AI audio for video: voices, effects and music, and their limits
Dialogue, effects and music generated with the picture: what the models really do in 2026, where they break in the edit and what to check on rights.
AI audio for video now comes from three places: video models that generate sound together with the picture, such as Veo and Kling, voice models such as ElevenLabs', and music generators such as Suno and Eleven Music. For drafts, animatics and short reels it is often good enough straight out of the model. For an ad or a brand film you still need someone to unify the tracks, rebuild the ambience, clean up the dialogue and check the rights.
This guide covers what the models actually generate as of November 2026, where generated audio works, where it breaks and what to check before you publish. Dates and features come from official pages or press coverage, linked in the text.
Native audio in video models: where things stand
Until 2025 video models were silent. With Veo 3, unveiled by Google in May 2025, came dialogue, sound effects and ambient noise synchronised with the picture; Demis Hassabis described it as leaving the silent-film era behind. In October 2025 Lightricks released LTX-2 with audio built in, as Wikipedia’s entry on text-to-video models records.
| Model | What it generates | Length per clip | Notes |
|---|---|---|---|
| Veo 3 and 3.1 | Dialogue, effects, ambience | 8 s for Veo 3 at launch | Veo 3.1 since 15 October 2025 |
| Kling 3.0 | Multi-character dialogue, different languages and accents | Up to 15 s | Lip-sync per character |
| Kling 4.0 | Stereo audio, more spatial ambience | 3 to 30 s | Voice references |
| Sora 2 | Video and audio | - | Shut down: app in April, API in September 2026 |
According to the Kling Video 3.0 page, you can assign lines to several characters in the same scene, each with their own voice and lip movement. The Kling 4.0 guide adds two-channel stereo, ambience with a clearer sense of space and voice references to lock a character’s vocal identity. Kling says 4.0 officially launched on 15 October 2026 and that audio options depend on the model and the account.
Then there’s the Sora lesson. OpenAI introduced Sora 2 on 30 September 2025 as a video and audio app; according to Wikipedia, citing OpenAI’s help centre, the app closed on 26 April 2026 and the API on 24 September 2026. An audio workflow built on a single engine can vanish within months.
AI voices: on-screen dialogue and voiceover
Voices are now generated in two ways. The first is inside the shot: the video model produces face, line and lip movement together. In the projects Kling showcased after the 4.0 launch, creators point to more accurate lip-sync, including on Cantonese dialogue.
The second is a separate voice model. ElevenLabs' Eleven v3, released in alpha in June 2025, supports more than 70 languages, multi-speaker dialogue and tags such as [excited] or [sighs] to direct the delivery.
Our rule of thumb: on-screen dialogue when the face is in close-up and the line is short; a separate voice model for voiceover, narration and anything the client might want to change. Rewriting a voiceover line takes a minute. Regenerating a talking close-up to change one word means remaking the shot. For other languages, see our guide to AI dubbing and lip-sync.
Sound effects and ambience: where they work and where they break
Footsteps, doors, traffic, wind: native audio puts them where they belong, at the right moment. In the Kling 4.0 guide, the sample prompt asks for footsteps, street ambience, background music and English dialogue in a single generation. As a guide for timing and rhythm, that’s valuable.
The trouble starts in the edit. In our experience it’s always the same list:
- Jumping ambience. Every clip generates its own room tone. Cut between two shots of the same scene and the background changes.
- Everything on one track. Voice, effects and music arrive already mixed. You can’t duck the music under a line or swap a single effect.
- Uninvited music. A bed generated inside the clip clashes with the soundtrack of the edit.
- Drifting voices. The same character can sound different from one generation to the next unless you lock a voice reference.
- Artefacts. Slurred words, repeated syllables, sounds with no source: listen for them on headphones, not on a phone.
Music: generated, licensed or composed
Music is where the rules bite hardest. On Suno only tracks downloaded on a paid plan carry commercial rights; trial downloads on the free plan are for personal use. Suno assigns its rights to subscribers but, Music Business Worldwide reports, gives no warranty that copyright will vest in the output. ElevenLabs launched Eleven Music in August 2025 with licences from Merlin and Kobalt, presenting it as cleared for broad commercial use.
The legal picture keeps moving: lawsuits from the majors, a German ruling on outputs that reproduce existing songs, cases over imitated voices. We cover it in our piece on AI music licensing. In practice: switch off the music generated inside clips, and choose the soundtrack separately, with a clear licence.
