Songs: procedural art, not a template
Every patch is a deterministic function of hash(text) ⊕ seed. Change the seed and
you get a sibling; keep it and you get the same song back. The layers, in order:
1 · Form detection
Tags already in the source ([Verse 1], [Chorus], [Bridge]…)
are stripped before anything else. A tag line becomes a role hint for the stanza under it, so a
chorus you marked stays a chorus. An intro or outro you tag with lines under it keeps them, and the
words in any tag ([Intro — spoken, low]) lead that section's tag; a word about the voice
replaces the engine's. Empty intro and outro tags are dropped and regenerated. An empty
[Solo], [Guitar Solo], [Instrumental], [Interlude]
or [Break] stays where you put it, as an instrumental break.
[Post-Chorus] is a section of its own, which the engine never guesses;
[Fade Out] ends the song as its outro and fades.
Brackets inside a lyric line are removed. Stanzas are split on blank lines. A stanza that repeats
(fuzzy match on normalized words) becomes the [Chorus]. A short stanza directly before
a chorus becomes [Pre-Chorus]. A late, unrepeated stanza with a different vocabulary
becomes [Bridge]. Single-stanza poems are folded into verse / chorus by repeating the
strongest couplet.
2 · Energy curve from value noise
A 1-D value-noise field (smoothstep interpolation between seeded lattice points) is blended with an archetype curve (rise, wave, collapse, plateau). Each section samples the field once. That number drives everything downstream: which mood word, how many instruments enter, whether the vocal is whispered or belted, whether the peak line is capitalized.
3 · Palette with weighted drift
Each genre family is a palette (instruments, grooves, mix words, adlibs, exclusions). The mutation slider is a temperature: at 0 the engine only draws from the family; higher values let it borrow one or two elements from a neighboring family, which is exactly how the field guide suggests escaping the "pop gravity well".
4 · Grammar for the Style box
A sentence grammar fills the community's master template: identity, vocal, core instruments, rhythm, arrangement arc, dynamics, atmosphere, mix, master. Sentences end with periods, not commas, because Suno parses a period as a hard boundary and a comma as "optional".
5 · Lyric-level typography
Rules from the "dirty tricks" series, applied by energy: ALL CAPS only on the single
highest-energy line (contrast is the whole point), (parentheses) echoes on chorus
tails, inline [adlibs] only in families that use them, syllable stretching
lo-o-ove on a ballad's peak word, *asterisk* cues for atmosphere.
6 · Sung or rapped
Each section gets a rhyme-and-meter read: syllables per line, how often line endings rhyme
(couplets or alternating lines) and how often words rhyme inside a line. Long lines alone are
treated as poetry; only dense, rhyming sections are tagged rapped, tight rhythmic flow
or talk-sung, and the Style box adds a sentence about it. Choruses need stronger
evidence. Set Delivery to "Always sung" to switch this off, or to "Always spoken" to speak
every section, hooks included: each is tagged spoken word, not sung, the Style box
says so, and the syllable advice is left out, as there are no bars to fill.
Families
Seven families from the original engine plus five vocal scenes from Ishkur's Guide: darkwave, ethereal wave, dark synthpop, electroclash and futurepop.