AI Video With Sound: How to Prompt Dialogue, SFX & Music

AI video with sound means the model creates the soundtrack together with the picture: dialogue, sound effects, background ambience and music in a single generation. Google's Veo 3.1 and Kling 3.0 both do this. You direct the audio in the prompt by describing each sound in its own sentence, writing speech without quotation marks and keeping lines short enough to fit the clip.

This guide covers the current dialogue syntax for both models, how many words fit in a 4 to 10-second clip, prompts for effects, ambience and music, when to switch sound off to save stars, and how to fix garbled speech or burned-in subtitles. For the visual side of prompting, see our AI video prompts guide.

What native audio means in AI video

Earlier text-to-video models produced silent clips, and you added sound in an editor and lined it up by hand. Native audio models generate picture and sound together, so a door slam lands when the door closes and a character's mouth moves while they speak.

Google's Veo prompt guide asks you to describe audio in separate sentences, and it helps to think in four layers:

• Dialogue: lines spoken by characters on screen.
• Sound effects: discrete sounds tied to actions, such as glass shattering or boots on gravel.
• Ambience: the continuous background of a place, such as rain, traffic or a busy café.
• Music: a background score or a song played in the scene.

Set expectations on speech. Google DeepMind's Veo page notes that natural-sounding speech, short lines in particular, is still in active development, so plan for the occasional re-roll on dialogue clips.

Which AI video generators make audio

At the time of writing (October 2026), native audio is standard on the leading models. All three models in the ToonlyAI studio can generate sound, and it's an optional toggle:

Outside ToonlyAI, Google's newer Gemini Omni Flash, MiniMax H3 and ByteDance's Seedance 2.5 also generate audio with the picture. OpenAI shut down Sora 2 and its video API on 24 September 2026, so older Sora audio tutorials no longer apply. For a side-by-side look at the two model families, read Veo 3.1 vs Kling 3.0.

How to add dialogue to an AI video

Veo 3.1: no quotation marks

Google's current Veo best-practice guide says not to put speech in quotation marks, because quotes can make the model print the words on screen. Instead, name the speaker, add “says” and a colon, then write the line.

This is a change. Google's October 2025 prompting blog recommended quotes, and many templates still copy that style. If your dialogue clips keep showing subtitles, the quotation marks are the first thing to remove.

Medium close-up of an old fisherman in a yellow oilskin coat on a foggy harbour at dawn. Slow push-in. Cold blue light, soft fog. The fisherman says in a slow, gravelly voice: The tide waits for nobody. Gulls cry in the distance. Water laps against the boats.

• Describe the speaker by something visible (the fisherman, the girl in the red scarf) so the model knows whose mouth should move.
• Put the delivery right before the line: whispering, laughing, out of breath, flat and bored.
• Add effects and ambience as separate sentences after the line.

Kling 3.0: speaker, line and delivery together

Kling's prompt guide asks for clear speaker labels, with each speaker's name, line and delivery kept together rather than describing voices in one place and lines in another. That matters most with two characters, which Kling's 10-second clips have room for.

Polished 3D animation: a lemon detective in a tiny trench coat questions a nervous banana across a kitchen counter at night. Static medium shot. Moody film-noir light. Lemon detective, low and unimpressed: Where were you at midnight? Banana, fast and squeaky: In the fruit bowl, I swear! A fridge hums in the background.

Multi-character comedy like this is the engine of fruit drama; our AI fruit drama guide builds a full episode clip by clip.

Keep voices the same across clips

Google's guidance for consistency is to name each character, give them a voice style and repeat the full description unchanged in every prompt. Paste the same sentence, such as “Milo, a small blue robot with a soft, cheerful voice”, into every clip of a series. Our guide to consistent characters covers the visual side.

How many words fit in an AI video clip

People speak at roughly 2.5 words per second. Treat that as a rule of thumb, not a model spec, and aim below it so the line isn't rushed or cut off.

Short lines also land better on social video. One punchy sentence with a reaction afterwards usually beats a paragraph squeezed into eight seconds. If the script needs more, split it across clips; how to make an AI video from text shows how to stitch clips into a longer story.

