Training real-time data augmentation

DeepSpeech
#1

Hello,
From what I can see from the current code, data is loaded at the beginning of training, then features are created and this dataset is kept in memory until the end of training. Is it possible to implement real-time data augmentation, that every time file is feed to the model during training I can apply random augmentation on it?

#2

I was thinking the same, any progress in this?
Real time data augmentation would be quite beneficial

#3

I’ve been working on a tf.data input pipeline that should allow online augmentation. I expect to merge it in the next couple of weeks.

1 Like
#4

Thank you!!!
I am looking forward to that update!

#5

Thanks for reply. This is amazing news! What kind of augmentation will be possible? Will it be possible to mix training data with noise data in runtime?