How accurate is it, really?

Key: a convolutional neural network now decides the key, with essentia and a chroma read running underneath as a cross-check. Graded on the GiantSteps+ expert re-annotation (596 tracks, one answer key for every detector), the progression is essentia 69.5% exact / 18.5% would-clash, then the first CNN 72.8% / 14.8%, then the model shipping today at 75.7% exact / 13.6% would-clash. The gain came from training data, not a bigger model: the current CNN also learns from full-length DJ library tracks, which is the domain customers actually have. The old 64.9% gauge was graded against the original GiantSteps labels, a different and older answer key the expert re-annotation was made to correct, so it is not comparable to this and we retired it. The shipped model (models/key_cnn.onnx) is held to this frozen score by a ship-blocking parity test (tests/test_keycnn_parity.py). One platform note: these are the CNN's numbers, and on Mac the CNN runs on Apple Silicon. An Intel Mac uses the essentia engine underneath instead, the 69.5% exact shown as the starting point above, because the key model has no Intel build; nothing else about the app differs. BPM: a convolutional neural network now reads the tempo from the audio, the same on-device approach as the key. Graded on the GiantSteps tempo benchmark (664 tracks, Schreiber and Muller's canonical labels), it reads the tempo correctly on 89.0% of tracks at the standard 4% tolerance (the MIREX Acc1 metric), up from 80.8% for the previous DSP detector on the identical set and metric. Both numbers sit on the same ruler, so the gain is the engine, not a looser tolerance. Octave-forgiving (Acc2, counting a half or double reading as right) it is 97.9%, up from 86.2%. These are measured through the shipped path, the app's own decode into a pure-numpy mel into the ONNX model, so the page reports what the app actually does, and the model (models/tempo_cnn.onnx) is held to the score by a ship-blocking parity test (tests/test_tempocnn_parity.py). Two honest limits: the benchmark is 2-minute previews, its own domain, so full-length tempo accuracy is unproven and we do not claim it; and detection runs on your machine, nothing is uploaded. GiantSteps still counts a half-time or double-time reading as fully wrong at this tolerance, a 174 track heard at 87 is a miss even though your pitch-fader math still works, and we report the strict number anyway. One caveat on the truth itself: GiantSteps' tempo labels come in two academic rounds, the 2015 set we grade against and a 2018 crowdsourced re-annotation, and the two disagree on about a third of tracks (68.1% agree within 1 BPM, 9.1% differ by an octave or level, 22.8% by more), so the 664-track ground truth is not fully settled and we grade against the 2015 canonical labels. Same harness, reproducible from the repo.

Essentia cross-check run measured 2026-07-15, frozen in the repo (tests/giantsteps/RESULTS_2026-07-15_gsplus.txt); high confidence means an internal essentia score of 0.7 or above. Truth is the GiantSteps+ expert re-annotation, scored on 596 tracks (8 of the 604-track key set have no mappable expert key). These 78.4% / 82.2% / 11.8% figures describe the essentia cross-check, not the key the CNN shows; for the shown key the numbers to hold are the CNN's own, 75.7% exact and 13.6% would clash in the gauges above.