Step by step: a talking clip in ToonlyAI

1. Open the Video Studio and choose Text to Video, or Image to Video if you already have a character image.
2. Pick a model: Veo 3.1 Lite for cheap drafts, Veo 3.1 Fast for a 1080p final, Kling 3.0 when you need 10 seconds for an exchange.
3. Set 9:16 for TikTok, Reels or Shorts and choose a length that fits your line.
4. Turn sound on. Write the scene first, then the line in the says-colon format, then effects and ambience.
5. Check the star cost, generate (usually 1–3 minutes) and download the MP4.
6. If the speech is off, change one thing and run it again.

AI sound effects and ambience prompts

Google's guide recommends describing audio in separate sentences, with effects as discrete sounds and ambience as the continuous background. Three habits make effects land:

• Tie each effect to a visible action: “As the lid snaps shut, a sharp click.” That gives the model a moment to sync to.
• Name the material and distance: boots crunching on gravel, a distant church bell, a close, crisp page turn.
• Keep it sparse: one ambience bed and one or two effects. A crowded mix turns to mush in a 6-second clip.

Close-up of a ceramic mug on a wooden café table as a barista pours steaming coffee into it. Static shot, warm morning light, shallow depth of field. Coffee splashes into the cup, then a spoon clinks against the rim. Quiet café chatter and a hissing espresso machine in the background.

A wooden door at the end of a dark hallway slowly swings open on its own. Slow push-in. Cold moonlight, dust in the air. The hinges creak long and low. Wind whistles through the empty house, and everything else is silent.

How to prompt music in an AI video

Describe music by genre, mood, instruments and tempo, and say whether it belongs to the scene or sits on top as a score. “A street busker plays a cheerful accordion tune” is in the scene; “soft piano score, slow and hopeful” is background.

• Keep music low under dialogue, or leave it out of talking clips altogether.
• Don't name real songs, artists or bands. Describe the sound instead: upbeat ukulele, lo-fi beat, brass fanfare.
• If you'll post with a trending sound, skip generated music; it will clash with the track you add later.

When to switch audio off (and save stars)

On ToonlyAI, sound is a toggle and the studio shows the star cost before you generate. Silent clips are cheaper on every model:

Switch sound off when:

• You'll post with a trending sound or a track from the platform's own library.
• You'll record one voiceover across several stitched clips, so the narration stays in one voice.
• You're still drafting framing and motion. Get the picture right silent, then generate the final with sound.
• The clip is B-roll or a montage beat with no on-screen speech.

Keep it on when a character speaks on camera, when a sound is the joke (a squeak, a crash, a perfectly timed thud), or when the clip will be posted on its own.

One thing to be clear about: ToonlyAI doesn't have a lip-sync or voice-cloning tool. If a character needs to speak on screen, generate the line with sound on in the same clip. A voiceover added afterwards works as narration, but it won't move anyone's mouth. Our talking object guide compares both approaches.

Fixing garbled speech, wrong speakers and burned-in subtitles

Change one thing at a time and keep the rest of the prompt identical, so you can tell what fixed it.

ToonlyAI refunds failed generations automatically. For dialogue, it pays to draft on a 4-second Veo 3.1 Lite clip with sound (25 stars) before spending 120 stars on an 8-second Veo 3.1 Fast version.

Music, voices and copyright

Generated audio is still new territory legally, so keep it simple. This isn't legal advice.

• Don't imitate real people. Prompting a celebrity's voice, or making a real person appear to say something they didn't, can break platform rules and invite right-of-publicity claims. YouTube requires disclosure for realistic AI depictions of real people; see our guide to monetizing AI videos on YouTube.
• Don't name songs or artists in music prompts. Describe genre, mood and instruments.
• Ownership is limited. The US Copyright Office has said prompts alone aren't enough to make you the author of AI output, though human selection, arrangement or editing can be protected. Check your generator's terms for commercial use.
• Want a trending track? Add it inside TikTok, Instagram or YouTube when you post, rather than trying to recreate it in a prompt.