conditionaudiogen checkpoint

Inspired by the practice in the image field which evaluates the location of generated object using a object detection model, we employ a SED system to provide event-based measures (Eb) and clip-level macro F1 score (At) to assess the temporal order control capability. These metrics evaluate the presence of sound events in the generated audio, as well as the onsets and offsets using the first-place SED system in DCASE 2022 Task 4, PB-SED (https://github.com/fgnt/pb_sed/), on AudioCondition test set.

The checkpint of PB-SED that we used for evaluating is publicly available here.