If that’s the case, the problem is either in the loop outside the transcribe function, or is it this…?
arr.to("cpu") # offload before deleting
del arr, audio_data
gc.collect()
torch.cuda.empty_cache()
1 Like