Two takes
Start with the vocal sketch to hear the words on the tune, then the instrumental to hear the melody clean.
Maeve — spoken over the full arrangement
Pick this oneBoth key words now land on their written note — “Treats” at 1.11s against the B4 at 1.15s, “you” at 3.71s against the held C4 at 3.75s. The bed is carved around the voice (a dip in the speech range, not just turned down) rather than fighting it. Full arrangement underneath: bells, pad, upright bass, shaker, vibraphone answer. This is the closest thing to a finished outro right now.
download mp3Pacing comparison
Why this takeSame voice, three reads. The first pass used an ellipsis in the prompt and came out slow — below natural conversational pace. The tight version above measures 3.7–4.5 syllables/second, which is the normal range for spoken English. This is what "conversational" sounds like next to "not conversational."
Original — 2.9–3.5 syl/s, slow
Slower alternate — 2.8–3.4 syl/s
The arrangement alone
InstrumentalBells, pad, bass, shaker, vibraphone — no voice. Use this if the tagline runs as on-screen text instead, and as the guide track for anyone recording a real vocal.
download mp3English vocal sketch
Synthetic singingA formant synthesizer actually singing the English words — glottal source shaped by vowel filters. It sounds robotic by construction. Judge whether the words sit right on the melody, not the voice quality.
download mp3Instrumental melody
Bells + padThe tune on its own. This is also the guide track — whoever sings it plays this in earbuds and sings along.
download mp3It ends on “you”
The last word of the whole reel points at the person watching. That turns a brand tag into a compliment — which is exactly the voice in Mimi's brief: neighbor, not brand.
Nine syllables, about six seconds. It works sung, spoken, or as plain on-screen text, so it can become a real audio signature rather than a one-off.
One arch, an octave wide
It climbs to the brand name, then walks all the way back down and settles on “you.” Rising is the brand announcing itself; falling is the warm aside.
The trick worth keeping: “Treats” lands on B4 — the leading tone, one half-step under home. Your ear wants resolution, so the second line feels necessary instead of tacked on. It finally lands on C4, a full octave below the peak.
Key of C major · 104 BPM · 6.4 seconds
| Syllable | Note | MIDI | Beats |
|---|---|---|---|
| Mi | E4 | 64 | 0.5 |
| mi's | G4 | 67 | 0.5 |
| Sweet | C5 | 72 | 1 — peak |
| Treats | B4 | 71 | 1.5 — leading tone |
| (rest) | — | — | 0.5 |
| not | A4 | 69 | 0.5 |
| as | G4 | 67 | 0.5 |
| sweet | F4 | 65 | 1 |
| as | E4 | 64 | 0.5 |
| you | C4 | 60 | 3 — home |
6.4 seconds, pinned to the notes
Every visual hit lands on a note rather than a stopwatch. Sizes: 1080×1920 for reels, 1080×1350 for feed.
- 0.0 – 0.6s · bells enter
The reel's last frame holds. A soft white bloom opens from the centre and washes the frame down to a clean dark ground.
- 0.6 – 1.6s · the C5 peak
Wordmark scales up from 92% with a slight overshoot and locks frame-exact on the peak note. Red underline draws left to right beneath it.
- 1.6 – 2.4s · B4 ringing
Wordmark settles. Three or four gold sparkles drift up and fade. Restrained — not confetti.
- 2.4 – 2.9s · the rest
Complete stillness. Nothing moves. This half-second is what makes the next line land — resist filling it.
- 2.9 – 4.4s · descending line
Tagline fades up under the wordmark in white, lightly letter-spaced: not as sweet as you
- 4.4 – 6.4s · C4, held
The word you warms from white to gold. Red footer bar slides up with 320 Main Street · Delta, Colorado. Hold, then fade over the last third of a second.
Getting a real voice on it
One fork decides everything, and it's easy to miss: some tools write their own melody, others sing yours. Suno and Udio compose their own tune from a text prompt — you'd lose the arch above. Singing-synthesis tools take MIDI plus lyrics and sing exactly the notes you hand them.
Record a real person Recommended
FreeIt's nine syllables — a thirty-second ask of one human. Mimi's whole voice is “neighbor, not brand,” and a real Delta voice with a little imperfection is more on-brand than a polished AI vocal. No licensing questions, and it keeps the melody exactly.
How: play the instrumental above in earbuds, sing along, record a phone voice memo somewhere quiet. Three or four takes. Send the file over and it gets aligned, tuned and mixed onto the bells.
Synthesizer V or ACE Studio
Small subscriptionPurpose-built singing synthesis: load the MIDI notes and English lyrics, pick a voice, and it sings our melody. The technically correct tool when a specific tune already exists. Verify the commercial terms on the particular voicebank — those vary per voice even inside one product.
Suno or Udio
Small subscriptionCheap, fast, good output — but it writes its own melody from the lyric. Right if you'd rather audition fresh musical ideas; wrong if you want to keep this one.
Suggested sequence: ship the instrumental sting with animated on-screen text now, so the outro exists and starts building recognition this week. Drop a real vocal into the same 6.4-second slot whenever a take arrives — melody, timing and storyboard don't change, only the vocal layer swaps in.
Why not VocalRender
It was the original candidate. Licensing is fine — Apache‑2.0, commercial use permitted. Three other things stop it:
Its released checkpoints primarily target Mandarin, so English output is genuinely untested rather than impossible — some English training data is in the mix, but nobody has published how it performs. It also requires two to eight seconds of reference singing on every run, meaning you need a singer before you can synthesize one. And it needs a CUDA GPU, which this workstation doesn't have.
Not “incapable” — just an expensive coin-flip to answer a question its own authors hedge on. Worth an experiment someday; not worth a client deadline.