The first cut of an AI-generated commercial usually fails the same way. The frame is clean, the light behaves, the movement is plausible. Then you play it on a phone speaker and the whole thing feels like a slideshow with a narration track laid on top. Nothing is wrong with the picture. Everything is wrong with the sound.
That gap is the most common reason a technically finished AI video still reads as a demo rather than an ad. Generation models produce images with astonishing detail and almost no acoustic information. They cannot tell you how big the room is, what is behind the camera, or whether the floor is carpet or tile. Sound design is the part of the job that answers those questions, and in our studio it routinely takes as long as the visual edit.
Why generated shots sound empty
A real shoot captures more than dialogue. It captures the room: the low hum of a ventilation system, the slap of a footstep, the way a voice changes when someone turns away from the lens. Those cues are how an audience locates a scene without thinking about it. An AI shot arrives with none of them, because the model was never there.
When we receive a generated sequence, we treat it as silent footage from a location we have never visited, which is exactly what it is. The first pass is diagnostic: we watch with the audio muted and write down what each shot would have sounded like if a crew had filmed it. Room size, surface materials, distance to the subject, what is off-screen. That list becomes the sound map, and it is far more useful than starting from a music bed.
There is a second problem the picture hides. AI shots often cut faster than a real set would allow, because generating a new angle costs nothing. Four cuts in three seconds is fine visually and exhausting acoustically, since every cut implies a new space. Sound design either has to smooth those transitions with a continuous ambience bed or the edit has to be slowed. We usually slow the edit.
The four layers, in the order we build them
We build sound in four layers, and the order matters because each one constrains the next.
Dialogue and voiceover come first, even when the video has no on-screen speaker. A commercial without a human voice still needs a voice, whether that is a narrator, a radio-style read, or a single line from a character. We record or generate it before touching music, because the voice sets the tempo of the entire edit. If the read is 78 seconds, you have 78 seconds of video, not the other way around.
Room tone and ambience come second. Every location gets a continuous bed, usually 18 to 22 dB below the voice, that never fully disappears. The common beginner error is dropping ambience to zero under dialogue, which makes the voice sound pasted on. Keep the bed running and duck it two or three dB under speech instead.
Hard effects and foley come third: footsteps, cloth movement, a product being set down, a door, a phone vibrating on a table. These are the details that make a generated close-up of a hand believable. If a shot shows a ceramic cup meeting a wooden table, that contact needs a sound with the right weight and the right room behind it. We often generate three variants and pick by ear rather than by rule.
Music comes last, and it is the layer most often used to hide a problem instead of solving it. If the track is doing the work that ambience and foley should be doing, the result sounds like a trailer for nothing.
Matching reverb to what the lens says
The single most technical decision in AI sound design is reverb, because reverb tells the audience how far away things are. A wide shot of a large hall needs a long tail, roughly 1.8 to 2.6 seconds depending on surface hardness. A tight close-up in the same hall needs almost none, even though the space has not changed, because the camera is now inside the acoustic bubble of the subject.
We set a reverb profile per location, not per shot, then override it for close-ups. When a sequence moves from an exterior to an interior within the same scene, the tail length has to change at the cut, and if it does not, the transition feels like a jump even when the image cuts perfectly. This is the kind of mismatch that viewers cannot name and always notice.
Generated dialogue needs extra attention here. Voice models tend to produce a very dry, close, almost anechoic read. Dropping that into a wide shot of a tiled kitchen with no processing is one of the clearest tells that a video was machine-made. A short plate reverb, a touch of early reflection, and a slight high-frequency roll-off fix most of it.
Level targets we actually deliver to
Loudness is where most AI video work quietly fails platform review. Our standard delivery targets: minus 14 LUFS integrated for social platforms with a true peak ceiling of minus 1 dBTP; minus 16 LUFS for web and podcast embeds; minus 23 LUFS only when a broadcaster asks for it. Dialogue sits around minus 18 to minus 20 LUFS short-term, music beds 12 to 18 dB below dialogue, and ambience below that.
Two practical rules save time. First, mix at low volume for the first pass, because balance problems show up there before they show up loud. Second, check the whole video on a phone speaker before you approve it, since a meaningful share of the audience will hear it exactly that way, in a room with traffic outside. If the voice disappears on a phone, the mix is wrong regardless of what the meters say.
Vertical delivery adds another constraint. On a 9:16 cut played in a feed, the first two seconds are usually silent by platform behaviour, so any critical audio information in that window is wasted. We design the opening so that the first meaningful sound lands at the moment the image earns attention, not before.
When sound should change the edit
Sound design is not a finishing step applied to a locked cut. In our workflow it frequently sends the edit back for changes, and that is a feature rather than a failure. A shot that looked fine at four seconds feels long once it has a full ambience bed and a footstep pattern. A cut that felt abrupt becomes invisible when a sound bridges it, which is why we routinely place a single continuous sound across a cut to carry the audience through.
The practical version of this: build a rough sound pass at the same time as the first visual assembly, not after approval. It costs less to move a cut in the assembly than to rebuild a mix.
If you are planning an AI video and want to know what the sound work will involve for your specific format, send us the script or the shot list and we will tell you which layers your project actually needs. KAMOD is an AI video studio based in Tashkent, working with clients worldwide, and we would rather have that conversation before generation than after.