Kouano AI
{{ credits }} credits
Self-hosted · open-weight

Two finished songs.
Before the kettle boils.

{{ heroSub }}

No card for your first 5 songs·You own the downloads·Prompts never harvested
{{ heroCaption }}
48 kHz stereo−14 LUFS masterBest-of-2 quality gateSeed-reproducibleInstrumentals on demandREST API

One prompt in. Two masters out.

Every request renders more takes than you see. An automated quality gate throws the weak one away, so you always choose between two good options — never between a good one and a dud.

  1. {{ st.n }}

    {{ st.title }}

    {{ st.body }}

    2.
Try it right here

Type something. Anything.

This box is the real thing — same defaults, same tier switch, same credit maths as the app.

ESTIMATEQUICK TIER
Takes kept2
Credits4
Typical wait~18 s
Master−14 LUFS

Where we actually win

We won't tell you our vocals beat Suno's — they don't, yet. Here's the ground we do hold.

Capability
Kouano
Suno
Udio
{{ r.label }}
{{ r.k }}
{{ r.s }}
{{ r.u }}

Vocal realism: Suno leads. We're not going to pretend otherwise.

Studio tier

Eight tools that keep the song yours.

Repaint

Drag a range on the waveform and regenerate only that. Keep the groove or go fresh.

Stems

Separate vocals and instrumental, rebalance, then generate a mastered mix at the new volumes.

Extend

Lengthen the song as an async job. Honest progress, no fake spinner.

Voices

Clone a voice you own or have permission to use, then sing any track with it.

Jam

Keep the loop going. Audition fast variations, keep the good ones, capture to a song.

Demo Elevation

Balance and master a rough take to −14 LUFS. Optional reference matching.

Crop

Trim to the exact range you want. Non-destructive, like everything else.

A real API, not a waitlist.

One POST returns two mastered takes with provenance attached. Self-host it, or use ours — either way your prompts stay yours.

{{ codeSample }}

Every track can show its receipts.

Provenance on every track

Model, seed, tempo, key, format, gate result, watermark and ownership — auditable, attached to the file.

Private by architecture

Self-host it, or use ours. Either way your prompts and audio are never used as training data.

Consent before cloning

Voice cloning requires an explicit affirmation that you own or have permission for the voice. We record it.

Clean, owned downloads

Watermarked, commercial-ready masters and stems. You keep the rights to what you make.

Kouano Self-hosted AI music. Your prompts, your masters.

What are we making?

Say as little or as much as you like. Defaults are already good.

Lyricsoptional

Choose Write them for me to use your direction, style and selected song length. You can edit the generated lyrics below.
Style
Length · {{ lenLabel }}
Tier

Quick always uses the optimal defaults (8 steps · 7.0 guidance). Use -1 for a random seed each run, or enter a fixed seed. Switch to Studio to edit steps and guidance.
{{ costLabel }}·{{ tierLabel }}·2 takes kept
{{ genPendingNote }}

{{ genTitle }}

{{ genMeta }}

{{ g.label }}
{{ g.note }}

We render more takes than you'll see. The gate scores them on tuning, mix balance and artefacts, then keeps your best two.
Two takes kept

{{ songTitle }}

Quality gate kept your two best takes. A third, weaker take was scored and auto-discarded — you're only choosing between good ones.

TAKE {{ t.n }} PICKED
{{ t.dur }}
{{ t.pendingLabel }}

Library

{{ libCount }}

{{ emptyTitle }}

{{ emptyBody }}
{{ s.tier }}
{{ s.pendingLabel }}

{{ s.title }}

{{ s.dur }}
{{ s.meta }}
{{ s.versionNote }}
{{ toolKicker }}

{{ songTitle }}

{{ sec.label }}
{{ playTime }}{{ durLabel }}

{{ lockTitle }}

{{ lockBody }}

Repaint a range

{{ regionLabel }}

Drag the pink handles on the waveform above — or use the arrow keys. Everything outside the selection is kept exactly as it is.

APPROACH
6 credits
BEFORE / AFTER · {{ regionLabel }}
Before
After · saved as {{ newVersion }}

Crop to a range

Same handles as Repaint. Nothing is destroyed — cropping saves a new version and the original stays in your history.

New duration {{ cropDur }}

Stems mixer

{{ stemsState }}

Separate vocals and instrumental, set new volumes, then generate a mastered mix. Live preview isn’t available — generate to hear and download the result.

WAV or MP3 · or use the current Studio track

{{ regionLabel }}

Drag the pink handles on the waveform above — or use the arrow keys.

