Open yousifa opened 9 months ago
actually there is misinformation in this repository but when I searched for information about the approach and resources they relied on for the model I found this official article, it may help you understand how things work here a little better
It seems that the number of training steps is 870000 and the number of trained samples is 192478464. This second number seems to be the total combinations audio+noise+rir trained, so front ground audios could be repeated. You can check this information creating a script loading and printing pretrained models. Training time depends on hardware architecture, so it's dificult to estimate.
Few questions: