AIVA composes instrumentals: what do you use when you need a sung song?

Forty cues into an indie RPG soundtrack, the brief changes. The director wants the end credits to roll over a song. Not a cue, a song: a voice, a chorus, words that pay off the last scene. You open AIVA, scroll the style list, and find what you already knew. Every preset is an instrumental.

That is the answer to the question in the title. AIVA composes instrumental music from styles and influences; its own site describes it that way and says nothing about singing or lyrics. When a cue needs a voice, you need two things AIVA does not produce: the words, and a generator that sings them. Keep AIVA for the score. Write the lyric separately, then render it in a tool built for vocals, Suno, for example.

The words are the part we sell. The lyrics generator takes a short form brief and returns a title and a full lyric with [Verse] and [Chorus] section tags already in place, ready to paste into Suno's Lyrics field. Generations run on plan credits: one lyric is 20 credits, and at 100 credits to the dollar that works out to $0.20 a lyric. There is no free trial, so the cheapest way in is the $5 monthly plan, which adds 500 credits (25 lyrics) every month. The other option is $19 once: lifetime access to the catalogue and the prompt generator plus 1 500 credits, enough for 75 lyrics, granted once and not refilled (pricing has the details). We generate no audio, we are not AIVA, and Suno is a separate company we do not resell or speak for.

What AIVA is good at, and why that is not a flaw

AIVA is a composer's tool. You pick a style from a long preset list (the site advertises more than 250), or you upload an audio or MIDI file as an influence, and it writes an instrumental piece in that manner. You can edit the result and export it, and the MIDI export is the part working composers care about: the cue leaves AIVA as notes, not as a baked recording, so it goes straight into your DAW, gets reorchestrated, humanised, rebalanced against dialogue.

That is a score workflow, and it is the right one for most of what a film or game needs. Underscore sits under something else. Its job is to support a scene without asking for attention, and a voice asks for attention by definition. A game with forty minutes of combat, exploration and menu music needs almost none of it sung.

So the useful framing is not "AIVA is missing vocals". It is that a sung song is a different deliverable, with a different input, and AIVA was never built to take that input. A text-to-song generator reads words and a style description. AIVA reads a style preset and a reference. The first can be told what to say. The second cannot, because there is nothing for it to say.

The cues that need a voice

Most of a soundtrack stays instrumental. A handful of moments do not, and they tend to be the ones people remember:

• The end-credits song. The obvious one. It plays once, it carries the emotional verdict of the whole story, and it usually needs words that echo something from the last scene.
• The diegetic song. Music that exists inside the world: a bard in the tavern, a radio in a car, a work song on a ship, a character humming at a sink. The player or viewer hears it as something a person in the story is singing, so it needs lyrics that belong to that person.
• The trailer song. A sung hook over a trailer cut, often a slower, darker cover-style arrangement. Two lines of lyric can do more marketing work here than two minutes of score.
• The lullaby or the anthem. A lullaby a mother sings in chapter one that returns broken in chapter five. A faction anthem a crowd chants before a battle. Both are tiny songs with very specific jobs.

Each of these starts with a lyric, and each lyric has a point of view, a structure and a hook shape that AIVA has no field for. That is where the brief comes in.

Step one: the words

Here is a brief for the tavern case. Every value below is one the generator's form actually offers:

theme      A drinking song for the harbour tavern in a fantasy RPG.
           Sailors mock the lighthouse keeper who never comes down
           from his tower, then raise a cup to him when the fog rolls in.
POV        collective
language   en
structure  loop
hook       chant
emotion    playful
vocal      male
age        any

A few choices in there are doing real work. collective puts the lyric in a "we" voice, which is what a room full of sailors sounds like on paper. chant asks for a hook built to be shouted back, the shape a tavern song needs. loop keeps the verses short and lets the hook do the work, which is how a song a room joins in on is built. The theme is 175 characters of the 300 allowed, and it hands the generator a turn (mockery into a toast) rather than a mood.

The end-credits song wants almost the opposite settings: first or narrator point of view, ballad or classic structure, a title-line or confession hook, an emotion like longing or hope. The theme should name the one image the last scene leaves behind, because a credits song that describes the plot in general terms feels like a book report set to strings.

What comes back is a title and a complete lyric with section tags. Read it as a draft. Change the line that does not sound like your character. Swap a generic noun for the object your film actually shows. That editing matters twice: it makes the song belong to your project, and it is where your own authorship begins, a point the rights section below returns to.

If you want the longer version of what happens between a finished lyric and a finished track, it lives in turning lyrics into a song.

