Add Sound Effects to a Video with AI, Timed to the Frame

Start with a silent clip

If you have a clip, upload it. We made ours in Ciyo: one 8-second vertical close-up of hands tearing a croissant over a plate, flakes falling, then a cup of coffee set down beside it. It cost 9 credits at the cheapest video setting, and we asked for no audio.

Silence matters. Sound that a video model adds by itself competes with the effects you place later, and you cannot take it out cleanly.

I run Crumb & Kettle, a small bakery. For an Instagram Reel, make me one silent 8-second vertical clip (cheapest video setting, no audio): close-up of hands tearing a fresh croissant open over a white plate, flakes falling, then a cup of coffee set down beside it. Keep the camera still. Don't add any sound yet; I want to add sound effects in the next step.

Name the moments, not only the sounds

A list of sounds is not enough. Say what each sound belongs to, so the agent can find it in the picture: the crust splitting, the halves pulling apart, flakes landing, the cup touching the table. Add a room tone for air, and say how quiet it should be. Ask for the timestamps, so you can check them.

The clip is good. Now add sound effects so they land on the actions: the crisp crackle when the crust splits, the soft stretch as the halves pull apart, flakes landing on the plate, and the cup touching the table, plus a quiet café room tone underneath. Find the exact moments in the video yourself, tell me the timestamp of each effect, and keep the room tone low so the actions stand out. No music, no voice.

How the agent found the moments

The agent stepped through the video and put a timing sheet on the board: 32 frames, one every 0.25 seconds, each stamped with its time. It is a working reference you can delete, and it is also the fastest way for you to check the result.

Then it generated five sounds and mixed them into a copy of the clip, keeping the original. The five sounds cost 7 credits; the mix cost nothing.

The timestamps it chose

The agent reported each effect with the reason for its timing, for example that the cup lands at 6.80 seconds “when the cup's downward movement stops”. The room tone runs the whole clip about 20 dB below the effects.

The frames below are taken from the finished video at those four times.

We measured the audio

We downloaded the finished MP4 and measured its loudness every 50 milliseconds. The room tone sits at about −46 to −48 dB. The cup hit is the loudest moment, −18 dB, and it starts at exactly 6.80 seconds. The flakes rise at 3.70 seconds. The crackle is placed at 1.95 seconds, but its loudest part comes at 2.5 to 3.0 seconds, when the crack is widest, which is still inside the span the agent chose.

So the room tone is about 28 dB below the loudest effect, a little quieter than the agent's “about 20 dB”, and the whole Reel measures −19.7 LUFS. If it sounds quiet next to your other Reels, ask for a louder mix; the agent offered to remix without new sounds, for free.

Check it with your ears, then adjust

An AI agent can measure a waveform, but it cannot tell you whether a crackle sounds like a croissant. Listen with headphones and on a phone speaker. For each effect, ask two questions: does it start on the frame where the action starts, and does it sound like the object? If a sound is early or late, give the agent the frame time from the timing sheet and the direction, for example “move the crackle 0.3 s later”.

Next steps for the Reel

Effects, music and voice work together when each has its own space. Add music under the effects or a short voiceover after them, and keep the silent original for the next edit.

• Add background music to a video
• Add a voiceover with Eleven v4
• Turn a photo into a video
• Make an Instagram Reel cover