Can finetune whisper-large-v2 with data augmentation SpecAugment, Stochastic Depth, and BPE Dropout?

[maowandong]

I'm fine-tuning Whisper and noticed that Whisper-large-v2, compared to other versions of Whisper, has been trained for 2.5X longer. It also includes data augmentation techniques such as SpecAugment, Stochastic Depth, and BPE Dropout. I wonder, doesn't this extended training time of 2.5X lead to overfitting? Also, are these data augmentation techniques suitable for use during fine-tuning? If they can be used, are there any recommended open-source implementations that I can refer to?

[sky-lc]

Continuous attention

[mizoru]

Yes, too bad the implementations have not been shared