Step two: translate your score into a style line

This is where composers trip. You think in keys, tempi and orchestration, because that is how you brief yourself. A text-to-song generator does not read that language well. Suno in particular does not reliably honour a written key or a numeric tempo, and packing both into the Style field tends to muddy the result rather than fix it. It responds to feel, instrumentation and vocal description.

So the translation runs like this:

And the credits song for that RPG, written as one line:

cinematic orchestral ballad, swelling strings, felt piano, solo cello intro, warm intimate female vocal, slow build, big final chorus

Our prompt generator builds lines like that one, and the catalogue holds nearly 2 000 style prompts and more than 1 000 artist-style prompts in the same shape. Both were built for Suno's format. Other vocal generators take a description too, but they phrase and weight it differently, so if you use something other than Suno, treat our line as a starting point and check that tool's current documentation for how it wants the style described.

One honest limit: a text-to-song generator writes its own melody. It will not quote the leitmotif you built in AIVA, however precisely you describe it. If the credits song must carry the main theme note for note, the realistic routes are either a human singer over your own arrangement, or an audio-upload feature in the singing tool, which Suno's upload and remix article covers. Before you upload an AIVA cue anywhere, read the next section.

Rights: two companies, two sets of terms, plus ours

A song built this way passes through at least two companies, and each one's terms apply to its own part.

AIVA. At the time of writing, AIVA's pricing page ties copyright ownership, credit requirements and where you may monetise to the plan you are on, and the free plan is the most restrictive of them. That matters twice here: once for the score in your film or game, and again if you upload an AIVA cue into another tool as a reference. Read the page for your plan before either happens, because those terms change.

The singing generator. Suno, Udio and the rest each set their own rules on commercial use, often different between free and paid tiers. Those rules govern the audio you render there.

Our lyric. Our disclaimer allows commercial use of the text and claims no rights in it. The rights you get are non-exclusive, the same model writes for every user so passages can coincide with someone else's, and we clear nothing against anyone else's work. Whether a generated lyric is protectable at all is unsettled and answered differently in different countries; the US Copyright Office, for one, has declined to treat prompt-only output as authored. Who owns AI-generated music walks through that in more depth.

This is general information, not legal advice. If a studio contract, a publisher or a festival asks you to warrant who owns the song, show those terms to a lawyer in your jurisdiction.

Where the other tools sit

AIVA is not the only instrumental-first tool, and Suno is not the only one that sings.

If you came here from Soundraw rather than AIVA, the same gap applies from a video editor's angle, and the Soundraw version of this question is written for that reader. If your constraint is running everything locally, some open-source models do generate sung vocals from lyrics; the open-source roundup lists what runs on which hardware and, more importantly, under which licence. And if you are still choosing a vocal tool at all, the best AI music generator by job sorts them by what you need to hand over at the end.

The part none of them replaces is the lyric. Every vocal generator will sing whatever you paste in, and a flat lyric comes back as a flat song no matter which model renders it.

A working order for one sung cue

This is the order that wastes the least of your time on a project that is otherwise instrumental:

1. Decide what the song does in the scene. Diegetic or non-diegetic, who is singing, whether the words must mention something the audience has seen. Write that down in one sentence. It becomes your theme.
2. Brief the lyric. Pick the point of view and hook shape from the job, not from taste: collective and chant for a crowd, narrator for a ballad that tells what happened, confession for a credits song that lands on one admission.
3. Generate two or three drafts and keep the strongest lines. Three drafts take 60 plan credits, $0.60 at plan rates, and the lines worth keeping are rarely all in the same one.
4. Edit by hand. Put your film's objects and your character's words into it. Check that the chorus survives being sung five times.
5. Write the style line from the score's feel, with no key and no BPM, and render in the vocal generator several times. Pick the take that sits with the rest of your soundtrack.
6. Keep AIVA for everything else. The underscore, the menus, the combat loops. For how game music has to be briefed differently from everything else, see music for video games.

Start with the one cue that needs a voice

Do not rebuild your pipeline. Take the single cue in your project that has to be sung, whether that is the credits, the tavern or the lullaby, and fill in the brief for that one.

Open the lyrics generator, set the point of view and hook for the job the song does, write the theme in one sentence, and run it. The cheapest way in is the $5 monthly plan: 500 credits a month, which is 25 lyrics at 20 credits ($0.20) each. If you can already see several cues that need words, the lifetime plan suits a soundtrack better: $19 once, the catalogue and the prompt generator for good, and 1 500 credits, which is 75 lyrics, granted once and not refilled.