Two people can have believable expressions and still exchange voices at the wrong moment. Treat speaker assignment, word accuracy, and handoff timing as three separate observations before judging whether a generated dialogue shot is ready to edit.
Build the test input
Set a landlord and a tenant on opposite sides of a kitchen counter. Give them the short lines โDid you leave the key?โ and โIt's in the blue bowl.โ Put a pause between lines. Both characters stay visible so a reviewer can see whose mouth moves during the first question.
Run the proposed comparison
Use the same text and visual reference across eligible audio-enabled operations. If one interface cannot take a required reference, put it in a separate track instead of silently dropping the reference from only that run. Save the exact line text with the export and review sound at the same volume.
Decide what passes
Transcribe both lines before checking faces. Then mark any oral movement before the second speaker's line begins. Require both words and speaker order to match. Flag a pause too short to separate the responses as a timing failure even if all words appear somewhere in the clip.
Act on the outcome
A clip with strong expressions but incorrect lines may be usable as a silent reaction. A clip where the speaking face stays turned away can support an editorial cutaway. A reliable multi-shot workflow may be cheaper than repeatedly seeking flawless close dialogue within one generated clip.

