Can Stable Audio make a whole song, or only the sound under it?

TrackType: Music, VocalType: Instrumental, slow indie folk, fingerpicked
steel-string acoustic guitar, upright bass, brushed snare, warm tape
saturation, small wooden room, wistful, 90 BPM

That prompt is the right way to use Stable Audio, and it will never give you a song anyone can sing along to. Stability's own prompt guide for Stable Audio 3 says it plainly: the model is built mainly for instrumental work, it can sometimes produce wordless vocal-like textures, and it is not meant to generate intelligible vocals. There is no lyrics field to fill in. Type a verse into the prompt and you are describing a sound, not handing over words to be sung.

So the short answer to the title: Stable Audio makes the sound under a song, and that is the job it was built for. Full instrumental tracks, isolated instrument parts, textures, risers, sound effects, and edits of audio you already have. The two things a sung song still needs are words and a voice, and neither comes out of this tool.

The two missing pieces, and where each one comes from

A song with a vocal is three layers stacked: the music, the words, the singer. Stable Audio covers the first. The other two have to come from somewhere else, and it pays to know exactly where before you spend an evening re-rolling beds.

The words. This is the part we sell. The lyrics generator takes a short form brief and returns a title and a full lyric with [Verse], [Chorus] and other section tags already in place. One generation costs 20 credits, which at 100 credits to the dollar works out to $0.20 a lyric inside a plan. Credits come with a plan, so there is no paying for a single lyric on its own: the cheapest way in is $5 a month, which brings 500 credits (25 lyrics) every month. The lifetime plan is $19 once and keeps the catalogue and prompt generator open for good, with 1 500 credits (75 lyrics) included once, not refilled. It does not sing and never will. Nothing on sunomarket produces audio.

The voice. A separate tool that actually sings from a lyric. For most people that means Suno, which is a different company: we do not resell it, control it or speak for it. We are not affiliated with Stability AI either. Each of these is its own product with its own terms.

That split is the whole map. Stable Audio for the bed, a lyric from the generator, a singing tool for the vocal. What follows is how to make each hand-off clean, because the joins are where the evening goes.

If you are still choosing between tools rather than trying to get Stable Audio to do more, our comparison of AI music generators by the song you are making is the more useful page. This one assumes you already like what Stable Audio does and want a sung song out of it.

What Stable Audio is actually good at

Read the guide and the product's shape becomes clear. It thinks in sound, not in songs.

• Full instrumental tracks. Tag the prompt TrackType: Music and describe a genre, instruments, mood and tempo. Stability's guide describes generation lengths of several minutes on the larger model, with a shorter ceiling on the small one, and it recommends choosing a duration that fits what you describe rather than always asking for the maximum.
• Single instruments. TrackType: Instrument asks for an isolated part: a solo guitar, a bass line, a drum loop. That is a stem you can drop into a DAW next to something else.
• Sound effects and texture. TrackType: SFX for a door slam, a riser, room tone, rain on a tin roof. The guide suggests setting a short duration for these, since most effects are brief.
• Working on audio you already have. Audio-to-audio (upload a clip and push it somewhere new), inpainting (regenerate one section and keep the rest), and continuation (extend from the end). These are editing tools, not only generation.

Notice what is missing from that list: a place to put lyrics, a choice of singer, a way to tell a voice what to sing on bar nine. That is not a gap in the prompt. It is the design.

Stability also publishes open model weights you can run on your own hardware. That is a different project with its own licence, and our page on open-source music generators covers the local route. Everything here is about using Stable Audio as a tool inside a song workflow, hosted or local.

Why typing lyrics into the prompt does not work

People try it, so it is worth spelling out what happens on paper.

TrackType: Music, pop ballad with female vocal singing
"I left the porch light on for you, it's burning through the night"

The model reads this the way it reads everything else in the box: as a description of audio. It was trained on audio paired with metadata, the kind of text a sound library would write about a file. Nowhere in that training is a line of lyrics mapped to the moment it gets sung. The best case, by Stability's own account, is a vocal-like texture with no intelligible words. The quoted line is not a script. It is a very long adjective.

The fix is not a smarter prompt. It is a different tool for that layer.

