Add Sound to AI Video: Music and SFX That Hit the Cut

Mute the last AI clip you were proud of and watch it again. Notice how much of the magic goes with the sound? Now picture the version you actually shipped: a stock track dragged under it, starting a beat late, peaking in the wrong place, and fading out a second before the final shot. That's how most AI video still gets its audio. The picture is generated. The sound is an afterthought.

It doesn't have to be. Sonilo builds models that read a video and score it: music that lands on your cuts, sound effects placed on the action, and an option to keep the dialogue you already have. All six Sonilo models are live on each::labs. This guide shows how to add music and sound effects to AI video with them, step by step, from a silent clip to a finished MP4.

Why Stock Music Never Fits AI Video

A stock track was written for nobody's footage. Its build lands at bar 17 because that's where the composer put it, not where your product reveal happens. So you trim, you nudge, you crossfade, and the result still feels pasted on, because it is.

AI video makes this worse in a sneaky way. Generated clips have their own rhythm: a camera push that eases in, a cut every two or three seconds, a character turning at a moment you didn't choose. Generic music fights that rhythm. Good scoring follows it.

That's the problem Sonilo set out to solve. When it introduced its first video-to-music model in March 2026, Sonilo said it "synchronizes beats, intensity, and structure with motion, cuts, and narrative arcs," with no prompt required. The v1.1 update in June went further, saying it "locks beats, builds, and transitions to the pacing of your visuals more precisely than v1.0." The point isn't more music. It's music that knows where the edits are.

The Six Sonilo Models, and Which One You Need

Sonilo on each::labs comes as six models, and the split is simple once you see it. Two questions decide which one you call: is the input a video or a text prompt, and do you want audio back or a finished video?

Video-to-Music reads your clip and returns a music track as audio (M4A by default, or WAV or MP3). Its sibling, Video-to-Video Music, does the same job and hands you back the finished MP4 with the soundtrack already laid in.

On the effects side, Video-to-SFX generates synchronized sound effects (Foley, ambience, action audio) as an audio file, and Video-to-Video SFX returns the clip with those effects mixed in.

Then there are two models with no video at all. Text-to-Music writes an original track from a prompt, from 5 to 360 seconds long. Text-to-SFX makes a single sound effect from a description, from half a second up to 180 seconds.

The rule of thumb: if you're editing in a timeline tool afterwards, take audio back. If the clip is going straight to a feed or an app, take the MP4.

Step 1: Start With a Clip Worth Scoring

Everything downstream depends on the picture having a shape. A clip with a clear beginning, a moment of change and an ending gives the model something to score. Ten seconds of a static product slowly rotating gives it very little.

Generate the clip with any video model on each::labs. For an image-led ad, Kling 4.0 Image to Video is a strong start. Our roundup of the best AI video generation models covers the alternatives. Write the motion with the soundtrack in mind:

A glass bottle of cold brew on a dark slate counter. The camera starts close on condensation, pulls back slowly, then a hand slides the bottle forward and ice drops into a glass beside it. Final beat: the full product, centered, warm light.

Notice the structure: a quiet opening, an action in the middle, a resolved end shot. That's three places for music and sound to do something. Upload the result somewhere with a public link, because every Sonilo video model takes the clip as a video_url. On each::labs, the video inputs accept files up to 300 MB, and the music models take clips up to six minutes.

Step 2: Lay the Sound Effects First

Start with effects, not music. Effects are the physical truth of the scene: the clink of ice, the scrape of glass on slate, the room tone underneath. Music sits on top of that, not the other way round.

Send the clip to Video-to-Video SFX with a short prompt. Sonilo describes its effects model as one that "reads a video's timing, pacing, and emotion," so you don't need to list every sound. Point it at what matters:

Close, crisp Foley. Glass on stone, ice cubes dropping into a glass, a soft pour. Quiet kitchen room tone. No music.

The each::labs page for the Text-to-SFX model gives advice that applies here too: describe the sound, the space and the pacing, not the visual scene. "Close, dry, small kitchen" tells the model far more than "a nice coffee shot."

