Describe a sound effect, download a pack that is ready to hand over.
Sona renders a batch of variations from one line of text and keeps the takes you mark. The download is a folder of WAVs with UCS filenames, metadata written into the files, matched loudness and trimmed heads and tails.
Free in early access. 150 renders a month, no card.
- UCS filenames
- BWF + iXML
- −23 LUFS
- 48 kHz / 24-bit
- CSV manifest
The loop
Six steps from a line of text to a folder
The keys are the ones the studio listens for. The mouse is optional after the first field.
- 1
Brief
A line of plain text, a count from 1 to 32, and a length up to 47 seconds.
/
- 2
Render
Sona picks the model, varies seed and phrasing across the batch, and streams takes in as they land.
- 3
Audition
A waveform per take. Move, play, loop, keep, drop, or re-roll a single take without the mouse.
↑↓SpaceLKXR
- 4
Name
A UCS category is scored from the brief and the filename pattern fills in. Both stay editable.
- 5
Finish
Loudness target, trim threshold, fades, sample rate, bit depth and channel layout, applied to every kept take.
- 6
Ship
A ZIP of named WAVs with metadata written in, plus a CSV manifest of the pack.
E
Output
What a pack contains
The render is the easy half. This is the half that is normally done by hand.
DOORWood_Heavy Door Slam 03_SONA_001.wav
- CatID
- DOORWood
- FXName + index
- Heavy Door Slam 03
- CreatorID
- SONA
- SourceID
- 001
| Filenames | UCS, from the pattern {CatID}_{FXName} {##}_{CreatorID}_{SourceID}. The pattern and the category are yours to change. |
|---|---|
| Metadata | BWF bext and iXML written into every WAV: description, category, originator, origination date and the brief that made it. |
| Loudness | Integrated to the target you set, −23 LUFS by default, with sample peaks held at −1 dBFS. |
| Edges | Silence below −60 dBFS trimmed from head and tail, then a 2 ms fade in and a 20 ms fade out. |
| Format | 44.1, 48 or 96 kHz · 16 or 24-bit · source, mono or stereo. 48 kHz and 24-bit by default. |
| Manifest | manifest.csv, one row per file: filename, category, brief, model, seed, duration, integrated loudness and peak. |
The studio
One brief, eight takes, five kept
A still from the screen where the work happens. Nothing on this page is interactive.
heavy wooden door, slow open, old hinges
- 01Heavy Door Slam 013.8 s−19.4
- 02Heavy Door Open 024.0 s−27.8
- 03Heavy Door Slam 033.6 s−18.1
- 04Heavy Door Creak 044.0 s−29.3
- 05Heavy Door Slam 052.9 s−17.6
- 06Heavy Door Hinge 064.0 s−31.0
- 07Heavy Door Close 073.7 s−21.2
- 08Heavy Door Slam 084.0 s−20.5
PACK
−23 LUFS · 48 kHz · 24-bit
5 kept · 5 files · manifest.csv
Export packWhat it runs on
Third-party models, local post-production
The models are not ours. Sona sends the brief to an audio model hosted on Replicate and gets a render back, and that step is the only AI in the product. Everything after it — decode, trim, loudness, fades, resample, WAV writing and the ZIP — runs in your browser, so the audio does not make a second trip to a server. Renders are stored privately against your account, in a bucket no other account can read.
Price
Free while it is in early access
150 renders a month, no card. A render is one take, so a batch of twelve spends twelve; auditioning, naming, exporting and re-exporting spend none.
Paid plans will come, and they will add renders, longer lengths and shared naming conventions rather than gate what is already here.
The free allowance will not be reduced without notice.
FAQ
Straight answers
Can I use the output commercially?
Yes. What you render is yours, including for paid client work and for products you sell. Sona claims no rights to it and asks for no attribution. The models Sona calls are ones whose licence permits commercial output; a model that changes its terms is removed from the registry rather than left in place.
Where does the audio come from?
It is generated, not pulled from a library. The brief goes to a model running on Replicate — Stable Audio Open, AudioGen or Tango, depending on what the brief describes — and the audio comes back as a render. Nothing in a pack is a stock recording.
Do you train on my briefs or my audio?
No. Sona trains nothing. Your brief is sent to Replicate as prompt text so a model can answer it, and your renders are stored in private storage tied to your account. They are not published, shared or used as training material.
Which naming conventions are supported?
UCS. All 753 categories ship with the app as data and are matched against the brief locally, with the top three offered and the best pre-selected. The filename is a template — {CatID}, {FXName}, {CreatorID}, {SourceID} and {##} — so a studio convention that reorders or drops fields is a change to the pattern. Any filename can also be typed by hand.
What happens to my packs?
They stay in the account: brief, takes, settings and audio. Re-exporting a pack at a different loudness target, sample rate or bit depth costs no renders, because the processing runs on your machine. Deleting a pack deletes its audio, and deleting the account deletes all of it.
Does it work offline?
No. Rendering needs the network, because the models run on Replicate. Everything after the render — decode, trim, loudness, fades, resample, WAV writing and the ZIP — runs in the browser, so re-exporting an existing pack works on a bad connection.
Is there a plugin?
Not yet. A JUCE shell that points a webview at the studio exists in the repository and has not been released. Sona runs in a browser today.