To get sound in an AI video, write it into the prompt as its own lines: who speaks, then a colon and the exact words they say, then the sound effects, then the background noise, then any music, each tied to the moment it should happen.
This only works on video models that generate audio, such as Veo 3.1, Kling, Seedance and Grok Imagine. If yours does not, make the clip silent and add the sound in an editor.
- Four layers: dialogue, sound effects, ambience (the background), and music. Write each one on purpose, or the model fills the gap with its own guess.
- Name the speaker, add a colon, then the words, with no quotation marks:
A woman says: My name is Clara.Google’s newest Veo guide asks for this so the words are not written on screen, and Kling’s guide writes dialogue the same way. Some older Google pages still show quotation marks. - Tie each sound to an action or a time. “The door slams as she leaves” beats “door slam sound”.
- Say what you do not want: “no background music”, “no dialogue”. Music you did not ask for is baked into the clip.
- Keep lines short. A clip is a few seconds long. One or two short sentences is usually all that fits.
- No audio in your model? Generate silent video and add effects, voice and music afterwards.
A film’s soundtrack is built from layers that a sound team mixes together: the actors’ lines, the footsteps and doors, the hum of the room or the street, and the score. A video model that makes sound is trying to do the whole team’s job from one paragraph. Give it the layers separately and it has far less to guess.
Which AI video models make their own sound?
The plain point: several of the major video models now generate sound together with the picture, often called native audio. The details differ, and some of these features were still in preview or early access on 12 October 2026.
| Model | Sound | What the docs say |
|---|---|---|
| Veo 3.1 (Google)preview models in the Gemini API | Dialogue, effects, ambience; always on in the Gemini API | Google lists native audio as always on for its Veo 3 models and gives a prompt syntax for speech, effects and background. 1 Google says the Veo 3.1 previews in the Gemini API were shut down on 22 October 2026, with gemini-omni-1.1-flash as the replacement (read 3 November 2026). 23 Veo 3.1 stays generally available on Google Cloud’s Gemini Enterprise Agent Platform. 4 |
| Kling 3.0 / 4.0 (Kuaishou)4.0 launched 15 Oct | Dialogue, ambience, effects, lip-sync | Kling says 3.0 has native audio in a single channel; 4.0, which Kling says rolled out to all users on 15 October 2026, moves to stereo and adds languages. 5 |
| Seedance 2.5 (ByteDance) | Dialogue, ambience, effects, music | ByteDance describes it as a joint audio and video model, and its prompt template ends with an audio line. 6 7 |
| Grok Imagine Video (xAI)API | Audio track by default; lip-synced speech on 1.5 | xAI says videos come with sound unless you switch it off. 8 |
Other models may make sound too, and features move quickly. If your tool’s documentation does not mention audio, assume the clip will be silent. Model-by-model syntax is on the Veo, Kling, Seedance and Grok Imagine cheat sheets.
Check the vendor’s page before you write a single line of dialogue. On a model with no audio, a quoted line does nothing, and on some models sound is on unless you turn it off.
The four layers of a sound prompt
The plain point: write dialogue, effects, ambience and music as separate pieces. Google’s Veo documentation sets out the first three with a label for each. 1
Google’s Veo 3.1 prompting guide shows the labels written straight into the prompt, as SFX: and Ambient noise:. 9
For dialogue, Google’s own pages now disagree. Its newest Veo guide, updated on 9 October 2026, says to put a colon after the speaker and leave out quotation marks, because quotation marks can make the model write the words into the picture as text. Its example is “A woman says: My name is Clara.” 10 Its Gemini API page (updated 8 October 2026) and its October 2025 prompting guide still put spoken words in quotation marks. 1 9 This page uses the colon style: it is Google’s most recent advice, it gives a reason, and Kling’s guide writes dialogue the same way (see below). Reasoning, 13 Oct 2026 — choosing the colon style is this site’s own reading of the three Google pages above; Google has not said which page replaces which.
- Dialogue: the name or description of who speaks, a colon, then the exact words, without quotation marks.
