BTC-HCQT — an HCQT variant of BTC for chord recognition

An HCQT (Harmonic-CQT) front-end variant of BTC (Park et al., ISMIR 2019) for automatic chord recognition / chord detection — turning audio into a time-stamped chord progression.

Honest result: on a reproducible mir_eval benchmark over held-out public data (GuitarSet, Schubert Winterreise), this model ties baseline BTC — it does not clearly beat it. Shared as a reproducible benchmark, an honest negative result, and an extensible HCQT base — not a new state of the art.

Benchmark (held-out, public data, duration-weighted %)

Model	GuitarSet root / 7ths	Schubert root / 7ths / mirex
baseline BTC	80.9 / 64.6	73.1 / 55.3 / 64.1
this model (BTC+HCQT, Beatles-FT)	80.5 / 63.0	73.8 / 55.6 / 65.3

A dead heat — BTC noses ahead on guitar, this model noses ahead on classical, every gap within the 95% confidence intervals.

Usage, code & full benchmark

Code, weights, the reproducible mir_eval harness, and a guide to extend the HCQT base to melody/bass/transcription: https://github.com/marcusfkelley/btc-hcqt

Limitations & honest notes

It ties BTC; for accuracy alone, baseline BTC is the simpler choice. An early in-house metric suggested big 7th gains — that was a recall artifact (over-calling 7ths); on frame-wise mir_eval it disappears, so always evaluate with standard metrics and confidence intervals. For pure note transcription, purpose-built tools (basic-pitch, MT3, Omnizart) will outperform a from-here build — HCQT is a strong base to extend toward melody/bass/transcription, not a finished transcriber.

Source

Built by Selekt — cleared-sample and music-analysis tools for producers and composers; chord recognition powers features like our chord-progression search. Honest writeup: https://selektaudio.com/chord-recognition

Built on BTC (Park et al., ISMIR 2019, MIT). HCQT: Bittner et al., ISMIR 2017.