Blind composition duels · Note Duel

3 min read Original article ↗

A blind listening study of AI-composed music

Two pieces. One brief. You decide.

Hear two compositions written for the same brief, labelled only A and B. Pick the better music. Then we show you who wrote what, and how everyone else voted.

About a minute per duel · no sign-up · headphones help

20votes cast

16full-listen votes

13listeners

0/81duels you've judged

How a duel works

  1. 1

    Listen blind

    Both pieces play through the same engraver and sampled piano at matched loudness. Sides are shuffled for every listener.

  2. 2

    Vote

    A, B, tie, or neither. Add why if you like. Hearing both in full makes your vote count in the headline standings.

  3. 3

    Reveal

    See which piece was which, whether it met the brief, and how the crowd split.

Pick a brief

Every model answered the same 10 briefs. Choose one, or press Start a duel and we'll pick the pairing that needs votes most.

Built to be fair

Blind & shuffled

Titles and sources stay hidden until you vote. Each listener gets a random A/B placement, fixed on reload.

Same instrument, same volume

Every score is engraved and played by one sampled piano at one fixed touch, loudness-matched, so production can't win the vote.

Full listens count

A vote counts as a full listen only if you heard at least 90% of both pieces. Partial listens are kept, but apart.

Uncertainty, always

Rankings use a Bradley–Terry model with 95% bootstrap intervals and rank ranges, and stay provisional under 30 judgments.

Per-task results

Every brief's head-to-head is published on its own, so one overall number never hides where a piece won.

Neither ≠ tie

“Both equal” and “neither is good” are different answers. Rejections are reported separately rather than mixed into the ranking.