I\u2019m researching how to build or fine-tune an audio-to-chord-recognition engine comparable in ambition to Song Master Pro \/ Auralis Sound Prism.
The goal is not basic major\/minor chord detection. I need reliable recognition of dense harmonic material: jazz, soul, funk, neo-soul, Brazilian music, film music, and arrangements with chords such as maj9, 6\/9, m9, m11, 13, altered dominants, slash chords\/inversions, secondary dominants, modal interchange, suspensions, passing harmony, etc.
Most public datasets I\u2019ve found seem too limited: either simplified chord labels, weak annotations, or repertoire that does not really cover sophisticated harmony. In particular, I need time-aligned audio + chord labels, ideally with beat\/downbeat information and a rich, consistent chord vocabulary.
My questions:
Which open datasets are genuinely useful for this level of chord-recognition work?
Are there any commercial\/licensable datasets with high-quality, detailed chord annotations that can legally be used to train a model and ship it in commercial software?
Is a dataset such as iReal Pro-style chord charts, Hooktheory, Ultimate Guitar, Chordify, or similar usable in any legitimate\/licensable way \u2014 or are they generally not viable due to rights and annotation quality?
For a serious model, is the realistic route to combine public datasets with a privately licensed\/hand-annotated corpus? If so, roughly how many accurately annotated tracks would be needed before it becomes meaningfully good at jazz-influenced harmony?
Are there papers, benchmarks, companies, or dataset vendors I should study before spending money?
I\u2019m specifically looking for practical, legally usable data sources\u2014not advice to scrape chord sites. Any experience from people who have trained MIR \/ chord-recognition models would be very valuable.<\/p>\n<\/div>\n
submitted by \/u\/DiscoramaMusic <\/a>
[link]<\/a><\/span> [comments]<\/a><\/span><\/p>
