sample-based bubblegum pop. the song is built around a single three-syllable hook — america · computer · dora — strung from clipped voices and varied into a library deep enough to audition for the ultimate version of the line.

three sample sources, one per syllable:

syllable	source	tone we're after
`america`	"America the Beautiful" performances (Ray Charles, Whitney, et al.)	sung-out vowel, "Aaaa-MEEE-ri-CAAAA!"
`computer`	macOS `say` voices (Fred, Zarvox, Cellos, Bubbles, Whisper, …)	flat synth-voice canon — 1990s mac kitsch
`dora`	Dora the Explorer theme + episode intros	bright kid voice, cartoon downbeat

bubblegum posture (decided 2026-05-26, per Fía's sub-researcher brief) #

bubblegum is earnest production around a goofy hook, not irony layered on irony. the three samples already carry the camp — our job is to play it straight around them.

tempo ≈ 108–116 BPM. fast enough to be kinetic, slow enough that the three syllables resolve as a line, not a tongue-twister.
key: major. D or E major are kind to Mac-voice formants and to most America the Beautiful performances (originally Bb/C). pitch-snap each clip to scale tones; let dora land on tonic or the major third — that's the resolution.
drums: dry kit, kick on 1 & 3, clap on 2 & 4, 16th tambourine. no half-time, no trap.
synths: DX-bell stabs on the hook downbeats, square lead doubling the syllables an octave up, toy-piano on the verse, sub-sine bass root-fifth.
stop-time: drop everything but vocal + claps for one bar before the final chorus. the oldest reliable move in the form.
structure (target): intro (4) → hook ×4 (4) → verse 1 (8) → hook ×4 (4) → verse 2 (8) → bridge (4) → stop-time (1) → hook ×4 (4) → outro (4) ≈ 40 bars, ~1:30 at 112 BPM.

what we are not doing: pitch-shifting everything into doll-voice (kills the human anchor); making the syllable string be the whole song (it's the chorus hook, not the verse); ironic-on-ironic stacking.

pipeline #

sources/<group>/*.mp3        # bin/fetch.mjs       — yt-dlp downloads (gitignored)
clips.json                   # extract.mjs seed    — timestamp annotations live here
utterances/<group>/*.wav     # bin/extract.mjs     — clipped syllables (or say-computer.mjs)
variations/<group>/<src>/*.wav  # bin/variations.mjs  — pitch / stretch / chipmunk / reverse
catalog.json + audition.html # bin/catalog.mjs     — browse + arrange in browser

one-time setup #

# 1. download sources (yt-dlp; copyright-bound, you fetch yourself)
node bin/fetch.mjs --list      # preview curated list
node bin/fetch.mjs              # download all
node bin/fetch.mjs america      # one group

# 2. fill in clip timestamps (start/end in clips.json). either by ear, or:
node bin/extract.mjs --scan sources/america/ray-charles-1972.mp3   # silencedetect candidate segments

# 3. cut utterances
node bin/extract.mjs

# 4. (auto-runs daily) generate TTS computer voices — no source needed
node bin/say-computer.mjs

# 5. build variation library (pitch / stretch / chipmunk / reverse)
node bin/variations.mjs

# 6. catalog + audition page
node bin/catalog.mjs
open audition.html

the audition page #

audition.html lists every clip grouped by syllable. shift-click or double-click to drop into the right-hand "arrangement" pane, then ▶ play in sequence to hear the string. export json dumps the arrangement so a follow-up render script can turn it into a fixed mix.

word isolation (bin/clean.mjs) #

the america/dora cuts came straight off full mixes — crowd, band and theme-music rode under every word, and the whisper transform smears that bed into the breath. clean.mjs fixes it at the source:

demucs (two-stems) separates each source recording into vocals / accompaniment. full recordings give the model context that sub-second clips can't; marathon sources (the 3-hour soy-dora compilation) get a ±15 s proxy window per clip so an 8 GB machine survives. stems cache in .cache/demucs/.
every clip in clips.json is re-cut from the vocal stem, denoised (afftdn), silence-trimmed on both ends — the music bleed is silence now, so the word boundary tightens itself — micro-faded, loudnormed.
results overwrite utterances/<group>/ under the same names.

