It has been done before, for eg the original RNNoise was trained on proprietary data, later there was crowd-sourced effort to record new data and have it under

It has been done before, for eg the original RNNoise was trained on proprietary data, later there was crowd-sourced effort to record new data and have it under libre licenses. https://github.com/xiph/rnnoise/ https://github.com/xiph/rnnoise/