I imagine the audio is transcribed before voice actors speak; they aren't translating on the fly.. it seems to me the natural solution would be to have the transcription the voice actor uses be the same as what's provided to the subtitle generation.

This seems to require two translation attempts; the main benefit being they can be done independently/parallel, but at the cost of duplicating the work. But it doesn't seem to save any work (assuming VAs work off a written transcription/translation, presumably completed by other staff members)