A vocal prompt framework: six slots, no adjective soup

The adjective soup problem

A typical vocal prompt reads like this: 'emotional powerful raspy soulful haunting beautiful female vocals'. Seven adjectives, most of them describing how you want to feel rather than how the voice should sound, several of them pulling in opposite directions. The model averages them, and the average of everything is nothing in particular.

The fix is to stop free-associating and fill fixed slots. Six of them cover almost every voice you'd want, each answering a different question, none overlapping. Fill each once. Leave one empty rather than guessing.

There's a second failure hiding in that example: most of those words describe the listener's reaction rather than the performance. 'Emotional' and 'beautiful' are outcomes you hope for, not instructions a model can act on. The slots below deliberately contain only things you could point at in a recording.

The six slots

Writing it out

Read in slot order that line is unambiguous. Every word is doing distinct work, and when the result is wrong you know exactly which slot to change — which is the real payoff. 'Too polished' is a texture problem. 'Too far away' is a distance problem. 'Wrong feeling' is attitude. You never have to guess which of seven adjectives to delete.

Keep the line as a single comma-separated string in slot order rather than scattering vocal words through the rest of the prompt. Descriptors that drift into the instrumentation section tend to get applied to the instruments instead, which is how people end up with a husky-sounding guitar and a perfectly clean voice.

Vocal: male baritone, husky, conversational, close-mic intimate, dry with light plate, resigned

Slots that conflict, and what wins

• Whispered delivery plus belted attitude — delivery usually wins, attitude gets lost. Pick one.
• Close-mic intimate plus heavy reverb — these fight; the reverb generally overrides the intimacy.
• Raspy texture plus clear texture — the model picks one arbitrarily, run to run.
• Distant plus dry — technically coherent but often sounds thin; distance implies some room sound.
• Multiple range values ('female alto to soprano') — you get an unstable, drifting performance.

When results are inconsistent between runs with an identical prompt, a slot conflict is the usual cause. Strip back to one value per slot and the variance drops noticeably.

Section-specific vocal direction

You can change the voice across a song without changing the singer. The productive move is to shift one slot at section boundaries — usually delivery or effect — while holding range and texture constant. That's what a real vocal performance does: same instrument, different intensity.

Keep the global vocal line in your style prompt and put section direction inline with the lyrics where the platform supports it.

Be conservative about how many sections get direction. Changing the voice at every boundary reads as inconsistency rather than dynamics. Two shifts in a song — usually one lift into the chorus and one drop in the bridge — is enough to create an arc without making it sound like three different singers took turns.

[Verse 1]   (delivery: conversational, close)
[Pre-Chorus] (delivery: rising, slight push)
[Chorus]     (delivery: belted, doubled, wider)
[Bridge]     (delivery: whispered, dry)

Testing a voice properly

• Write the six-slot line and generate twice to see its natural variance.
• Change exactly one slot and generate twice more.
• Listen only to the slot you changed — ignore the mix, ignore the melody.
• Log the slot, the old value, the new value, and the audible effect.
• Build a personal shortlist of values that reliably work in your genres.
• Reuse that shortlist as your starting vocal line on every new project.

FAQ