drone-v1

A 109,128-parameter key classifier: 24 classes (12 tonics x major/minor) from a constant-Q spectrogram. K10 (MIREX 2026) architecture, retrained on 1,363 Beatport EDM tracks. The smallest model to beat the native libkeyfinder reference (69.35% MIREX) on the held-out GiantSteps 604.

Results

Held-out GiantSteps 604, single evaluation per seed:

seed	exact	MIREX
13	68.5%	74.8%
37	66.1%	73.0%
73	65.7%	72.9%
mean	66.8%	73.58%

The shipped drone-v1.pt is the seed-13 checkpoint. drone-v1.onnx is its export (opset 18, dynamic batch, CPU parity verified).

Input contract

• Audio: mono, 22,050 Hz PCM (decode with any standard decoder, e.g. ffmpeg).
• Features: librosa.cqt, 24 bins per octave, 208 bins from two semitones below C1 (fmin ~ 25.957 Hz), hop 4096, log1p amplitude, shape (208, frames). The 192 core bins are rows 8..199; the 16 outer rows are the +/-(4) semitone shift pad.
• Model input: (batch, 192, frames) — the core window of the CQT, any length

  = 1 frame (the head pools over time).

• Output: 24 logits. Class i = tonic i // 12 semitones from C, mode i % 2 (0 = major, 1 = minor).

Usage

Python (PyTorch):

from drone import Predictor

predictor = Predictor()  # resolves drone-v1.pt by default
print(predictor.predict("track.mp3"))  # (tonic, mode, softmax probability)

Or raw, with the ONNX file:

import numpy as np
import onnxruntime as ort
import librosa

y, _ = librosa.load("track.mp3", sr=22050, mono=True)
cqt = np.log1p(np.abs(librosa.cqt(
    y, sr=22050, hop_length=4096, fmin=25.95661061576629,
    n_bins=208, bins_per_octave=24)))
x = cqt[8:200].astype(np.float32)[None]
logits = ort.InferenceSession("drone-v1.onnx").run(["logits"], {"features": x})[0]

Training

Beatport EDM Key Dataset previews (1,363 of 1,486 usable; 933 minor / 430 major), K10 protocol: 320-frame crops, +/-(4) semitone shift as an exact re-slice with a transposed label, AdamW 1e-3, cosine 120 epochs, batch 32, plain cross-entropy. Audio is (c) Beatport; annotations are CC BY-SA 4.0. Full story in docs/training.md in the source repository.

Caveats

Trained on 2-minute EDM previews; expect lower accuracy on long, genre-diverse recordings. Softmax values are uncalibrated.