- Sound effects: the specific sounds that belong to actions on screen. Footsteps on gravel, a cup set down on a saucer, a car door.
- Ambience: the bed of sound under everything. Rain on a window, a busy cafe, wind in tall grass, the hum of a fridge. It is what makes a scene sound like a real place.
- Music: say whether you want it at all. If you do, say the style, the instruments and whether it rises, holds or ends. If you do not, say “no music”.
Four layers, four lines. If you leave a layer out, the model decides it for you, and music is the layer it most often adds when nobody asked.
How to write dialogue: who speaks, and what they say
The plain point: say who is speaking, how they say it, and the exact words. With two or more people, pair each line with its speaker so the model knows whose mouth to move.
Kling’s 3.0 guide shows this pattern: the character, then the way they speak in brackets, then the line. 11
- Name or describe each speaker the same way every time. “The man in the grey coat” in the description and “the man in the grey coat says” in the line. A new description can read as a new person.
- Give the delivery, briefly. “Whispering”, “laughing”, “tired”, “firmly”. This is the same direction a voice actor needs; more on that on prompting music and voice.
- Keep it short enough to say in the clip. Read the line aloud with a timer. If it does not fit comfortably in the clip length, the model may rush it, cut it off or speed up the mouth.
- Name the language or accent if it matters. Kling says 3.0 speaks five languages and translates others into English. 11
- Spell hard names the way they sound. Models mispronounce names, brands and acronyms just as text-to-speech voices do.
Speaker, delivery, a colon, then the words, with no quotation marks. One short line per person per clip is the safe default.
Effects, ambience and music: tie each sound to a moment
The plain point: a sound attached to an action lands in the right place; a list of sounds floats. ByteDance’s Seedance guide puts it as a rule. 7
The same guide asks you to state the sound you want rather than hoping for “good sound”, including telling the model what to leave out. 7
Saying no matters more for sound than for pictures, because sound generated with a clip is mixed into it. Stray music cannot easily be pulled out afterwards, which is also a point made on how AI films are made. If you plan to add your own music in the edit, ask the model for “no music” every time.
Timing sound across several shots
Google’s Veo 3.1 guide shows a timestamp method: split the clip into timed sections and give each its own picture and sound. 9
Attach every sound to something that happens, and say “no music” unless you want the model’s music in your final cut.
What to do when your video model has no audio
The plain point: generate the picture silent, then build the sound separately: a voice from a text-to-speech tool or a real recording, effects from a sound-effects generator or a library, and music from a music tool. You then line them up in any video editor. Working this way is slower, but you can change one layer without regenerating the whole clip.
- Switch generated audio off when you plan to replace it. xAI’s API has
generate_audio=False, and Runway’s API has an audio switch for the Veo 3.1 it hosts. 12 - Sound effects: a sound-effects generator turns a text description into a short clip. ElevenLabs’ documentation is one example. 13
- Voice: a text-to-speech voice, or a real person’s recording. Direction for synthetic voices is on prompting music and voice and the ElevenLabs cheat sheet. To make a face move to a separate voice track you need a lip-sync tool; see how to make a talking photo or lip-sync video.
- Music: structure first, mood last, and ask for a real ending. The method is on prompting music and voice and the Suno cheat sheet.
One more reason to keep a silent version: Google says Veo 3.1 sometimes refuses a clip because of its audio. 1 If a prompt with dialogue keeps failing, try it without the spoken line and add the voice afterwards.
Silent picture plus separate sound is the slow, controllable route. Use it when your model has no audio, when you need a specific voice, or when the generated sound keeps getting in the way.
Real voices: consent and labels
The plain point: making a video character speak in a real person’s voice, or making a real person appear to say something, is a question of consent and disclosure before it is a question of prompting. The law and the dates are on voice and likeness law, and the platform labelling rules are on how to make a talking photo with consent. Background: voice cloning and deepfake.
Questions people ask
Why does my AI video have music I did not ask for?
Because you left the music layer empty and the model filled it. Say “no music” or “ambience only”, as ByteDance’s Seedance guide advises. 7
How do I make an AI video character say exact words?
