When people say an AI ad "feels fake", they usually point at the picture: the skin, the hands, the light. But watch one with your eyes closed and you'll often hear it first. There's a voice in a perfectly silent room. The words sound read, not said. A café shot has no café in it. The music is loud enough to fight the lines, then cuts off mid-bar.
We learned this the hard way. Our first product ads were short, beautifully lit and almost silent, and the verdict on them was "no background noise, no naturalness". When we then studied the top creator gadget ads frame by frame, nearly all of them had a person talking over their own sound. None of them were silent.
Sound matters more on some platforms than others, but on TikTok it's close to the whole point. In a study TikTok commissioned from Kantar, 88% of TikTok users said sound is essential to the experience, and 73% said they'd stop and look at an ad with audio.
A real place always has a sound.
Here are the seven fixes we now build into every talking ad our AI influencers make. The numbers are the ones we actually use; treat them as a starting point, not a rule.
1. Give every place its own sound
A real phone video is never silent. A laundromat hums and the drums thump. A street has traffic and footsteps. A kitchen has a fridge and a tap. Take that away and the viewer can't say what's wrong, only that something is.
Most disappointing UGC videos start with a weak brief: a product link and 'make something fun'. Here's a one-page brief template for product videos, field by field, with what to include, what to leave to the creator, and the rules to put in writing.
The fix: for each location in the ad, make one sound bed of that place: its room tone plus a few things that happen there. Then lay it under every shot set in that place, as one continuous track, so the sound doesn't restart at every cut. Text-to-sound-effects models are good at this when you describe the place, not a sound: "a small laundromat at night, machines humming, one drum turning, a door closing far away" works better than "ambience".
It's also the cheapest fix there is. In our itemized cost breakdown, 33 seconds of laundromat sound cost ten cents.
2. Remove speech nobody wrote
Video models that make their own audio sometimes make up lines too. In one of our b-roll shots, a clip of our influencer at a washing machine, the model had her say a cheerful "hi, it's me" that was in no script. It sounded like her, it was in sync, and it was wrong.
The fix: transcribe every generated clip before it goes into the edit. If a shot that should have no dialogue has words in it, drop its audio and put the place's sound bed underneath instead. It takes seconds per clip and catches the most embarrassing kind of mistake: a person saying something nobody approved, in an ad with a brand's name on it.
3. Write lines to be heard, not read
Text-to-speech reads what you give it. Lines written like a caption come out sounding like someone reading a caption.
The fix: write for the ear, then punctuate for the voice.
Short sentences. One idea each. "Both hands busy. Phone's on the case, face down." sounds like a person; one long sentence with three commas sounds like a script.
Punctuation is direction. A full stop is a beat. An ellipsis is a longer pause. A question mark lifts the end. Use them on purpose.
Spell out what should be said. "Thirty degrees" rather than "30°", "seven p.m." rather than "7pm". Models guess at symbols and units, and they don't always guess right.
Check brand names by ear. We had a product name pronounced wrong in a finished ad because it looked like a common word. Listen to every name before you render the video; respelling it as it sounds usually fixes it.
4. Keep one voice for one person
If the first line comes from one voice model and the rest from another, viewers notice, even when they can't say what changed.
The fix: decide on one voice per character and make every one of their lines with it. If you clone a voice from a short recording, make all the lines cloned, including any you made earlier with a different voice. If something else speaks in the ad, like a gadget answering, give it a clearly different voice so nobody mistakes it for the person.
A clean voice track is too clean. Real phone audio has a little room in it and a slightly forward, present sound; synthetic voices often come out boomy and flat at the same time.
The fix: a small, standard voice chain, applied to every line:
High-pass at about 80 Hz to take out rumble no phone mic would pick up.
A gentle cut around 200 Hz (we use about 3 dB) to remove the boxy, muddy part.
A gentle lift around 3 kHz (about 3 dB) for presence, which is where speech is easiest to understand.
Light compression (we use a 2.5:1 ratio) so quiet words don't disappear and loud ones don't jump out.
Then the room tone from fix 1 does the rest: once the voice sits on top of a real place, it stops sounding like a studio booth.
6. Keep the music under the words
Music helps an ad feel finished. It's also the most common reason viewers can't follow what's said.
The fix: three things.
Ask for the right music. Describe it like a brief: style, tempo, a few instruments, and "no vocals". For a talking ad we ask for something like "quiet lo-fi, about 85 BPM, soft drums, warm bass, a few muted piano notes, no vocals". Vocals in the music always fight the voice.
Duck it. Ducking (sidechain compression) turns the music down automatically while someone speaks and lets it come back up in the gaps. Ours runs at a ratio of 9:1 with a slow attack and release, so it breathes instead of pumping.
Make room for the voice. We cut the music by 3 dB around 3 kHz, the same range the voice was lifted in, so the two aren't competing for it.
Underneath all of that, the music stays low: about a fifth of full volume before ducking.
Check the mix on headphones and on a phone speaker.
7. Level the whole ad once, at the end
Generated clips come in at different loudness levels. Cut them together as they are and the ad jumps up and down, and the platform's own normalization may then turn the loud parts down further.
The fix: mix everything first, then level the finished ad once:
Target about −14 LUFS integrated, with true peaks under −1.5 dB. YouTube is widely reported to normalize playback to about −14 LUFS, turning louder videos down (though YouTube doesn't publish the figure itself). TikTok and Instagram don't publish a target, so −14 is a sensible middle that won't be squashed anywhere.
Add short fades. A third of a second of fade in and out at the very start and end stops clicks and an abrupt last frame.
Finally, listen on a phone speaker. That's where most of your audience hears it.
Check the result like a viewer would
One last step catches most of what's left: transcribe the finished ad and compare it with the script. If a meaningful part of the script can't be heard, something is burying it, usually the music or a loud sound effect. We flag an ad when less than 80% of its script comes through in the transcript, and we fix the mix before anyone sees it.
None of this needs a sound engineer. Every step above is a setting or a check, which is why we put them into Influgen's video pipeline instead of leaving them to luck: each talking ad gets its sound beds, a cleaned and levelled voice, ducked music and a transcript check before it's shown to its owner for approval. Sound is also the cheap part: the whole sound bed of a 42-second ad cost us ten cents, while a single talking shot cost dollars. Plug your own lengths into the AI video cost calculator to see where your budget goes.
A good AI UGC ad is decided before anything is generated: in about 60 words of script. Here is the five-beat structure our AI influencers' producer writes to, how many words fit in 20 to 30 seconds, and the mistakes that make a script sound like an ad.
A sponsored reel for a small wearable gadget took us six attempts over two days. Each one failed in a different way, from an invented product to a video that stopped mid-sentence. Here is what broke, why, and what finally worked.
