I want to use onsets_frames_transcription for serving, but the preprocessing of the audio example proto is related to the data.provide_batch, and it returns a dataset object