For longer clips, use segments. It's a JSON timeline that lets you treat parts of the video differently, for example silence for the opening shot and full Foley from the moment the hand enters. If you'd rather mix everything yourself, call Video-to-SFX instead and take the audio file into your editor.

Step 3: Score It Without Losing What's Already There

Now take the MP4 from step 2 and send it to Video-to-Video Music. This is where three settings earn their keep.

keep_original_sound keeps the whole original audio track, which is exactly what you want after an effects pass. Without it, your carefully placed ice clinks disappear under the new score. preserve_speech is the narrower option: it keeps only the speech from the source, which is right for a talking-head clip with messy background noise. And ducking lowers the music while someone is talking, so a voiceover or an on-screen line never has to fight the soundtrack.

Then write a prompt for the music itself. Up to 2,000 characters are allowed, but short and specific wins:

Minimal warm electronic, slow build. Soft pulse under the opening, a lift when the ice drops, resolve on the final product shot. Premium, calm, no vocals.

Use prompt_influence (0 to 1) to decide how hard the prompt steers versus how much the model follows the picture. Lower values let the cut structure lead. Higher values pull the track toward your described style. Start in the middle and move from there.

Step 4: Generate Variants, Then Choose

Here's the habit that separates a passable soundtrack from the right one: never take the first take. The music models accept variants_num from 1 to 10, so one call returns several different scores for the same clip.

Listen to them against the picture, not on their own. A track that sounds great in isolation can feel busy under a product shot. The right one usually disappears a little. You notice the clip feels better, not that the music is good.

For longer edits, the same segments idea applies to music. Sonilo's v1.1 announcement describes slicing the timeline and assigning a separate prompt to each segment, like an intro, a verse, a chorus and a bridge. That's how you score a 60-second brand film with a quiet opening and a big finish without stitching two tracks together.

When You Don't Have a Video Yet

Sometimes the sound comes first. A music-led social edit, an app intro, a sonic logo. That's what the text models are for.

Sonilo Text-to-Music takes a prompt and a duration and returns an original track. It can hand back up to ten variants at once, as M4A, WAV or MP3. Generate the track, then cut your video to it. On its own app, Sonilo added vocals to Text to Music in October, so lyrics can become a song there too.

Text-to-SFX is the sound library you never have to search. Need a whoosh for a logo reveal, a two-second notification chime or forty seconds of rainy street ambience? Describe it, set a duration, and drop it in.

On rights, Sonilo says its models are trained only on licensed catalogs and that its output is "cleared for commercial use." Check Sonilo's terms for your specific use, as you would with any audio source.

Put the Whole Chain in One Flow

The full pipeline is four calls: generate the clip, add effects, add music, pick a variant. Done by hand, that's a lot of downloading and re-uploading. Done as an each::labs flow, the output of each step feeds the next automatically, and you can rerun the whole thing on a new product shot with one input change.

A practical setup for an ad team: image in, Kling 4.0 for the motion, Video-to-Video SFX for the Foley, Video-to-Video Music with keep_original_sound and three variants, and a finished MP4 out. Swap the video model, keep the sound chain. Every model involved is behind the same API, and the wider set of video-to-video models slots in the same way.

Frequently Asked Questions

How do I add music to an AI video automatically?

Skip the stock library. Send the clip to Sonilo Video-to-Music or Video-to-Video Music on each::labs, which reads the cuts and motion and returns a synced track as audio or as a finished MP4.

Can I add sound effects to an AI video without recording Foley?

Yes. Sonilo Video-to-SFX generates effects matched to what happens on screen, and Video-to-Video SFX returns the clip with them mixed in. A short prompt about the sounds and the space helps it focus.

Will new music drown out the dialogue in my clip?

It doesn't have to. Turn on preserve_speech to keep the original speech, and ducking to lower the music whenever someone talks.

How long can the video be?

On each::labs, the Sonilo video models accept files up to 300 MB, and the music models take clips up to six minutes. For a longer film, score it in sections.