# What is native audio in AI video?

URL: https://dreambench.ai/glossary/native-audio

Native audio is sound that an AI video model generates in the same pass as the picture: dialogue, sound effects, ambience and sometimes music, timed to what happens on screen. It differs from sound added afterward, by a separate voice, effects or music model or by hand in an editor, which has to be fitted to finished frames.

Written and reviewed by the DreamBench team. Published 2026-10-29; updated 2026-10-29.

## Native audio at a glance

- **What it is:** Sound made by the video model itself, in the same pass as the frames.
- **The layers:** Dialogue, sound effects, ambience and, on some models, music.
- **How you direct it:** With words in the prompt: a spoken line in quotes, the sounds of the place, the effects you want to hear.
- **Strong at:** Ambience and effects that land on the action; a short line said by someone on screen.
- **Weak at:** Long or fast speech, several speakers at once, an exact voice, a particular song.
- **Not the same as:** Lip sync, which moves a mouth to match a voice track you already have.

## How native audio works

A model with native audio is trained on clips with their soundtracks, so it learns that a door shuts with a bang, a crowded street hums and lips move with the words they say. When it generates, it makes the frames and the sound together, both steered by the same prompt and laid on the same timeline, so each sound starts on the frame that causes it.

*Diagram: Native audio against sound added afterward. Generated together, each layer lands on the frame that causes it; fitted afterward, the same sounds have to be placed, and can drift off their frames.*

The older way, still common, makes a silent clip first and gives it sound afterward: a voice from a text-to-speech model, effects from a sound model that watches the clip, music chosen in an editor. Each step is separate, so each has to be lined up with the picture by hand or by another model. Google describes [Veo 3.1](https://dreambench.ai/models/veo-3.1) as [generating all of its audio natively](https://deepmind.google/models/veo/), and ByteDance calls [Seedance 2.5](https://dreambench.ai/models/seedance-2.5) [an audio-video joint generation model](https://seed.bytedance.com/en/seedance2_5).

#### Under the hood: two streams that listen to each other

Published joint models run two generators side by side, one for the picture and one for the sound, and let them exchange information at every step. [MM-Diffusion](https://arxiv.org/abs/2212.09478) (2022) was an early design: two subnetworks that turn noise into a matching clip and soundtrack together. [Ovi](https://arxiv.org/abs/2510.01284) (2025) joins two transformer towers block by block, sharing timing and meaning between them, which its authors say gives synchronization without a separate alignment step.

The [LTX-2 paper](https://arxiv.org/abs/2601.03233) (2026), from Lightricks, the makers of [LTX-2.3](https://dreambench.ai/models/ltx-2.3), sets out the same idea at scale: a 14-billion-parameter video stream and a 5-billion-parameter audio stream, joined by attention in both directions, with speech, background sound and foley, and published weights. The other route, sound made after the picture, is still studied too: [MMAudio](https://arxiv.org/abs/2412.15322) takes a finished video, with optional text, and generates sound aligned to it frame by frame.

## The four kinds of sound in an AI video, and how to prompt each

A clip’s soundtrack is built from layers, and a prompt works best when it names each layer it wants. Google’s guidance for [Veo 3.1](https://dreambench.ai/models/veo-3.1) is to [say clearly whether you want audio, and describe it in sentences of its own](https://docs.cloud.google.com/vertex-ai/generative-ai/docs/video/video-gen-prompt-guide); [its prompting guide](https://cloud.google.com/blog/products/ai-machine-learning/ultimate-prompting-guide-for-veo-3-1) labels effects “SFX:” and the background “Ambient noise:”.

**The layers of an AI video’s sound**

|  | What it is | How to ask for it | Example |
|---|---|---|---|
| Dialogue | Words spoken by someone on screen or by a narrator | The line in quotes, straight after the person who says it, with how they say it | The old chef says quietly, “Still not perfect.” |
| Sound effects | Single sounds caused by something in the scene | Name the sound and the moment it happens | SFX: a cup set down hard on the counter as she stands. |
| Ambience | The steady sound of the place, under everything else | Two to five words drawn from the location | Ambient noise: rain on the window, a quiet room. |
| Music | A score or song in the clip | Genre, instruments and mood, or “no music” if you will add your own | No music: only the scene’s own sounds. |
| Silence | What should not be heard | Say it, or the model fills the gap | No dialogue; the only sound is wind. |

Ambience and effects are what video models do most reliably: a place’s sound is long and forgiving, and an effect only has to land on one frame. Speech is the hardest layer, because every syllable has to match a mouth.