How to write a Stable Audio prompt that holds

Stability's guide asks for the same ingredients in roughly the same order across its examples. Here is the order as a working checklist, with what each slot does.

1. Track type tag first. TrackType: Music for a full piece, TrackType: Instrument for a stem, TrackType: SFX for an effect. The repository notes also list VocalType: Instrumental next to TrackType: Music as a pairing that tends to give more coherent output. For a bed you plan to sing over, you want that pairing anyway.
2. Genre. Direct and specific. slow indie folk beats folk. The notes allow Genre: tags, and even several of them.
3. Instruments and rhythm. Name each instrument and how it is played. fingerpicked steel-string acoustic guitar is a sound. guitar is a category.
4. Mood and arrangement. One or two words for the feeling, and a note on how the track moves if it matters.
5. Production. Room, tape, reverb, how close the mic is. This is where the track gets its air.
6. BPM as a number. This is where Stable Audio differs from what we tell Suno users. Stability's guide lists BPM with an example written as a number, 120 BPM. In Suno's Style field we advise describing tempo as a feel instead, because numbers there tend not to hold. Different model, different rule.

Three prompts built on that checklist, one per track type:

TrackType: Music, VocalType: Instrumental, slow indie folk,
fingerpicked steel-string acoustic guitar, upright bass, brushed snare,
warm tape saturation, small wooden room, wistful, 90 BPM

TrackType: Instrument, a lone upright bass walking slowly,
close-miked, dry room, soft finger attack, 90 BPM

TrackType: SFX, a slow reverse cymbal swell rising into silence,
wide stereo, about four seconds

The first is a bed for a verse and chorus. The second is a part you could layer under a vocal track later. The third is the kind of transition that sits in front of a final chorus. None of them asks the model for anything it was not built to do, which is most of the trick.

Turning a Suno-shaped prompt into a Stable Audio prompt

Our prompt generator and our catalogue of nearly 2 000 styles and more than 1 000 artist-style prompts were built for Suno's format. They are not written for Stable Audio and we do not claim they drop in unchanged. The raw material is useful, though, because a good Suno Style string already names genre, instruments and production. It needs three edits to fit Stable Audio.

Suno Style (as written for Suno):
wistful indie folk, fingerpicked acoustic guitar, brushed drums,
breathy female vocal, warm tape, intimate, slow ballad feel

Adapted for Stable Audio:
TrackType: Music, VocalType: Instrumental, wistful indie folk,
fingerpicked acoustic guitar, brushed drums, warm tape, intimate,
76 BPM

The three edits, and nothing else changed:

• Remove every vocal word. breathy female vocal describes something this model does not make, so it spends a slot of the prompt on nothing. The whole voice, every word about it, moves to the singing tool.
• Add the track type tags at the front: TrackType: Music, VocalType: Instrumental.
• Turn the tempo feel into a number. slow ballad feel became 76 BPM. Pick a number that suits the genre and write it down, because you will need the same number later when you fit a lyric to the bed.

Genre, instruments, texture and mood carried over word for word. That is the useful part of a Suno prompt for this job: the band description survives, the singer does not. And if the Suno version came with a lyric sheet full of [Verse] and [Chorus] tags, leave that sheet for the vocal step. Stable Audio has nowhere to put it.

If you have been using Suno prompts in other tools generally, the page on moving Suno prompts into other generators covers the wider pattern.

Three ways from a Stable Audio bed to a sung song

There are three honest routes. They differ in how much control you keep and how much work the joins take.

Route one: let the singing tool make the song, use Stable Audio around it. Write the lyric, generate the whole song in a tool that sings, then use Stable Audio for the parts that tool is weakest at: a textured intro, a riser into the last chorus, a stem to thicken the bridge, ambience under a spoken outro. You layer these in a DAW. This is the route to pick for most people, because the vocal and the arrangement are born together and never fight.

Route two: bed first, vocal on top. You made a Stable Audio instrumental you love and you want someone singing over exactly that. Suno's help centre documents an Add Vocals feature that layers a generated vocal onto an uploaded instrumental, using lyrics you supply. Which plans include it, and what you may upload, are on Suno's own pages and change over time, so check there. This route gives you the bed you chose, and in exchange the lyric has to fit a structure that already exists (more on that below).