Name the speaker, add a colon, then write the exact words, on a model that generates speech. Google’s newest Veo guide gives the example “A woman says: My name is Clara.” and says to avoid quotation marks so the words do not appear as text in the video. 10 Google’s Gemini API page still says “Use quotes for specific speech”, so if the words turn up written in the picture, switch to the colon style. 1
Can I turn the sound off?
On some models. xAI’s API takes generate_audio=False, while Google’s Gemini API lists Veo 3 audio as always on, so there you would remove the track in an editor. 8
QUICK CHECK · 3 QUESTIONS
Pick an answer to see if you got it. Nothing is recorded.
1. Which four layers does the page say to write into a sound prompt?
- Volume, speed, pitch and length
- Dialogue, sound effects, ambience and music
- Intro, middle, ending and credits
Show the answer
Answer: Dialogue, sound effects, ambience and music. The page says to write each layer on purpose, or the model fills the gap with its own guess.
2. Why does the page say to write no music when you do not want it?
- Music the model adds is mixed into the clip and is hard to pull out later
- Music makes every clip longer
- Video models are not allowed to make music
Show the answer
Answer: Music the model adds is mixed into the clip and is hard to pull out later. The page says music is the layer models most often add uninvited, and sound generated with a clip is baked into it.
3. What does the page suggest if your video model makes no sound?
- Put the dialogue in quotation marks anyway
- Give up on sound for that clip
- Make the clip silent and add voice, effects and music in an editor
Show the answer
Answer: Make the clip silent and add voice, effects and music in an editor. The page says to generate the picture silent and build the sound separately, which also lets you change one layer without regenerating.
Related on this site: anatomy of an AI video prompt · first frame and last frame · text-to-video · multimodal · What Is Lip Sync in AI?
- 1Google AI for Developers, Generate videos with Veo 3.1 in Gemini API, last updated 8 Oct 2026, read at source 12 Oct 2026, re-read 13 Oct 2026“Natively generates audio with video.”, listed as “Always on”, and “You can provide Veo with cues for sound effects, ambient noise, and dialogue.” ↑a“Dialogue: Use quotes for specific speech.”, “Sound Effects (SFX): Explicitly describe sounds.” and “Ambient Noise: Describe the environment’s soundscape.” ↑b“Dialogue: Use quotes for specific speech.” ↑c“Veo 3.1 will sometimes block a video from generating because of safety filters or other processing issues with the audio.” ↑d“Dialogue: Use quotes for specific speech.” ↑e
- 2Google, Gemini API deprecations, last updated 12 Oct 2026, read at source 13 Oct 2026veo-3.1-generate-preview, veo-3.1-fast-generate-preview and veo-3.1-lite-generate-preview, shutdown “October 22, 2026”, recommended replacement “gemini-omni-1.1-flash”; the page notes that its dates “indicate the earliest possible dates on which a model might be retired”. ↑
- 3Google, Video generation in the Gemini API, read at source 3 Nov 2026“Veo 3.1 Preview (veo-3.1-generate-preview, veo-3.1-fast-generate-preview, and veo-3.1-lite-generate-preview) in the Gemini API was shut down on October 22, 2026; migrate to Gemini Omni Flash (gemini-omni-1.1-flash) or the GA models available through the Gemini Enterprise Agent Platform to avoid service interruptions.” ↑
- 4Google Cloud, Veo 3.1, Gemini Enterprise Agent Platform documentation, last updated 10 Oct 2026, read at source 13 Oct 2026veo-3.1-generate-001 and veo-3.1-fast-generate-001, “Launch stage: GA”, retirement date “November 17, 2026 or later”; veo-3.1-lite-generate-001, “Launch stage: Preview”. ↑