## Native audio vs text to speech, AI sound effects and lip sync

Native audio is one of several ways to give a video sound:

**Native audio and the other ways to give a video sound**

|  | What makes the sound | In time with the picture | Choose it for |
|---|---|---|---|
| Native audio | The video model, in the same pass as the frames | Yes, from the start: one timeline | Clips where the sound belongs to the action: a line, a slam, a room |
| Text to speech | A voice model, from a script, with no picture | No: you place the voice, and a mouth on screen won’t match it | Narration over footage, or an exact voice you have chosen |
| AI sound effects (video to audio) | A sound model that watches a finished clip, such as MMAudio | Close: it times effects to the frames it is given | Effects and ambience for a clip that was made silent |
| Music in an editor | A track you choose or license, laid under the cut | Only where you cut it to the beat | A score that runs across several clips, or a song you need exactly |
| Lip sync | A model that moves a face’s mouth to match a voice track you already have | Yes, for the mouth: the voice comes first | Dubbing, translation, or a voice recorded separately |

The difference from [lip sync](https://dreambench.ai/glossary/lip-sync) is the order. Lip sync starts from a voice and changes the picture to match it; native audio makes the voice and the moving mouth at the same time, from the words of the prompt, so neither is fitted to the other.

## How to add dialogue to an AI video

Dialogue is written into the prompt, not recorded. This 30-second take was made on DreamBench with [Seedance 2.5](https://dreambench.ai/models/seedance-2.5) with sound on; the brief is in English and the chef’s one line is in Japanese, quoted in Japanese script with its meaning beside it.

*Video (seedance-2.5, 16:9, 720p, 30 s, with sound): A spoken line on [Seedance 2.5](https://dreambench.ai/models/seedance-2.5): “says quietly in Japanese”, then the words in quotes. The shop’s hush comes with it, unasked.* [Watch](https://media.dreambench.ai/templates/tpl-g-o05mz7aa8g.mp4)

> Brief: One unbroken take inside a tiny late-night ramen shop, steam everywhere. An old chef ladles broth with total focus, tastes it, pauses, and says quietly in Japanese: 「五十年、同じスープ。まだ完璧じゃない。」("Fifty years, the same soup. Still not perfect.") Without a cut he turns, sets the finished bowl in front of a young customer at the counter, and watches him take the first sip — allowing himself the smallest smile as the camera drifts slowly closer through the steam. No cuts, no cutaways: one continuous shot from ladle to smile.

- Put the line in quotes, straight after the person who says it, so the voice belongs to someone on screen.
- Say how it is said: quietly, out of breath, laughing. The delivery is part of the line.
- Name the language when it isn’t the language of the brief, and write the words in that language.
- Keep it short: one or two short sentences fit an 8-second clip; a speech does not.
- One speaker at a time. Two people talking over each other is where voices get mixed up and mouths go wrong.

> Google’s model reference lists English as [the prompt language for Veo 3.1](https://docs.cloud.google.com/vertex-ai/generative-ai/docs/models/veo/3-1-generate). A line in another language is worth testing on the model before you rely on it.

## Prompting effects and ambience

Sound without words still needs directing. This clip on [Veo 3.1](https://dreambench.ai/models/veo-3.1) asks for one sound that changes once, on cue: the brief ends “Muffled underwater ambience, then one clear breath”, and the breath comes as the turtle breaks the surface.

