← Work

2021 · 2026

AutoTab v2

Guitar tablature from audio, via CNN

Keras 3 over CQT features, 6 strings × 21 classes. Modernised in 2026: Python 3.12, uv, CLI, Streamlit app, Docker, pytest, GitHub Actions. Calibrating the silence class raises the F-score roughly threefold — with no retraining.

Source →

Problem

Guitarists learn pieces from tablature, but most recordings don't come with any. Picking the frets out of audio by ear is tedious – and an interesting problem for a neural network, because six strings ring at once.

Solution

A convolutional neural network in Keras, trained on the GuitarSet dataset from Queen Mary University. Constant-Q spectrograms as input, six outputs with 21 classes each – a fret or silence per string. Built in the 2021 bootcamp, fully modernised in 2026: Python 3.12, Keras 3, uv, a CLI, a Streamlit interface, Docker, 18 tests and CI.

Result

The biggest improvement came not from more training but from calibrating the silence class: roughly three times the F-score with the same model. With Demucs the tool now separates the guitar out of full songs, and adapters for six datasets make further training possible.