Sound design for AI video: why silence gives it away

AI video sound design is the work of building a soundtrack for generated footage: ambience, effects, music, voice and a mix that survives a phone speaker. It matters more than in filmed video, because a generated clip often arrives with no sound at all, or with sound that does not match the clip next to it. Silence is the quickest way to tell a viewer that nothing in the frame ever existed.

Why does AI video feel fake without sound?

A camera on a real street records traffic, wind, footsteps and a door somewhere behind it. Nobody notices this sound, but everybody notices when it is missing. A silent shot of a woman walking through a market looks like a render. Add a murmur of voices, a scooter passing and her shoes on the stone, and the same pixels read as footage.

Sound also hides things. Generated video has small visual faults: a hand that flickers, fabric that moves oddly. A well-timed sound pulls attention to the action the viewer is meant to follow. And it sets the pace. A cut that lands on a beat or on a door slam feels intended, while the same cut in silence feels like a mistake.

What are the layers of a soundtrack?

Layer	What it is	What it does in generated video
Ambience	The continuous sound of the place: room tone, street, wind, café	Convinces the viewer the place exists and glues separate shots into one scene
Foley and spot effects	Sounds tied to actions: steps, cloth, a cup set down, a lid clicking	Gives objects weight and makes generated movement feel physical
Music	Score or a licensed track	Sets mood and pace, and carries the viewer across cuts
Voice	Voiceover or dialogue, recorded or synthetic	Carries the message; everything else makes room for it
The mix	Levels, equalisation, compression, loudness	Makes all of it clear on a phone speaker and consistent from start to end

Beginners use one layer, usually music. It is the layers under the music that make footage feel filmed.

How is sound added to a silent generated clip?

1. Lock the edit first. Sound is timed to frames. Designing sound before the cut is final means doing it twice.
2. Lay one ambience per scene, not per shot. A single continuous bed under five shots tells the ear they happened in one place at one time.
3. Spot the effects by hand. Go through the cut and mark every visible action: a step, a pour, a touch. Place a sound on each, from a sound library or recorded on a phone. A real recording of your own product’s click or fizz is worth the ten minutes.
4. Add music and cut to it, or cut it to the picture. Move edit points by a frame or two so that they meet the beat.
5. Record or generate the voice, then clear space for it. Lower the music when someone speaks.
6. Mix and check on a phone. Details are below.

Text-to-sound-effect tools can now generate a specific effect from a description, which helps when a library has nothing close. They do not remove the need to place it on the right frame.

Which AI video models generate sound themselves?

Until 2025 almost every video model produced silent clips. As of September 2026 the leading ones generate synchronised audio with the picture:

• Veo 3.1 (Google). Google’s Veo documentation says the model natively generates audio with video and takes prompts for dialogue, sound effects and ambient noise. It also notes that English is fully supported and other languages have not been evaluated.
• Kling 3.0 (Kuaishou). Native audio with dialogue, effects and music in one generation, and lip movement matched to each speaker. Kling lists Chinese, English, Japanese, Korean and Spanish. Russian is not on that list.
• Seedance 2.5 (ByteDance). Released on 31 July 2026 with joint audio and video generation, clips of up to 30 seconds, and audio files accepted as references.
• Sora 2 (OpenAI) also generated audio, but OpenAI closed the Sora app in April 2026 and scheduled the API shutdown for 24 September 2026.

This is real progress, and for a single clip posted on its own it can be enough. In an edited film it creates a new problem. Each clip is generated separately, so each has its own room tone, its own level and sometimes its own idea of how the character’s voice sounds. Cut five of those together and the background jumps at every edit. Generated music cannot be traced to a licence. And the audio comes as one mixed file, so you cannot turn the music down under the voice.

So at Zero Shot, an AI video studio in London, we treat native audio as a useful sketch. We keep the synchronised effects that are hard to fake, such as a footstep that lands exactly right, and rebuild ambience, music and voice in the edit. The rest of that process is described in how AI video is made.

AI voice or a voice actor?

Synthetic voices are now good enough for most short narration, and they have practical advantages: a changed line takes a minute, the voice is the same in every video, and one script can be produced in English and in Russian. A second language takes more than a second voice track, as we explain in one reel made in English and Russian. They need directing all the same. Generate each line several times, fix brand names with phonetic spelling, and cut out the breaths that fall in odd places.

Use an actor when the read carries emotion or comic timing, when the audience will hear the voice for years, or when an accent has to be exactly right for a local audience. Cloning a real person’s voice needs that person’s written consent, the same as using their face. For a presenter who appears on screen, see AI presenters and avatars, and for the trust questions around synthetic spokespeople read AI UGC ads explained.

Music licensing: what can a business account use?

This is where small businesses most often get into trouble, because the rules for a personal account and for a business are different.

Instagram. Meta’s help page on access to the licensed music library says the music in it is intended for personal, non-commercial use, and that certain business accounts and types of post do not have access. Meta points those accounts to its Sound Collection, which it describes as more than 14,000 royalty-free songs and sounds that can be used for commercial purposes, including ads.

TikTok. TikTok’s page on its Commercial Music Library says businesses cannot use the general music library for commercial use and should use the commercial library for organic posts as well as ads. Business accounts see only those commercial sounds when they add music.

Two consequences follow. A trending song is rarely available to a brand, even on an unpaid post. And each platform’s library covers that platform only, so a reel scored with a Sound Collection track is not cleared for TikTok. If one video will run everywhere, or as a paid ad, license the track from a production music library or a composer, and check that the licence names paid advertising and the platforms you use. With AI-generated music, read the tool’s terms on commercial use. This is general information, not legal advice.

How loud should it be, and how do you mix for a phone?

None of the three big platforms publishes an official loudness target for short video. A common working convention is about -14 LUFS integrated, with true peaks below -1 dBTP so the platform’s compression does not distort the loudest moments. More useful than the number are four habits:

• Keep the voice clearly above everything else, and lower the music whenever someone speaks.
• Do not rely on deep bass. A phone speaker reproduces very little of it, so a sound that lives only in the low end vanishes.
• Check the mix in mono, because a phone held upright gives almost no stereo separation.
• Listen on an actual phone at half volume in a room that is not silent. If the words are clear there, the mix works.

Why add captions if the sound is this good?

Because a large share of viewers will never hear it. People scroll on trains, in offices and in bed next to someone asleep. Instagram and TikTok can both generate captions automatically, but burned-in captions that you style and proofread are safer, especially for brand names and for Russian or mixed-language speech. Keep them inside the safe area, two lines at most. The exact margins are in our reference on Reels safe zones and export settings. The reel has to work muted and reward the viewer who turns the sound on. Our guide to making Instagram Reels with AI covers where captions can sit.

If you have generated footage that feels lifeless, or need reels and short-form video delivered with finished sound, send us a short brief and we will reply with a fixed price.