*Video (veo-3.1, 16:9, 1080p, 15 s, with sound): Ambience on [Veo 3.1](https://dreambench.ai/models/veo-3.1): a muffled underwater bed for most of the clip, then one sound placed at the moment the picture changes.* [Watch](https://media.dreambench.ai/templates/tpl-g-p5iq7agjo4.mp4)

> Brief: Underwater at sunrise: light shafts sweep across a coral reef as a sea turtle glides through a school of silver fish that parts around it like a curtain. A ray lifts from the sand in a slow wingbeat of silt. The turtle rises toward the glittering surface and breaks it into the sun. Muffled underwater ambience, then one clear breath.

- Tie each effect to an action in the same sentence: “the door slams as she leaves” lands better than a list of sounds at the end.
- Describe the ambience as the place’s own material: wind over a rooftop, a café’s murmur, traffic two streets away.
- Say how the sound ends: it stops on a cut, carries on, or fades.

## Should the model make the sound, or should you add it?

**Native audio or sound added afterward?**

- **A character says a line on screen:** Native audio. Write the line in quotes after the person who says it: the voice and the mouth are made together, so they start together.
- **A narrator speaks over the pictures:** Either. A video model can voice an unseen narrator, but a text-to-speech voice gives you the exact words, voice and timing, and you can change it without remaking the clip.
- **You need one particular voice:** Add it afterward. A video model’s voice comes from the words that describe it; record the voice, or use one you have the rights to, and use lip sync to match the mouth.
- **B-roll with the sound of the place:** Native audio. Ambience and effects are what video models do most reliably, and they land on the action without any lining up.
- **A film of several clips under one score:** Add the music afterward. Each clip is generated on its own, so music made inside a clip stops at its cut; one track laid under the whole edit runs across the joins.

## Where native audio breaks

- Lip sync drift: the mouth and the words start together, then slide apart over a long line, or the mouth moves when no one is speaking.
- Garbled speech: words come out slurred, invented or in the wrong order, more often in long or fast lines. Google says it is still working to [remove instances of incoherent speech](https://deepmind.google/models/veo/) from Veo, and its guide calls the audio features [experimental and actively improving](https://cloud.google.com/blog/products/ai-machine-learning/ultimate-prompting-guide-for-veo-3-1).
- The wrong speaker: with two or three people in frame, a line can come from the wrong mouth, or an extra voice can appear.
- Voices that change between clips: each clip is made separately, so a character described differently in two prompts can sound like two people; see [character consistency](https://dreambench.ai/glossary/character-consistency).
- Music that stops at every cut: a score made inside one clip ends where the clip ends, and the next clip starts a different one.
- Music you can’t vouch for: nothing in a generated clip tells you whether its melody resembles a recording that exists. For anything you publish commercially, use a track whose license you know.
- Loudness that jumps: each clip comes out at the level the model chose, so a cut between two clips can be suddenly louder or quieter. Level the sound in an editor before you publish.

#### Voices and consent

A generated voice can sound like a real person’s. Don’t prompt for a named person’s voice, and don’t put words in a real person’s mouth without their consent; for a voice you need exactly, record it with permission.

## Which AI video models have native audio

Most current video models generate sound. On DreamBench, some always do and some have a switch, and the table below is read off the composer’s own settings:

**Video models on DreamBench: sound always on, or on or off, beside first and last frame, resolutions and aspect ratios**

| Model | Sound | First and last frame | Resolutions | Aspect ratios |
|---|---|---|---|---|
| Seedance 2.5 | On or off | Yes | 480p, 720p, 1080p | 1:1 · 16:9 · 9:16 · 4:3 · 21:9 |
| Seedance 2.0 | On or off | Yes | 720p, 1080p, 4K | 1:1 · 16:9 · 9:16 · 4:3 · 21:9 |
| Veo 3.1 | Always on | Yes | 720p, 1080p, 4K | 16:9 · 9:16 |
| Kling 4.0 | Always on | Yes | 720p, 1080p | 1:1 · 16:9 · 9:16 · 4:3 · 21:9 |
| Kling 3.0 | On or off | Yes | 720p, 1080p | 1:1 · 16:9 · 9:16 · 4:3 · 21:9 |
| Wan 3.0 | Always on | Yes | 720p, 1080p | 1:1 · 16:9 · 9:16 · 4:3 |
| Wan 2.7 | On or off | Yes | 720p, 1080p | 1:1 · 16:9 · 9:16 · 4:3 |
| Grok Imagine 1.5 | Always on | First frame only | 720p, 1080p | 1:1 · 16:9 · 9:16 · 4:3 |
| MiniMax H3 | Always on | Yes | 720p, 1080p | 1:1 · 16:9 · 9:16 · 4:3 · 21:9 |
| HappyHorse 1.1 | Always on | First frame only | 720p, 1080p | 1:1 · 16:9 · 9:16 · 4:3 · 21:9 |
| LTX-2.3 | On or off | Yes | 720p, 1080p, 4K | 16:9 · 9:16 |
| Vidu Q2 | On or off | First frame only | 720p, 1080p | 1:1 · 16:9 · 9:16 · 4:3 |

