has released RE-USE, a speech restoration model for reducing noise and reverberation. It has about 9.6 million parameters, 38.4MB of weights and support for 8–48kHz mono audio. In the official evaluation’s 16kHz subset, character recognition accuracy rose from 76.54% to 87.09% after enhancement. It does not lead every metric, however, and improves the main speaker rather than separating multiple speakers. Long recordings need chunking, and deployment should check supported hardware and quality on representative input.
Back to Tech
Nvidia releases a 9.6-million-parameter speech restoration model
RE-USE reduces noise and reverberation, but does not separate speakers