- 5Kling AI, Kling 4.0 vs 3.0, 30 Sep 2026, read at source 12 Oct 2026“Kling 3.0 already supports Native Audio, including dialogue, ambience, sound effects, multilingual speech, and lip-sync.” and “Kling 4.0 expands the audiovisual side with two-channel stereo audio and improved lip-sync accuracy.” · Kling AI, Kling 4.0 at Work, 27 Oct 2026, read at source 3 Nov 2026: “Since Kling 4.0 rolled out to all users on October 15, creative teams across film, advertising, and animation have been putting these capabilities to work.” ↑
- 6ByteDance Seed, Seedance 2.5, read at source 12 Oct 2026“a next-generation audio-video joint generation model”. ↑
- 7Dreamina (ByteDance), Seedance 2.5 prompt guide, 18 Aug 2026, read at source 12 Oct 2026template audio options “dialogue/ambience/SFX/music”. ↑a“Attach each sound cue to an action or timecode and define whether music rises, holds, or resolves”. ↑b“For audio: no background music, no dialogue, or ambience only—state the intended audio behavior instead of simply asking for ‘good sound’.” ↑c“For audio: no background music, no dialogue, or ambience only”. ↑d
- 8xAI, Video generation, read at source 12 Oct 2026“Generated videos include an audio track by default, with lip-synced speech on `grok-imagine-video-1.5`” and “Pass `generate_audio=False` to request a silent video”. ↑a“Pass `generate_audio=False` to request a silent video”. ↑b
- 9Google Cloud Blog, The ultimate prompting guide for Veo 3.1, 16 Oct 2025, read at source 12 Oct 2026, re-read 13 Oct 2026“Sound effects (SFX): Describe sounds with clarity (e.g., SFX: thunder cracks in the distance).” and “Ambient noise: Define the background soundscape (e.g., Ambient noise: the quiet hum of a starship bridge).” ↑a“Dialogue: Use quotation marks for specific speech”, with the example A woman says, “We have to leave now.” ↑b“By assigning actions to timed segments, you can efficiently create a full scene with multiple distinct shots, saving time and ensuring visual consistency.” ↑c
- 10Google Cloud, Best practices for generating videos, Gemini Enterprise Agent Platform documentation (the older Vertex AI address redirects here), last updated 9 Oct 2026, read at source 13 Oct 2026“To prevent the model from rendering text in the video, use a colon ( : ) after the speaker’s action to denote speech and avoid using quotation marks”; its recommended example is “A woman says: My name is Clara.” ↑a“To prevent the model from rendering text in the video, use a colon ( : ) after the speaker’s action to denote speech and avoid using quotation marks”; its recommended example is “A woman says: My name is Clara.” ↑b
- 11Kling AI, Kling VIDEO 3.0 Model User Guide, 6 Feb 2026, read at source 12 Oct 2026“When inputting instructions, you can pair character directly with their respective dialogue.” Its example line: “Mom (softly, in a surprised tone): Wow, I didn’t expect this plot at all.” ↑a“Video 3.0 supports dialogue output in five languages: Chinese, English, Japanese, Korean, and Spanish.” and “If dialogue is entered in a language other than those listed above, the model will translate it into English.” ↑b
- 12Runway, API reference, image to video, model
veo3.1, read at source 12 Oct 2026“Whether to generate audio for the video.” (default on). ↑ - 13ElevenLabs, Sound effects, read at source 12 Oct 2026the API “turns text descriptions into high-quality audio effects with precise control over timing, style and complexity”, up to “30 seconds per generation”, with an option to “Enable seamless looping for sound effects longer than 30 seconds”. ↑
Kling AI, Kling VIDEO 3.0 Model User Guide, 6 Feb 2026, and Kling 4.0 vs 3.0, 30 Sep 2026, both read at source 12 Oct 2026.
ByteDance Seed, Seedance 2.5, and Dreamina, Seedance 2.5 prompt guide, 18 Aug 2026, both read at source 12 Oct 2026.
The sound-team comparison and the example prompts are this site’s own, not claims from these sources.
Free printable AI cheat sheets, and AI news every week.
Sixteen one-page AI guides and checklists in one free PDF, no sign-up. And if you want to keep up, the newsletter covers what changed in AI each week, in plain English, with the sources.
SUBSCRIBE FREE ↗ GET THE CHEAT SHEET PACK