Sound is always on with [Veo 3.1](https://dreambench.ai/models/veo-3.1), [Kling 4.0](https://dreambench.ai/models/kling-4.0), [Wan 3.0](https://dreambench.ai/models/wan-3.0), [Grok Imagine 1.5](https://dreambench.ai/models/grok-imagine-1.5), [MiniMax H3](https://dreambench.ai/models/minimax-h3) and [HappyHorse 1.1](https://dreambench.ai/models/happyhorse-1.1): their makers offer no way to turn it off. [Seedance 2.5](https://dreambench.ai/models/seedance-2.5), [Seedance 2.0](https://dreambench.ai/models/seedance-2.0), [Kling 3.0](https://dreambench.ai/models/kling-3.0), [Wan 2.7](https://dreambench.ai/models/wan-2.7) and [LTX-2.3](https://dreambench.ai/models/ltx-2.3) have a switch. [Vidu Q2](https://dreambench.ai/models/vidu-q2)’s only sound is background music, which a DreamBench film leaves off.

Turning sound off lowers the price only on Kling 3.0, whose provider bills a silent clip for less; on every other model, Seedance 2.0 among them, though ByteDance describes it as [a joint audio-video model](https://seed.bytedance.com/en/seedance2_0), the switch changes what you hear, not the credits. Each model’s page on [the models list](https://dreambench.ai/models) gives its lengths and settings.

## Native audio on DreamBench

On DreamBench you [describe the film](https://dreambench.ai/), and DreamBench writes every clip’s prompt, sound included. Each clip with sound gets a sound line of its own: the ambience of the place, each speaker’s voice, how many lines are spoken, by whom and in which language when your brief names one, and how the sound ends at the cut.

- A character’s voice is decided once and written word for word into every clip they speak in, because a voice described in new words each time comes out as a new voice.
- A film’s clips carry their own sounds and voices, never music: each clip renders on its own, so a score made inside one would stop dead at its cut.
- The Audio switch sits beside the model in the composer. On a model whose sound is always on, it shows on and can’t be turned off.

To write a single prompt with sound yourself, the [AI video prompt generator](https://dreambench.ai/tools/ai-video-prompt-generator) has a sound part, from ambience alone to one spoken line in quotes, and the [Veo 3.1 prompt generator](https://dreambench.ai/tools/veo-prompt-generator) writes for a model whose sound is always on. Read more on AI video generation, [text to video](https://dreambench.ai/glossary/text-to-video) and [multi-shot video](https://dreambench.ai/glossary/multi-shot-video).

