We share a link to denoiser's github, Facebook's real-time noise reduction technology that was announced at Interspeech 2020.
It is implemented in Pytorch and the title of the original paper is “Real Time Speech Enhancement in the Waveform Domain”. As you can see from the title, the main feature is that speech enhancement is performed directly in the waveform area, and it is said to have been developed with a relatively low computational capacity to enable real-time processing even in a notebook CPU.
In particular, it seems that it can handle not only stationary noise but also non-stationary noise, so if it is used in the preprocessing part of a speech recognizer, performance improvement can be expected. The basic structure uses Demucs, which was proposed for music source-separation. I also share a link to Demucs' github below.
