[0_____0]

The result here seems a bit like what I imagine "Audio Description" for visual media would read like as a transcript. Maybe it would work better if the model was trained on novels, rather than a more general corpus?