## Questions

- **Which AI video generator has sound?** Most current video models generate sound, among them Veo 3.1, Seedance 2.5, Kling 4.0, Kling 3.0, Wan 3.0 and LTX-2.3. On DreamBench, Veo 3.1, Kling 4.0, Wan 3.0, Grok Imagine 1.5, MiniMax H3 and HappyHorse 1.1 always make sound, while Seedance 2.5, Seedance 2.0, Kling 3.0, Wan 2.7 and LTX-2.3 have an Audio switch.
- **How do I add dialogue to an AI video?** Write the line into the prompt in quotes, straight after the person who says it, with how they say it: the old chef says quietly, “Still not perfect.” Keep it to one or two short sentences per clip, one speaker at a time, and name the language if it differs from the rest of the prompt.
- **Can Veo 3.1 speak other languages?** Google’s Veo 3.1 model reference lists English as its prompt language, so a line in another language is not documented and is worth testing before you rely on it. Other models take lines in other languages: the ramen clip on this page has Seedance 2.5 speak a line in Japanese from an English brief.
- **What is the difference between native audio and lip sync?** Native audio makes the voice and the moving mouth together, from the words of a prompt. Lip sync starts from a voice track you already have, recorded or from text to speech, and changes the mouth in a video to match it, which is how dubbing and translation are done.
- **Why is the speech in my AI video garbled?** Speech is the hardest layer for a video model: every syllable has to match a mouth, and Google itself says it is still working to remove incoherent speech from Veo. Shorten the line, give it one speaker, write the words in the language they should be spoken in, and generate again.
- **Why don’t the lips match the words in my AI video?** The mouth and the voice usually start together and drift apart over a long or fast line, or when several people are in frame. Keep each line short, keep one speaker in frame when they talk, and put the line straight after the person who says it so the model knows whose mouth moves.
- **Can I turn off the sound in an AI video?** On some models. On DreamBench, Seedance 2.5, Seedance 2.0, Kling 3.0, Wan 2.7 and LTX-2.3 have an Audio switch; Veo 3.1, Kling 4.0, Wan 3.0, Grok Imagine 1.5, MiniMax H3 and HappyHorse 1.1 always make sound, so the switch shows on and can’t be changed. You can still mute or replace the sound in an editor afterward.
- **Does AI video with sound cost more?** It depends on the model. On DreamBench, turning sound off lowers the price only on Kling 3.0; on every other model the Audio switch changes what you hear, not the credits, and on the models whose sound is always on there is no silent version to choose.
- **Should I let the AI video model make the music?** For a single clip it can. For a film of several clips, add one music track afterward: each clip is generated on its own, so music made inside a clip stops at its cut. A track you license yourself also comes with terms you can read.
- **Can I get only background sound, with no talking?** Yes. Describe the place’s sound and say there is no speech: “No dialogue; rain on the window, a quiet room.” If you leave the sound out of the prompt on a model whose sound is always on, it still makes a soundtrack, chosen by the model.
- **How do I keep a character’s voice the same across clips?** Describe the voice in the same words in every clip the character speaks in, for example “a low, dry voice, slow and amused”. Each clip is made on its own, so new words for the voice give a new voice. On DreamBench the voice is decided once and written into every clip for you.

## Sources

- [Veo 3.1](https://deepmind.google/models/veo/), Google DeepMind: read 2026-10-29
- [Veo 3.1 prompting guide](https://cloud.google.com/blog/products/ai-machine-learning/ultimate-prompting-guide-for-veo-3-1), Google Cloud: read 2026-10-29
- [Video generation prompt guide](https://docs.cloud.google.com/vertex-ai/generative-ai/docs/video/video-gen-prompt-guide), Google Cloud: read 2026-10-29
- [Veo 3.1 model reference](https://docs.cloud.google.com/vertex-ai/generative-ai/docs/models/veo/3-1-generate), Google Cloud: read 2026-10-29
- [Seedance 2.0](https://seed.bytedance.com/en/seedance2_0), ByteDance Seed: read 2026-10-29
- [Seedance 2.5](https://seed.bytedance.com/en/seedance2_5), ByteDance Seed: read 2026-10-29
- [MM-Diffusion: Learning Multi-Modal Diffusion Models for Joint Audio and Video Generation](https://arxiv.org/abs/2212.09478), arXiv: read 2026-10-29
- [Ovi: Twin Backbone Cross-Modal Fusion for Audio-Video Generation](https://arxiv.org/abs/2510.01284), arXiv: read 2026-10-29
- [LTX-2: Efficient Joint Audio-Visual Foundation Model](https://arxiv.org/abs/2601.03233), arXiv: read 2026-10-29
- [MMAudio: Taming Multimodal Joint Training for High-Quality Video-to-Audio Synthesis](https://arxiv.org/abs/2412.15322), arXiv: read 2026-10-29

DreamBench is an independent platform and isn’t affiliated with or endorsed by the companies named on this page.