Benchmark

By kind of recording

How fast it locks on

Each engine hears only the first few seconds of every clip. A line shows how many it names correctly by then; an engine has locked on once it is right and stays right with every extra second, up to 30. Tune Companion is replayed exactly as the Tune finder runs live, scoring once a second from the first second. JigRipper scores 8-second windows, so it answers from 8 s; FolkFriend and Tunepal answer whenever they can hear notes. The answer an app shows can come later still, as each waits until it is confident.

By instrument

Top-1 accuracy. A shaded cell is a clear win: on the same clips, the leader got right significantly more than the runner-up did (exact McNemar test, p < 0.05). A bold cell without shading leads, but by too little to be sure.

By popularity

Bands follow each tune's rank on The Session by how many members' tunebooks hold it, across the whole catalogue, so "rare" means rare. The clips so far are almost all of popular tunes, which were the easiest to find and label: these rows will fill in as rarer recordings are added.

How it was measured

• Every engine hears the same audio: the whole clip, usually 30 seconds to two minutes. An answer is right when it is the same tune on The Session, in any setting. Bars show top-1 accuracy; the thin line across each bar is its 95% interval. Hover or tap a bar for counts and top-5.
• Tune Companion: model.onnx, the Tune finder's own model, 23,305 tunes.
• FolkFriend: command-line build (commit d1faa84), 23,385 tunes. It is open source; its audio is upsampled to the 48 kHz it expects.
• JigRipper: web model #1006, 22,410 tunes. It merges some duplicate tunes; naming the tune a duplicate was merged into counts as right. Replayed offline with its own published web model and front end.
• Tunepal: its published method on its own tune index, 12,241 tunes. 1 benchmark tune is missing from its index, so it cannot name it. Its search and tune index come from its author's open-source rebuild (TunepalGodot); the app's own pitch tracker is not public, so a standard one (YIN) stands in for it, with the same settings on every clip.
• The clips are stored at 16 kHz, which removes the highest overtones for every engine alike. Recordings stay on YouTube; this page only links to them. Last run 1 November 2026.

Every clip