pop / flwe — "Flower Eater" #

The start of an emo-rock lane for flwe — "Flower Eater", written and performed by Jeffrey Alan Scudder (Whistlegraph, 2021, 74K views across eight tagged posts). A solo sung whistlegraph — one voice, one pen, no accompaniment — about a girl who eats flowers and gets crazy hallucinations.

What's here: the measured foundation (tempo, key, the melody per word, a word-level transcript, a vocalized click track) and the aesthetivox — every vocal through the WORLD speech-to-singing chain onto a derived note chart, per the house rule (cult → loner → here): no lead vocal ships raw; every line is a sung NOTE, not a chopped speech hit.

The takes #

All eight tagged posts are downloaded (bin/fetch.sh re-pulls them from the assets mirror; source/ stays out of git):

bin/takes.py → analysis/takes.json is the per-phrase comparison (loner's take-rotation move): every canonical phrase matched into every other take by fuzzy transcript, measured for voiced coverage, cents to the derived grid (in each take's own fitted tonic — they cluster 130–137 Hz, all near C +20–60¢), and HF noise floor. Verdict: the render stays on the primary for every phrase (the unbroken-take philosophy — one warp, no seams); better_elsewhere marks where a future dub pass should look first: Mystery flower (6950816…, more voiced, cleaner, closer to grid) and Bury them in the ground (6949737…, the primary's is nearly all fry).

What the source gave us #

Tempo — rubato, median ≈ 110 BPM. Onset autocorrelation over the sung region gives 117.2 global, but 10-second windows (octave-folded) drift 97 → 137 BPM around a median of 110.3. This is a hand-and-voice performance: it breathes phrase by phrase. Verdict in analysis/tempo.json: "rubato".

Key — C, about +40 cents sharp, minor-leaning and hand-drawn. The closing "looping every day" chant circles ~134 Hz = C3 +40¢, and the final cadence is "flower" (G3) resolving up a fourth to "eater" (C4) — 5 → 1 in C. The opening strain circles D♯/E♭ minor ("Looked" D♯3, cadence "love" D♯3) before the piece settles onto the C center; E♭ (♭3) outweighs E in the pitch-class weights, with heavy C♯/D neighbor motion. The voice sits ~+20–40 cents sharp of equal temperament throughout — tune any accompaniment to TONIC ≈ 134 Hz, not concert C. Receipt: analysis/key.json.

The melody, per phrase (per-word pyin receipts in analysis/melody.json; times are the primary take's):

The words (whisper.cpp ggml-small.en, -ml 1 word-level; full timestamps in analysis/transcript.json, raw JSON in analysis/whisper-raw/):

Looked for so long for the one that I love. Left with my arms up in the air. My lollipop was the stem I grabbed. Hold up with both of my hands. Mystery flower please don't cower. I am hatin' green now. I'm gonna take you so I can break you up into pieces. Yum yum yum. Now I think I wanna have you every day. I'll spit three of your seeds. Bury them in the ground. Harvest moon I'll come collect your spawn. They're gonna be all mine cuz I'm looping every day, looping every day, looping every day — flower eater.

Cross-take notes: the recital and live takes both hear "three of your seeds" (the primary's "twee" is delivery, not lyric); the live take hears "I am angry now" where the primary gives "hatin' green" — a line to confirm with @jeffrey before words get re-sung.

The click track #

out/flwe-clickvox.wav / .mp3 — the vocalized click study (the kickvox move from cult/loner): a steady click marks the measured grid — soft kick on downbeats, 1.8 kHz tick on beats, 110.3 BPM, 4/4, two-bar count-in — and the primary take's own phrases sit on it dry, each phrase start snapped to the nearest beat (all shifts ≤ ±0.25 s; receipt with measured-vs-snapped beats in analysis/clickvox.json) with the internal rubato left intact. The point: hear where Jeffrey's phrasing agrees with a grid and where the arrangement will have to bend instead.

