esysc69/btc-chord-finetuned

BTC chord recognition — fine-tuned (Magic Chords)

A fine-tuned variant of BTC (Bi-directional Transformer for Chord Recognition, ISMIR 2019), packaged for the Magic Chords pipeline. It improves the chord timeline (root presence and exact label accuracy) on full mixes while keeping the same 25-class major/minor vocabulary as the base model.

• Base model: puar-playground/btc-chord (MIT) — 25-class (large_voca=False) checkpoint.
• Format: a PyTorch state_dict for the inner BTC module, loadable via model.model.load_state_dict(torch.load("btc_finetuned.pt")) on the HF wrapper.
• File: btc_finetuned.pt (~12 MB, 221 tensors).

Intended use

Loaded by the Magic Chords app as an optional primary chord model:

BTC_FINETUNED_CKPT=/app/data/btc_finetuned.pt

When present it drives the chord timeline; key detection is computed separately (an ensemble of the fine-tuned and baseline timelines). When absent, the pipeline falls back to the base model.

Training (summary)

Fine-tuned on a license-clean corpus: the owner's own music library (confidence-gated BTC pseudo-labels; single round, +5,586 chunks), plus synthetic full-mix data and CC-BY symbolic progressions rendered to audio. Improvements (benchmarks, 20 songs): root-presence ~+6 pts, exact labels ~+7 pts, roughly half as many segments.

Licence and attribution

Released under MIT, in line with the base model. Please attribute BTC (Park et al., ISMIR 2019) and this fine-tune.