{{ stemUploadLabel }}
{{ stemUploadTiming }}
{{ stemUploadError }}

{{ stemRangeNote }} Only this range is mixed and mastered. Playback and downloads keep the full song.

{{ sm.name }} {{ sm.db }}
8 credits · first run ~60 s
Separating sources…{{ stemsPct }}

Extend the song

Adds length from the end of this track using the same prompt and lyrics already on the song — no extra direction needed. Runs as a multi-hop async job (a few minutes); the new version lands in your history when it finishes.

{{ extendBudget }}

{{ extendCost }}
{{ extendJobLabel }}
{{ extendJobEta }}
hop {{ extendHopNum }} · polling every 2 s · you can navigate away

Voices

{{ voiceCount }}

Upload or record 30–60 seconds of clean singing. Save up to four voices, select one, and generate a new vocal version with gentle vocal smoothing while preserving the song’s melody.

{{ v.name }}
{{ v.date }}

Edit reference · {{ voiceEditName }}

Adjust your recording before cloning. The processed preview updates after you adjust a control. Tone controls run in your browser; vocal polish uses the server CPU. No cloning-model generation. Choose 30–60 seconds of clean singing where possible.

Pitch amount 0 is off. Chromatic tuning may alter vibrato and expressive slides; it does not know the song’s key. These controls edit the reference, not the final cloned performance. Accent conversion is not supported.

{{ voiceEditStatus }}

Generating vocals with your selected voice…

Default keeps the existing 77% reference-range adjustment. Source performance uses your converted voice as the full lead over the instrumental, preserving the source pitch range. This mode selects and separates the strongest 20 seconds of your voice reference, then applies source-guided tone, compression, level and subtle ambience. It regenerates the accompaniment with SFT before mixing, like Gradio, without an additional mastering pass. This takes longer and can change the backing arrangement. Effects approximate the original production; they do not reproduce it exactly. No original lead is deliberately blended in. Both modes use 50 conversion steps. Some original vocal may remain if stem separation leaks.

Create a voice

Use 30–60 seconds of clean singing without background music.

Consent required before we process anything

Sing your selected song

{{ recordSongTitle }}

Choose a verse or chorus and sing naturally for 30–60 seconds. These are saved lyrics, not a timed transcript of the audio.

{{ recLabel }}

Recording stops at 60 seconds. Select Done, then name and save your voice below.
RECORD
{{ recLabel }}

Sing naturally in a quiet room. Tap again to stop; recording ends automatically at 60 seconds.
OR UPLOAD
Generating vocals with your selected voice…
{{ voiceHint }}

Jam session

Generate tracks up to four minutes long. Sections play consecutively as one session. Reorder them to update the full playback and download. Select a section to jump to it or continue from it.

{{ jamBudget }}

seconds · {{ jamDurationLabel }}
0:104:00

{{ jamStatus }}

{{ jamError }}

Updating session audio…
JAM SECTIONS · SESSION ORDER
SECTION {{ j.position }} · {{ j.range }}
{{ j.label }} · {{ j.dur }}

Demo elevation

Re-master this track to −14 LUFS. If you’ve already separated stems, vocal and instrumental levels from the Stems tab are applied before mastering.

SOURCE
{{ elevCost }}
{{ elevStage }}{{ elevPct }}
BEFORE / AFTER
{{ ab.label }}
Integrated{{ ab.lufs }}
True peak{{ ab.tp }}
Crest{{ ab.crest }}

Elevation complete — the elevated master is loaded in the studio player. Use Download for the WAV.

Voice → instrument

BETA · VALIDATING

Hum or sing a melody and we'll play it back on an instrument as a stem you can drop into the track. Still validating — pitch detection struggles with fast runs and breathy takes.

1 · SING IT
{{ v2iRecLabel }}
2 · DETECTED MELODY
{{ v2iRange }}{{ v2iConf }}
Record or upload to see the pitch contour
3 · PLAY IT ON
{{ v2iHint }}
Rendering {{ v2iInstName }} from your melody…
{{ v2iInstName }} stem · {{ v2iBars }}

{{ authTitle }}

{{ authSub }}

OR

{{ authSwapText }}

We don't train on your prompts or audio. Self-hosting instructions ship with every account.
{{ onboardStepLabel }}

{{ onboardTitle }}

{{ onboardSub }}

“Two takes, eighteen seconds, and one of them was the record.”

FIRST SONG · FREE · NO CARD

Developers

Keys, usage and request logs for the compose API.