When AI audio for video is enough, and when you need a sound designer
Not everything needs a mixing stage. Here’s how we think about it by type of project:
| Project | Generated audio | What post adds |
|---|---|---|
| Animatic and previz | Enough | Nothing |
| Reels and social content | Often enough | Consistent levels, clean-up |
| Ads and brand films | As a guide | Rebuilt ambience, mix, masters per channel |
| Language versions | Not enough | Separate voice, music and effects stems |
When post is needed, the sequence is:
- Decide early which sounds to generate and which to build, at script stage.
- Generate with audio on and use it as a reference for the timing and rhythm of the edit.
- Lock the edit before touching the sound.
- Rebuild the ambience scene by scene, with a continuous bed under the cuts.
- Fix the dialogue: clean up the good takes, redo the weak ones with a voice model.
- Lay in the final music, licensed or composed, and mix.
- Deliver masters and stems separately for each channel and each version.
Rights: what to check before you publish
Every platform sets its own terms. Kling’s Terms of Use, updated on 21 April 2026, bar commercial use of output without written permission and point to separate terms for paid services; they also ask you to label distributed output as generated with Kling, unless you have written permission. Always read the terms of the plan you’re on.
On voices the line is clear. Spotify removes unauthorised voice clones, and in November 2026 Elvis Presley’s rights holders sued Suno, alleging it let users imitate his voice. Real people’s voices need written consent: see our guide to voice and likeness rights, and for transparency duties our guide to the AI Act.
This is general information, not legal advice: for a specific project, have a lawyer check the terms and releases.
Sound is finished in post
In our method, generated audio works during production, as a guide; the final sound is built in post-production, alongside the edit, colour and versions, and the client signs it off before delivery.
If you have a video that needs to sound like a film rather than a demo, let’s talk.
Frequently asked questions
Do AI video models generate audio too?
The leading ones do. Since Veo 3 launched in May 2025, models generate dialogue, effects and ambience in sync with the picture, and Kling 4.0, released in October 2026, outputs stereo audio. What you get depends on the model and the plan.
Can I use an AI-generated voice in an advert?
Yes, if your plan’s terms allow commercial use and the voice doesn’t imitate a real person without their written consent. Spotify, for example, removes unauthorised voice clones. This is general information, not legal advice.
Why isn’t generated audio enough for an ad?
Because every clip has its own ambience and arrives already mixed: in the edit the background shifts at every cut and you can’t separate voice, music and effects. An ad needs rebuilt ambience, a proper mix and separate stems for the versions.
Sources
- Veo (text-to-video model) · Wikipedia, 2026-11
- Text-to-video model · Wikipedia, 2026-11
- Kling Video 3.0 · Kling AI, 2026-11
- Inside Kling 4.0: A Complete Guide to AI Video Creation · Kling AI, 2026-09-30
- AI Video Generation News: Industry Updates and Kling 4.0 · Kling AI, 2026-10-26
- Kling 4.0 in Practice: Eight Creators Show What Sets It Apart · Kling AI, 2026-11-03
- Sora (text-to-video model) · Wikipedia, 2026-11
- ElevenLabs · Wikipedia, 2026-11
- An update to our downloads policy and Terms of Service · Suno, 2026-08-10
- Suno launches v6 AI music models in partnership with WMG, BMG and Believe · Music Business Worldwide, 2026-09-09
- ElevenLabs Launches Eleven Music in Collaboration with Music Industry Partners · ElevenLabs via Business Wire, 2025-08-05
- Kling AI Terms of Use · Kling AI, 2026-04-21
- Spotify updates AI policy to label tracks, cut down on spam · TechCrunch, 2025-09-25
- Elvis rights holders and Eminem’s publisher sue Suno, testing the ELVIS Act · Los Angeles Times (via Yahoo), 2026-11-09
Keep reading.
Kling 3.0: a practical guide to multi-shot, audio, Omni and limits
Multi-shot, native audio, Omni elements, 4K and costs: how Kling 3.0 works in November 2026, what it can’t do, and when it still beats Kling 4.0.
Veo 3.1: a practical guide to audio, references, Flow and SynthID
Native audio, reference images, first and last frames, extension, access and limits: how to use Google Veo 3.1 in November 2026, and when to switch to Gemini Omni.
AI Christmas ads in 2026: who uses AI, who refuses it and why
Prada draws by hand, Levi’s goes back to claymation, Dior faces AI accusations: Christmas 2026 shows what audiences punish and accept from AI.