Binary noise classification using CRNN mean teacher algo [R]

Hi all, so im making a binary noise classification model. I have training data with timestamps of noise and some with verified some with unverified. The performance parameters are F1 and Dice scores and their combined score. I am following AST backbone (Stage 1 ast frozen and stage 2 ast unfrozen along with BiGRU/CRNN head in stage 1 and then the mean teacher in stage 2).

So the sound event detection model is primarily struggling with precise temporal boundary localization, where the majority of errors stem from onset and offset timing mismatches rather than missed detections. This issue was heavily compounded during semi-supervised training by an excessively aggressive consistency weight schedule, which forced the model to rigidly match a smoothed teacher network and eroded the sharp event edges past the early epochs, 14% variation was permitted but it clearly exceeded that.

Please suggest what i should do. I divided the audios into 5s window size and 8 batch size.

I want the combined score F1+ Dice to reach atleast 1.5,

current- Event F1 of 0.284, a Dice score of 0.635, and a Combined score of 0.919.

Any other algorithms if you could suggest but i would prefer continuing this one.