{{ u.label }}{{ u.pct }}
{{ u.value }}
{{ u.note }}

API keys

{{ k.name }} {{ k.masked }} {{ k.used }}

Recent requests

STATUSENDPOINTMSCR
{{ r.status }} {{ r.path }} {{ r.ms }} {{ r.cr }}
60-SECOND QUICKSTART
{{ quickstart }}

Account

Signed in as {{ sessionUser }}
CURRENT PLAN
Creator£19 / mo

400 credits a month, Studio tools on demand, commercial rights on everything you download.

CREDITS
{{ credits }}
of 400 · resets 1 Oct

Invoices

{{ iv.date }} {{ iv.amount }} PAID

Privacy & team

Your audio, prompts and voice models are never used as training data. Enterprise plans can self-host the weights entirely.

Pay for compute, not for permission.

Every tier includes commercial rights and clean downloads. Credits buy generations; nothing is locked behind a licence upsell.

{{ p.name }}

POPULAR
{{ p.price }}
{{ p.credits }}
{{ f.label }}

Quick generation = 4 credits · Studio = 14 · Repaint = 6 · Instrumental = 8 · Extend = 5–10 · Voice model = 12
For developers

One POST. Two mastered takes. Provenance attached.

No waitlist, no partner form. Generate an API key and ship. Or run the weights on your own GPUs and never send us anything at all.

Sync and async

Compose returns inline. Extend and stems return a job id you poll — documented, not hidden behind a spinner.

Provenance in the payload

Model version, seed, tempo, key, gate result and watermark id come back with every take.

Self-host the weights

Open weights, published container, no phone-home. Your prompts never leave your infrastructure.

Keys in one click

Create and revoke from the dashboard. Separate live and test keys, per-key usage.

{{ codeSample }}

Honest about what this is.

Generative music raises real questions about consent, ownership and disclosure. Here's exactly where we stand, including the parts that aren't finished.

Provenance

Every track carries an auditable record: model version, seed, tempo, key, format, quality-gate result, watermark id and ownership. Reuse the seed with the same prompt, lyrics, model, and settings to reproduce a generation. Results can vary across model or software versions.

Privacy

We do not use your prompts, lyrics, audio or voice models as training data. Enterprise customers can run the weights inside their own network so nothing reaches us at all.

Consent for voices

Cloning a voice requires an explicit affirmation that you own the voice or have the singer's permission. We store that affirmation with the model. We do not permit cloning of public figures without rights.

Ownership and watermarking

Downloads are watermarked and commercial-ready — you own the master and can release it. We say commercial-ready rather than guaranteeing legal outcomes, because licensing law varies by territory and we will not pretend otherwise.

What is still unproven

Two features carry a BETA tag because we are still validating them: reference matching inside Demo Elevation, and Voice-to-Instrument. They may change or be withdrawn.

We also will not claim our vocals beat Suno's. They lead on vocal realism; we compete on speed, openness, API access, privacy and ownership.
{{ docKicker }}

{{ docTitle }}

{{ docIntro }}

{{ docApiStatusText }}

Base URL: {{ docApiBase }}

This page documents the worker API. When your local or deployed worker is running, open Swagger for the live OpenAPI schema (same routes the UI calls).

{{ b.heading }}

{{ b.body }}

{{ b.code }}

The Kouano system

One vocabulary across marketing, app and developer surface — editorial serif for voice, neutral sans for interface, mono for anything a machine produced.

Spectrum

{{ c.name }}
{{ c.hex }}

Gradient: hero moments, active waveforms, primary CTAs. Never behind body text.

Neutral ramp

{{ c.name }}
{{ c.hex }}

Type scale

{{ t.spec }} {{ t.sample }}

Icon set

play
pause
repaint
mix
extend
mic
jam
elevate
v2i
crop
download
share
shield
history
lock
key
book
plus
search
check
folder
sliders
dice
lyrics
code
chart
note
loop

Components

BUTTONS · LIGHT
BUTTONS · DARK
BADGES & PILLS
BETA · VALIDATING PICKED 6 credits STUDIO
WAVEFORM PLAYER
STEM CHANNEL
Vocals −4 dB
ASYNC JOB
Generating continuation
~70 s

Motion

{{ m.name }}
{{ m.spec }}
{{ m.use }}

Why it looks like this

Notes on the revamp — the calls made, and what's deliberately unfinished.

{{ r.title }}

{{ r.body }}
{{ generationPendingText }}
{{ toast }}