3AM
AI karaoke where every song is a world premiere. Nobody knows the melody, so performance becomes shared discovery — a party game first; the AI is just the onboarding.
The insight
Traditional karaoke judges you against an original everyone knows. Remove the original and the whole social contract changes: nobody knows the melody, failure is funny instead of embarrassing, and the room experiences a brand-new song together. The pitch bars on screen are the melody's first performance anywhere. 3AM is a social party game that happens to contain a music generator — not the other way round.
The bed system
The architecture's key move is taking generation out of the hot path. Instrumental beds are pre-generated offline across genres, vocals stripped, and each original vocal analysed once into a structural fingerprint — BPM, key, section layout, per-line syllable counts, stress positions, held notes, rhyme scheme.
At party time, a request just selects a bed and has an LLM write personalised lyrics into that constraint grid — "four lines, 8/8/9/8 syllables, line 2 rhymes with line 4, the held note lands on an open vowel" — with a programmatic syllable-validation retry loop, because LLMs are only okay at counting. New words align to the existing melody; the pitch highway renders from the original vocal's contour; the singers follow the guide over the instrumental.
Why the beds matter
Song delivery drops from minutes to seconds, because the expensive generation became a Sunday-afternoon batch job. A provider outage — or a provider lawsuit — can't kill a live party. And the same instrumental reads as a completely different song once it's titled and themed, because the lyrics carry all the personalisation. Live full generation stays available as the slow, premium-chaos option for a room that doesn't mind waiting a round.
Status
Design stage. The first milestone is deliberately, embarrassingly small: a handful of generated tracks, a TV, and a room of friends. Everything else waits on whether that night is fun.