Closed pandysong closed 1 year ago
Yes exactly. We don't have results on real recordings. We would ideally need a training set generated in a real environment and a corresponding test set. Do you have any in mind?
I have no such recording for training. Data argumentation might be good enough to add noise or reverberation on to the clip?
But before that I am curious about the performance of the current available model on a real recording (which I do not have as well)
Thanks for your update.
I see. For now I don't have any test data recorded in real conditions. This would be something to do in the future.
If I could capture recording data for a mic array, e.g. 4 Mic in a circular pattern. Would it help?
Sure. For now the network has been trained to perform on 2 speech sources, so you could record two sources at different direction of arrival. Thanks!
If I understand correctly, the sample clip is generated uinsg librispeech.
I wander if there are test result on a real recording clip?