Technology
Sonilo on Kubeez: AI Sound Effects and Music Written for Your Video
Sonilo is a video-native audio model: give it a clip and it writes effects and score against the timeline. The eight configurations on Kubeez, the real duration limits, the billing floors, and how to call it from the API.
· Kubeez
Most AI audio tools hand you a file and leave the hard part to you. You get a nice explosion, a decent loop, a usable ambience bed, and then you open an editor and spend twenty minutes sliding it around until it lands on the frame where the thing actually happens.
Sonilo is built the other way around. It is a video-native audio model: you give it the clip, it watches what happens on screen, and it writes audio against that timeline. On Kubeez it runs in eight configurations across three tools, and this post covers what each one does, what the real limits are, and how to reach it from the API.
Two ways in: a sentence, or a clip
Every Sonilo product on Kubeez starts from one of two inputs.
From a prompt. You describe a sound and get the sound. No video needed. Two models sit here, one for effects and one for instrumental music.
From a clip. You upload a video and get audio composed against it. Six configurations sit here, which sounds like a lot until you see that it is three creative modes times two delivery formats.
From a prompt
Sound effects from text
Describe what you hear in your head and you get it back. A door on a wet hinge, a crowd two rooms over, a server rack spinning up, a whoosh under a title card.
The useful part is that you pick the length, and the model respects it. The range is 1 second to 180 seconds, prompts up to 2,000 characters. A short impact stays short instead of arriving padded with three seconds of silence you then have to trim, and a three minute room tone is one generation rather than a loop you have to hide the seam on. One caveat: 1 second is the shortest you can ask for, not the shortest you can be billed for. Effects bill from a 3 second floor, so anything under that still costs 3 seconds. See the costs section below.
Run it at /audio/sound-effects.
Instrumental music from text
Same idea, different output. You describe a mood, an instrument set, a tempo feel, and you get an instrumental track. Range is 5 seconds to 360 seconds, and again the duration you ask for is the duration you get. Music bills from a 10 second floor, so a 5 second track and a 10 second track cost the same.
That last point matters more than it sounds. A lot of music models pick their own length and hand you 32 seconds when you needed 20, which means every single use ends in an edit. If you are scoring a fixed slot, a model that honours the number is a different workflow, not a nicer one.
Run it at /audio/music.
From a clip
Upload a video, up to 360 seconds long, and pick what you want written onto it.
Effects. The model reads the motion and the cuts and generates time aligned sound effects. Footsteps land on footsteps. The door closes when the door closes. This is the mode that fixes silent AI generated footage, screen recordings, and gameplay captures.
Music. A score written against the clip's pacing and structure instead of a stock loop dropped underneath it. It follows where the clip builds and where it settles.
Mixed. Both at once, effects and score, in one pass. This costs more than either single mode because it is genuinely two pieces of work, not a toggle.
Each of those three arrives in one of two shapes:
- Audio track. A bare audio file you drop onto your own timeline. Pick this when you are already in an editor and want control over the mix.
- Muxed video. The same generation, delivered as your clip with the audio already in it. Pick this when the next step is uploading, not editing.
Same generation either way. Priced separately because the muxed version does the mux.
Run any of them at /audio/soundtrack.
The limits, in one place
| Thing | Value |
|---|---|
| Text to sound effects | 1 to 180 seconds |
| Text to music | 5 to 360 seconds |
| Video input length | up to 360 seconds |
| Prompt length | 2,000 characters |
| Delivery for clip based modes | audio track, or your clip with sound muxed in |
These are the provider's own ceilings, not Kubeez limits layered on top.
What it costs, structurally
Kubeez prices Sonilo per second of output, not per generation. A 5 second effect and a 90 second score are not the same job and are not billed as if they were.
Two things worth knowing before you budget:
- There are billing floors. Anything that writes music, which includes text to music, video to music, and both mixed products, bills from a 10 second minimum. Anything that writes effects only bills from a 3 second minimum. So a 2 second effect is billed as 3 seconds, and a 4 second music clip is billed as 10.
- Mixed costs more than single mode. It runs both engines.
Live rates for every model sit on /pricing and in /docs/available-models. Those are the source of truth, and they move, so this post does not repeat them.
Sonilo from the API and from MCP
Every Sonilo configuration is available outside the web app under the generate:media scope, through a single tool: generate_sound.
You pass the model id, a prompt, and either a duration or a video URL depending on the model:
generate_sound(model="sonilo-text-to-sfx", prompt="heavy rain on a tin roof", duration_seconds=12)
generate_sound(model="sonilo-video-to-video-sound", video_url="https://.../clip.mp4")
The eight model ids are:
sonilo-text-to-sfxsonilo-text-to-musicsonilo-video-to-sfx,sonilo-video-to-music,sonilo-video-to-soundsonilo-video-to-video-sfx,sonilo-video-to-video-music,sonilo-video-to-video-sound
The video-to-* ids return an audio track. The video-to-video-* ids return your clip with the sound in it. The *-sound pair is the mixed mode.
Two things the tool enforces before it spends anything: a clip longer than 360 seconds is refused before credits are reserved, and a video mode called without a clip never reaches the provider. You get the error, not the bill.
Connect it at /settings/mcp.
When Sonilo is the wrong tool
Be honest about the boundary. Sonilo writes instrumental audio and effects. It does not write songs with vocals, it does not do voiceover, and it does not clean up bad source audio.
- Vocals and full songs: use the music tools at /audio/music.
- Narration and dialogue: /audio/dialogue.
- Pulling an existing track apart into stems: /audio/separation.
Start here
If you have a silent clip sitting in a folder right now, the fastest honest test is the clip based path: upload it, pick mixed, take the muxed delivery, and watch whether the effects land on the frames you would have landed them on yourself.
The full walkthrough with playable samples of all four modes lives on /model-intro/ai-sound-effects-generator-for-video.