alberto chord recognizer (CRNN, triads and sevenths)

A small convolutional-recurrent network that labels guitar audio with one of 73 classes per frame: no chord, or a major, minor, dominant-7th, major-7th, minor-7th or half-diminished-7th chord on any of the 12 roots. It is the neural recognizer of alberto, a guitar songwriting helper that listens to a progression, estimates the key, and suggests what to play next.

Parameters: 482,025 (checkpoint under 2.5 MB)
Input: log-magnitude constant-Q transform, C2 upward, 6 octaves, 24 bins per octave, hop 512 at 22.05 kHz (config.json has every value)
Output: per-frame logits over 73 classes (config.json lists them in order)
Decoding: alberto runs Viterbi over the frame posteriors with a bigram transition matrix stored in the checkpoint (transitions, estimated from the training labels); plain argmax works too, at a small accuracy cost
Runs comfortably on CPU: ~16 ms for model + Viterbi on an 8-second window on an M-series Mac

Training

Trained on GuitarSet (Xi, Bittner, Ye, Bello, ISMIR 2018; CC BY 4.0), mono-mic recordings, using the performed chord annotation mapped to the 73-class vocabulary (references with other qualities — 9ths, 6ths, sus — reduce to the triad their third implies). Players 00–04 (300 excerpts, strummed and solo) were used for training with random pitch shifts of up to ±5 semitones applied by rolling CQT bins; player 05 (60 excerpts) was held out. 30 epochs, Adam 1e-3 with cosine decay, 8-second chunks, cross-entropy.

Evaluation (mir_eval, held-out player 05, 30 strummed excerpts)

metric	this model, viterbi	this model, argmax	binary-template baseline
root	0.729	0.724	0.541
majmin	0.689	0.675	0.516
triads	0.525	0.516	0.399
sevenths	0.575	0.566	0.400
mirex	0.721	0.721	0.506

Scored against the performed annotation. The previous 25-class version scored majmin 0.684 and sevenths 0.450, so the seventh vocabulary lifted sevenths by 12.5 points without costing majmin.

Usage

The model definition lives in the alberto repository (alberto/chords/model.py). With it on your path:

from huggingface_hub import hf_hub_download
from alberto.chords.model import recognize

weights = hf_hub_download("shawon/alberto-chord-crnn", "crnn_majmin.pt")
segments = recognize(y, sr, weights=weights)   # y: mono float32 numpy array
for seg in segments:
    print(f"{seg.start:6.2f} {seg.end:6.2f} {seg.chord}")

Or load the raw checkpoint into your own copy of the architecture:

import torch
ckpt = torch.load("crnn_majmin.pt", map_location="cpu", weights_only=True)
model = ChordCRNN(n_bins=ckpt["n_bins"], n_classes=ckpt.get("n_classes", 25)).eval()
model.load_state_dict(ckpt["state_dict"])
logits = model(torch.from_numpy(features).unsqueeze(0))   # (1, n_frames, 73)

Frames whose RMS is below -60 dBFS should be treated as no-chord regardless of the prediction; alberto decodes the posteriors with Viterbi (see transitions above).

Limitations

Six qualities only; sus, diminished and augmented chords map to the nearest class by their third, and extended chords (9ths, 6ths, 11ths) likewise.
Trained on one acoustic guitar recorded with one microphone. Expect lower accuracy on electric guitar, distortion, or full mixes.
Assumes the root is the lowest note, as in ordinary guitar voicings.

Downloads last month: 43