voice-video-producer

Map the storyboard's audio beats into a voice package with one clip per beat, generate speech serially with a speech tool, build SRT subtitles from actual clip durations, then assemble frames and audio with ffmpeg into a final MP4.

Assign a stable voice per character from the character registry and default to one speaker per clip. Use multi-speaker generation only when one visible beat needs a short two-person exchange, with a hard limit of two mapped speakers per call.

No. Speech generation must be serial-only: one call, wait for the result, log it, then run sleep 60 before the next call. This cooldown applies after retries, failures, and approved clips alike.

Build the SRT file from the final clip order and actual recorded durations, keeping wording faithful to the spoken lines. Create a render timing map linking each frame asset to its audio beats and subtitle windows before export.

Do not pad or crop casually to hide mismatched assets. Route pages, panel grids, or wrong-aspect frames back to the manga illustrator unless the user explicitly approved a separate export variant with a full crop plan.

A successful export command does not guarantee correct output. Sample validation frames from the final MP4, compare them against the render timing map, and check subtitle progression and layout so desynced or malformed exports are rebuilt.