How can I preserve speaker turns and cue boundaries in short rap translations?

I am testing localization for very short two-person clips. The source is a two-photo AI rap video workflow that generates original English vocals in 5, 10, or 15 second scenes. Each line is assigned to one of two speakers and aligned to a beat window.

If I send the lyrics through the MotaWord API, is there a recommended way to keep speaker labels and line breaks from being merged during translation? Ideally the response would preserve each cue as an independent segment so the translated text can be reviewed against the original timing.

Would you recommend one project per clip with a structured JSON file, or a subtitle format such as WebVTT with speaker notes? I am especially interested in whether translators can see the timing constraint without altering the cue boundaries.