node bin/clean.mjs --resnap   # separate + re-cut + rebuild snapped/ + hooks.html

falls back to a denoised mix-cut (and says so) if a stem comes up empty for some span. demucs installed via pipx install demucs + pipx inject demucs torchcodec.

the hook machine (bin/hooks.mjs) #

every utterance gets autotuned to every target note its slot needs across the hook variants, plus a whispered twin (FFT phase randomization — the word's formants survive, the pitch dissolves into breath). that's the snapped/ library, indexed by hooks.json. from there every combination of the full phrase — currently 22,218 (14 america × 69 computer × 23 dora) — is playable instantly, no per-combo render.

node bin/hooks.mjs                  # build snapped library + hooks.html
node bin/hooks.mjs flight --count 24 --seed 7   # random stitched mp3s → out/hooks/
node bin/serve.mjs                  # then open http://localhost:7777/hooks.html
node bin/hooks.mjs render --fav hooks-favorites.json  # stitch kept combos

hooks.html: three columns (america / computer / dora), per-slot whisper toggles, melody-variant selector, beat-grid playback. space plays, s surfs a random combo, f keeps it, arrows/,/. cycle each slot. export favorites json and feed it straight to render.mjs --fav.

out/hooks/_flight-all.mp3 is the contact sheet — every flight combo in one listen with a breath between.

bachiamatrixian mode (render.mjs) #

the hook sections now ride a serious four-on-the-floor kick (42 Hz body, tanh drive) with a deep sidechain pump (duck to 0.22, cosine recovery across the beat). a bach arp — 16th-note baroque figuration, chord tones + the diatonic ninth, plucked like a harpsichord pixel — rains above the vocal and takes the full pump. three bach-* hook variants in melody.json give the vocals baroque contours to follow; the variant rotation reaches them on the second chorus. the bridge is the matrix breakdown: kick + sub + arp alone in 8ths an octave down.

the canonical mix (the old bubblegum-only cut is deprecated):

node bin/render.mjs \
  --pick america=whitney-houston-w1 \
  --pick computer=samantha-slow,fred-slow,karen-slow,daniel-slow,moira-slow,ralph-slow,kathy-slow,albert-slow \
  --stretch computer=1.4 --decap computer --whisper computer \
  --pick dora=theme-song-w1 \
  --out out/americomputadora-bach-whisperz.mp3

comma lists in --pick are a roster that rotates per hook phrase — a different computer voice every utterance.
--stretch group=1.4 time-stretches (pitch kept): "cawwwmputerrrr".
--decap group chops the leading consonant — vowel-onset detection (energy + zero-crossing) finds where /k/ ends, so "computer" becomes "ahmputer" and pours straight out of america's open "aaaah". detection runs on the voiced twin even when the whispered file is used.
every vocal clip is RMS-matched after processing so the three words sit at the same loudness in the same space.
--fav hooks-favorites.json --fav-index 2 renders a kept combo.

render pulls clips from snapped/ directly (already tuned + loudnormed, whisper twins included) and only live-shifts when a snap is missing.

current state (2026-06-09) #

✅ all three sample groups cut: america 14, computer 69, dora 23 utterances (+1,802 variations)
✅ america/dora words demucs-isolated from their mixes, denoised, boundary-tightened (bin/clean.mjs, 37/37 clean — no fallbacks)
✅ bin/hooks.mjs — snapped/whispered library (940 wavs), hooks.html combinatorial audition page, flight + render --fav stitchers
✅ bin/render.mjs — full mix with bubblegum kit, bach arps, serious sidechained kicks, --pick/--whisper/--fav hook selection
✅ melody.json — 4 classic + 3 bachiamatrixian hook variants
⏳ pick THE hook: surf hooks.html, keep favorites, render each with --fav, A/B until one sticks

album #

ID3 album: pixsies (lowercase). same shelf as the other pop/ singles.

acknowledgements #

bubblegum-pop genre brief by sub-researcher 2026-05-26 — captured the Kasenetz-Katz lineage through Max Martin to K-pop. see also ../booch/README.md for the boom-bap counterpart.