The aesthetivox #

The full process, in order (study source: pop/loner/bin/halo3.py — the v4pid regulation engine — plus pop/loner/bin/aesthetivox.py and pop/cult/bin/sing.py's Saitou recipe):

1. The notes (bin/notes.py → analysis/notes.json + analysis/melody-chart.json). The scale is derived from the data, not assumed: a frame-level pyin histogram over the sung region, folded into pitch classes in the take's own frame (tonic 133.9 Hz = C3 +40¢), degrees kept above the histogram's own largest gap (≥ 7% of voiced mass). What falls out is not natural minor — it is a hand-drawn 8-degree set:

1 ♭2 2 ♭3 · 5 · 6 ♭7 7 — a chromatic cluster around the tonic, an empty middle (3, 4, ♭5, ♭6 all under 6%), the dominant, and a chromatic cluster under the octave.

Per-word targets are re-measured (median of the last 60% of each word's confidently-voiced frames — the sustain, past the scoop; melody.json's whole-word medians called "that" D♯2 when the note is A2). One pin: the closing "eater" → C4, the 5→1 cadence. The chart quantizes each phrase onto the 110.3 BPM clickvox grid, word durations to 8ths — the chart is a score, not a transcription, and it is what any future arrangement reads.

2. Plosives vs vowels + the render (bin/aesthetivox.py). Per charted phrase of the primary take: WORLD analysis with a fitted floor (probe first — the phrases live anywhere from 67 to 140 Hz); word boundaries pulled to real acoustic events (whisper times the transcript handover, not the note change); trailing decay trimmed, not stretched; 30 ms of attack runway kept, borrowed only from silence. Then the regulation: consonants ride 1:1 and are never stretched; the voiced nucleus carries the note — stretched to fill its beat slot and pulled to the chart target at snap 0.92 (45 ms smoothing; the snap fades out where the pitch genuinely slides). Each vowel onset lands on its beat slot with the consonant leaning in ahead, the way a singer places a word. A nucleus stretched past 1.8× flattens to its target with vibrato fading in over 0.4 s; a phrase that hard-stops gets a synthesized WORLD release (ping-pong of the last 120 ms, faded). Voiced regions are WORLD audio; the warped original is composited back through unvoiced regions (5 ms seams) so /s/ /t/ /k/ stay real, then unvoiced-bright frames get +8 dB back (halo3's sibilant restore).

Receipts in analysis/aesthetivox.json (per word: runway ms, stretch ratio, holds, boundary moves, trims). QC: the rendered vocal sits at a median 9.7 ¢ from its chart targets (75 measurable words).

Out:

• out/flwe-aesthetivox.wav/.mp3 — the sung-note vocal on the click grid (the clickvox study, but every phrase is now NOTES)
• out/flwe-aesthetivox-halo.wav/.mp3 — + octave halo (vowels-only, darkened, ±6¢ pair) and low self-backup at −2/−4 scale degrees, low gain
• vox/NN-slug.wav + vox/.manifest.json — the phrase bank: each phrase's dry render with its lead_in_s noted, so beat 0 is placeable in any arrangement

Re-running #

bin/fetch.sh                                  # pull + decode all 8 takes
../../.venv/bin/python3 bin/analyze.py        # tempo.json, melody.json, transcript.json
../../.venv/bin/python3 bin/clickvox.py       # out/flwe-clickvox.wav+.mp3, clickvox.json
bin/transcribe.sh                             # whisper-raw/ for every take (needs whisper-cli
                                              #   + ~/Models/ggml-small.en.bin, or WHISPER_MODEL=…)
../../.venv/bin/python3 bin/notes.py          # notes.json + melody-chart.json (the score)
../../.venv/bin/python3 bin/takes.py          # takes.json (per-phrase take comparison)
../../.venv/bin/python3 bin/aesthetivox.py    # the render + vox/ bank + aesthetivox.json

Whisper raw JSONs come from whisper-cli -m ggml-small.en.bin -f <16k wav> -ml 1 -oj (whisper.cpp). Everything python runs on pop/.venv (librosa 1.0 + soundfile + pyworld); the key receipt (analysis/key.json) is derived from melody.json.