A nonverbal cue is useful when it lands in its intended beat. It becomes a problem when a model reads the tag aloud, adds a second reaction, or changes a character's next line. Use a minimal cue test before judging broader acting quality. Resemble documents paralinguistic tags for Chatterbox Turbo.
Build the test input
Write a fictional exchange in which a sleepy librarian briefly chuckles before saying โI'll bring it downstairs.โ Locate the chuckle before the reply, with no other expressive directions. Prepare another copy without a cue. Keep both in your notes alongside a human reading reference or a simple timing sketch.
Run the proposed comparison
Use each model's documented cue syntax only where supported, and evaluate the untagged baseline too. A feature present on one platform need not be exposed by another. Match overall playback level. Mark the onset and duration of nonverbal sound and transcribe the adjacent sentence word for word.
Decide what passes
Reject the requested-cue version if the model audibly pronounces a tag, starts the reaction before the intended moment, or adds a second reaction. Pass if the sound is short enough to leave the full response intelligible. Keep tone, words, and timing in separate score columns.
Act on the outcome
Treat strong spoken output without a good laugh as useful voice material. A separately recorded or generated reaction can be timed in a track while preserving the spoken sentence you already approved. Choose the combined model pass only when it reduces editing without making the line harder to understand.

