Film sound splits on whether a person in the picture could have heard it. Dialogue on set, footsteps, a phone in shot, a song playing from a car radio: diegetic. Score, a narrator, a whoosh you added on the cut: not. Native-audio models bake both kinds into the same generate unless the prompt says which is which. Leaving the field blank is how you get a sung whoosh, or a presenter whose voice is in the room and also floating over it.
SDH is how a mute viewer gets the diegetic layer that is not speech: [door slams], [music in bar], [phone buzzes]. A speech-only caption track transcribes the line and drops the knock the joke is built on. Voiceover is the other door — a read under picture that nobody in the scene is saying. Mixing them is how a product demo grows a host it did not need, or a VO gets labelled as if the bottle were talking.
When the model owns the bake, write the in-world sounds as events in the prompt and keep score and VO as separate files you can slip. Replacing baked diegetic audio later is an edit, not a toggle. If the brief is silent turbo plus a designed mix, you are authoring diegetic and non-diegetic on a timeline, which is the slower path and the one that still lets you mute the radio independently of the narrator.
In practice
- Name in-world sound in the prompt as events in the room, not as a mood ('cinematic audio').
- If a mute viewer must get a non-speech beat, write it into SDH. Auto captions will not invent [glass breaks].
- Keep score and voiceover off the native bake when you will need to recut them.
The mistake to avoid
Calling every soundtrack diegetic because it came out of the video model. Native audio is a bake. Diegetic is whether the scene contains the source.
Where you will run into it
- Add a Voiceover to a Video — Type the script. Get a narrated video back.
- AI Video Generator — Text-to-video, image-to-video, and story chaining in one studio.
Related terms
Native audio
Native audio meaning: a video model that generates soundtrack (dialogue, effects, ambience) in the same pass as the picture.
SDH
SDH is a caption track that includes speech and non-speech audio — [door slams], [music], speaker IDs — so the file is usable without sound.
Voiceover
A voiceover is a spoken track laid under picture, rather than lipsync that drives a mouth or a caption track that visualises speech.
Text-to-speech
Text-to-speech converts written text into spoken audio using a synthetic voice you choose before generating.
Voice cloning
Voice cloning meaning: building a reusable synthetic voice from a real sample so new scripts can be spoken in that voice later.
The all-in-one AI studio for creators. 60+ models for video, image, voice, music and lipsync, in your browser or on your phone.
Free on the web and on iPhone, one account.