Give it a sentence about what the track is for and the words are written first, then sung. Writing words runs a language model rather than a GPU, so it costs a flat 5 seconds of the allowance and it needs a language model configured on the deployment. Without one, the other two ways in still render.
Hand it a lyric and those exact words are sung, verbatim. Section markers the engine understands survive into the arrangement, so a chorus you marked as a chorus is arranged as one rather than being read as a line of text.
No words at all, as a flag on the render rather than a hopeful line in the caption, which is the difference between no vocals and probably no vocals. It is also the cheaper rate: an instrumental second costs 1 second of the allowance where a sung second costs 1.25.
Style words, tempo, key and a seed are fields on the same request rather than adjectives hidden in a sentence, and every one of them comes back on the finished track. A caption that produced something you liked is a caption you can send again.
Opening theme for a two person podcast about cycling across a city at night. Warm indie folk, female lead, acoustic guitar and hand claps, ends on the show name.
A word count is wrong by a factor of two between one word and another, and a line count says nothing at all. A digit is its own beat, so a phone number is counted digit by digit rather than as one word, and a run of capitals is spelled rather than read: FM is two beats, and a W on its own is three.
A :10 holds about 19 of them and a :60 about 112, and neither figure is the whole length: a track has an instrumental lift at the front, a breath between sections and a button on the end, and a lyric written for every second leaves nowhere for any of it.
The engine knows a small set of section markers and SINGS anything it does not recognise, so a body containing the word tag in brackets puts the word tag in the vocal. Tag is what a producer calls the sung logo at the end, so that is the word in the interface; on the wire it is [outro].
The sheet shows both, because that mismatch is the one thing about writing for this engine nobody can guess. The markers it does understand are [verse], [chorus], [bridge], [outro], [inst], and anything else in square brackets is a line of the lyric.
The estimate is reported before anything renders, and past 15 percent either side of the target the studio warns rather than just printing the number. That is the honest moment to find out a lyric is long: on the page, not after you have listened to it run over the end of the slot.
A chorus carries more than its share because it is the part that has to be remembered, and a tag carries far less because it is two or three words. At :30 it is given 8 syllables where an even split across the plan would hand it 19, which is the difference between a sung logo and a mouthful.
The longest shape is the last one on the strip, and a render longer than 4:00 gets the same plan with more room inside each section rather than more sections. The engine goes to 10:00, and which plan you are on decides how far up that envelope you can go.
A set of style words and an optional reference track, saved and reused, so the fourth episode theme sounds like the first. NOT a cloned voice: this engine has no speaker embedding for singing, and nothing here promises a particular person’s voice.
Pin the sampler and the same take comes back. A seed reproduces a render only while everything else is unchanged, which is why the caption, the duration and the model are all on the record beside it: a track made a year ago must not claim to have come from today’s checkpoint.
Ask for alternates and they arrive as versions of one track rather than as separate entries: one title, one place in a playlist, several things to listen to. Each take is a render and is billed as one, and the cost of the whole set is stated before you press it.
Re-render a track you own in a different style, keeping the words and the shape. One approved lyric as country, as pop and as rock is three finished tracks off one set of words. It covers only what your workspace owns, which is a licensing boundary rather than a missing feature.
Continue a finished track past its ending. You are charged for the seconds you added and never for the whole result, so a :30 extended by another :30 claims 30 seconds of audio rather than 60. The continuation is a new track with the original on its record, not an edit that overwrote it.
A draft is kept in a library rather than inside one render, and writing it runs a language model rather than a GPU: a flat 5 seconds of the allowance however long the words, with no audio made. Redeeming one copies the words onto the track, so deleting the draft next month leaves the finished piece holding what it was actually sung from.
The engine's ceiling on one track. Every plan sits somewhere inside it, from 1:00 upwards.
At the 4:00 ceiling and the sung rate, which is the dearer of the two. Shorter tracks go further.
About 8 tracks at the free tier's own 1:00 ceiling, for evaluation rather than for release.
| Plan | Longest single track | Stems | Commercial use | Saved style recipes |
|---|---|---|---|---|
| Free | 1:00 | Not on this plan | Evaluation only | 1 |
| Creator | 4:00 | Included | Included | 10 |
| Station | 5:00 | Included | Included | 40 |
| Group | 10:00 | Included | Included | 200 |
Vocals, drums, bass and the rest, pulled out of a finished track in one separation pass however many files come back. It costs 0.75 of the allowance per second of the track it separates, and re-exporting overwrites rather than accumulating.
Renders stop and we ask you to move up a plan. There is no overage line and nothing you do can run up a bill overnight. The whole of that promise, including the refund rule for a render that failed on our side, is stated once on the pricing page.