Route three: a human singer. Record over the bed. The lyric still has to come from somewhere, and the fitting problem is the same as route two.

Before routes two and three, check two sets of terms: whether your Stable Audio plan lets you use its output commercially, and whether the vocal tool's rules allow uploading that audio. They are separate permissions from separate companies. This is not legal advice, just the two documents to read before you publish anything.

Fitting a lyric to a bed you already have

Here is where most bed-first songs break. The bed has a fixed length and a fixed tempo, so it has a fixed number of bars, and the lyric has to be sized to them.

The arithmetic, for a bed at 90 BPM in 4/4 that runs 2:40:

one beat  = 60 / 90   = 0.667 s
one bar   = 4 beats   = 2.667 s
2:40      = 160 s
160 / 2.667           = 60 bars

[Intro]        4 bars
[Verse]       16 bars
[Chorus]       8 bars
[Verse]       16 bars
[Chorus]       8 bars
[Outro]        8 bars
               ------
              60 bars

Sixty bars, no room for a bridge. That tells you what to ask the lyrics generator for. In the form, structure has a no-bridge option, which is the shape this bed can actually hold. A brief for it might look like this:

theme:     letter to a friend who moved to another city, the porch light
           she used to see from the street, still left on out of habit
POV:       second
language:  en
structure: no-bridge
style:     Folk
hook:      title-line
vocal:     female
emotion:   longing
age:       25-34

vocal tells the lyric writer whether a female or a male singer is in mind, or leaves it unspecified. It does not pick a voice in the singing tool; you still describe the vocal there. style is a closed list of root genres in the form, and Folk is one of them.

Two sizing habits help once the lyric comes back:

• Count lines against bars. A 16-bar verse holding eight lines gives each line about two bars, which is room to breathe. The same verse holding sixteen long lines will be rushed or clipped by whatever sings it. Trim before you upload, not after.
• Keep the chorus short enough to repeat. Eight bars is four comfortable lines. If the generated chorus runs to six, cut to the four that carry the title.

None of this guarantees how a vocal model will phrase the words. It only removes the most common reason a bed-first song sounds crammed: more text than the bars can carry. Our guide to turning a lyric into a song goes deeper on formatting the lyric itself, including why anything outside square brackets gets sung.

What Stable Audio costs, and what you may do with the output

We are not going to quote a price or a plan limit here. Stability sets both and can change them whenever it likes. What matters for a song is the question each plan answers differently: can the audio you generate be used commercially, and under what conditions? Read that on Stability's current pricing and licence pages for the plan you are on, not in a blog post, including this one.

The open-weight models are licensed separately from the hosted product, which is another reason the local route is its own article.

Our position on the words is on our disclaimer: commercial use of the generated lyric is allowed, we claim no authorship over it, the rights are non-exclusive, and two users can end up with overlapping lines. We do not check a lyric against anyone else's rights. None of this is legal advice.

Questions people ask

Can Stable Audio sing? Not in the sense of singing words. Stability describes it as primarily instrumental, able to produce wordless vocal-like textures but not intelligible vocals.

Can I paste lyrics into Stable Audio? You can type them into the prompt. The model treats them as a description of sound, not as words to perform.

Does your prompt generator work in Stable Audio? It was built for Suno's format. The output is a good starting point after the three edits above: remove vocal words, add track type tags, write BPM as a number. Keep the lyric sheet with its section tags for the vocal step.

Is Stable Audio better than Suno? They do different jobs. Stable Audio is strongest at beds, stems and sound design. Suno sings a lyric. One song can use both.

Do I need a DAW? For route one, yes, something that lets you line up the Stable Audio parts against the sung track. Route two can stay inside the vocal tool, subject to that tool's features.

Start with the words

The bed is the part Stable Audio handles for you. The lyric is the part that decides whether anyone remembers the song, and it is the cheapest layer to get right. Open the lyrics generator, set structure to fit the bars your bed actually has, and run one brief. Inside a plan, each title and full tagged lyric costs $0.20. The cheapest way in is $5 a month for 25 lyrics; if you would rather pay once, $19 lifetime brings 75 lyrics and keeps the catalogue and prompt generator open for good.