Direct Your AI Vocalist: Tell It How to Sing, Not Just What to Sing

Your Singer Is Not a Mind Reader

Here's a thing most people get backwards: they spend twenty minutes picking vocal cues — gravelly, low, aggressive, female, whatever — paste in lyrics, hit generate, and then wonder why every section sounds like the same person standing in the same room saying words.

That's because you told the AI who the singer is, but you never told it how to perform. Big difference.

Toxic Sound generates your song in a single pass — the music, the words, the voice, all at once. That means the [VOCALS] cues in your prompt describe one singer for the entire song. You can't swap in a different person for the chorus. The AI can't do a two-step — it reads the prompt once and commits. So the [VOCALS] block is where the singer's identity lives: their register, their grit, their tone. That's the casting decision.

The performance — the whispering, the shouting, the laid-back drawl — that lives in your lyrics, in brackets, section by section. Think of it like directing an actor: you cast them in the prompt, and then you give them stage directions in the script.

The Trick: Bracketed Performance Notes

Look at this. It's from a real track on the platform, "Shatter the Gilded Cage," a rock track about tearing down false faith:

`

[Intro]

[whispered]

Breathe in the dust of the altar floor

Count every promise that locked the door

...

[Verse 1]

[sung]

I bent my knees until the marrow bled

...

[Chorus]

[shouted]

Break through the ribs of the holy cage

...

`

See the pattern? Every section gets a structural tag ([Intro], [Verse 1], [Chorus]) and a delivery tag ([whispered], [sung], [shouted]). That whisper-to-shout arc is the whole track. The intro creeps in close, the verse tells the story, the chorus blows the doors off. Same singer. Three completely different energies.

That's the lever. Short tags, right under the section header, telling the voice what to do. More examples that work:

• [delivered fast] — pushes the flow forward, great for rap verses and metal
• [aggressive] — leans into the attack without changing the pitch
• [laid-back] — lets the bars breathe, good for phonk and boom-bap
• [growled] / [hushed] / [chanted] — each one is a different animal

You already speak this language if you've ever read a screenplay. It's just stage directions for a voice that only exists in your prompt.

Rule One: One Singer, One Show

This is the mistake that wastes the most credits. Somebody writes [VOCALS] soft female folk singer in the prompt, then tags the chorus [demonic deathcore growl], and gets confused when the result is a confused murmur.

The model can't be two people. It hears two conflicting identities and splits the difference — which is how you end up with the vocal equivalent of lukewarm soup.

Here's the line: register contrast is fine, identity whiplash is not. Whispered-to-shouted works, because it's the same voice at different intensities. Sweet-sung-verse-to-snarled-chorus works, because it's the same person getting angrier. But a whole new vocalist in the bridge? That's not how a single-pass engine works. Cast your singer once in the [VOCALS] cues, then direct them.

Rule Two: Short Tags, Not Paragraphs

The generator reads your bracketed notes more like punctuation than poetry. So:

✅ [delivered fast, building anger]

❌ [this verse should be delivered fast with mounting urgency, like the singer is running out of time as the drums crescendo beneath them and the walls are closing in]

The second one won't hurt anything, exactly — but the model mostly keys on the shape of the instruction, not the novel. Two to four words. Deliver the emotion, not the essay. Your long paragraph eats attention the model could have spent on actually singing the line well.

Rule Three: Tag the Changes, Not Everything

Contrast is information. If every single section of your song is tagged [sung], you haven't told the model anything it didn't already know — that's the default. Tags earn their keep when they mark a shift.

Untagged sections fall back on whatever your [VOCALS] cues describe. That's not a bug, that's a feature. It means you can write:

`

[Verse 1]

[whispered]

[Verse 2]

[Chorus]

[shouted]

`

Verse 2 inherits the base singer — no tag needed, no tag wasted. The whisper and the shout pop harder because Verse 2 didn't get in the way. Think of tags like hot sauce: a few drops on the right sections, not poured over the whole plate.

Rule Four: Voice Notes, Not Production Notes

Keep your bracketed notes about the voice. [beat drops here] or [guitar solo] inside your lyrics is a coin flip — the engine decides production from the music side of your prompt, not from stage directions in the lyric block. Sometimes it'll do something with it, sometimes it'll just... sing the words "beat drops here," which is objectively funny the first time and infuriating the second.

The same goes for [pause] and [silence] — use them sparingly and only where a real singer would actually breathe. The generator already handles musical pacing; your job is the human on top of it.

Rule Five: Pair Notes With Your Cue Choices

The vocal cue pickers in the studio aren't decoration. The range cues in particular — the ones that let you push a voice toward whisper-quiet or full-throated roar — exist to set the ceiling and floor of what your singer can do. If you picked a hushed, intimate vocal set and then tag every chorus [shouted], you're asking the singer to do something you didn't cast them for.

Work with the grain: pick cues whose range covers the performance you're writing notes for. Whisper notes want low, breathy sets. Shout notes want aggressive, high-energy sets. The note tells them to move; the cue choice decides how far they can go.

Try It on Your Next Track

Next time you're in the studio, do the whole thing in one pass: cast the singer in [VOCALS], write the lyric story, then go back through with a director's eye and tag the shifts. Intro whispers. Verse settles in. Pre-chorus tightens. Chorus detonates. Outro exhales.

Then generate and listen for the arc. If the track sounds flat, don't touch the cues first — check your notes. Nine times out of ten, the song was fine. It just didn't know it was allowed to perform.