How Binturong R.E.D turns a guitar take into tab

Here’s what Binturong R.E.D does today: give it a recording of a guitar, and it writes back a score, with notation above tab, in bars, with the bends and hammer-ons marked. The image above is its first pass on a blues jam I recorded, untouched.

This post walks through how it gets there, at the level of “what problem does each step solve”, without the math.

The pipeline

Binturong R.E.D works in stages. Each stage takes what the one before it found and adds one kind of understanding:

Hearing the notes. The first step listens for pitches, and it has to hear several at once, because guitarists play chords, double-stops, and notes that ring into each other. This step uses basic-pitch, an open-source neural network from Spotify’s audio research team, which is good at exactly this. Everything after it is Binturong R.E.D’s own.

Finding picks and strums. A neural network hears pitch well, but it’s less sure when a note was struck. Binturong R.E.D looks separately for the sharp burst of energy each pick makes, and uses it to split notes cleanly and to group the notes of a strum into one chord.

Recognizing techniques. A bend looks like a note sliding up in pitch. Naively, that becomes two notes (or one wrong one). Binturong R.E.D follows each note’s pitch over time to tell a bend from a hammer-on, and to spot vibrato by how evenly the pitch swings.

Choosing string and fret. This is the step that makes guitar transcription harder than piano. The same note can often be played in four or five places on the neck. Binturong R.E.D follows where your hand probably is, the way a real player moves: staying in position is cheap, and shifting costs more the farther the hand moves.

Finding the rhythm. Notes arrive as times in seconds; a score needs eighth notes and triplets in bars. Binturong R.E.D finds the tempo from your strumming, then decides where every bar starts by weighing the whole take at once. An earlier version decided bar by bar, and one misplaced note could shift every bar after it.

Writing it down. Everything lands in one score, Binturong R.E.D’s own format, and small exporters turn that into files other tools open. Fixing a wrong note happens in MuseScore or Guitar Pro, which you already know how to use.

What it sees

Here’s a short Drop D riff in triplets. Below is a spectrogram of the recording, a picture of sound where height is pitch and darkness is loudness. The orange lines are the notes Binturong R.E.D heard; the green ones are what I actually played.

And here’s the score it wrote from the same recording, triplets and all:

Measured, every time

Transcription is easy to fool yourself about. A change that fixes the clip you’re staring at can quietly break ten others. So no change is kept on how it looks; it’s kept on what it measures:

The test recordings come from three places:

• GuitarSet, 360 recordings of acoustic guitar from the music research community, each annotated with exactly what was played.
• Guitar-TECHS, electric guitar played by professional players, recorded with a pickup that knows which string every note was on.
• My own reference recordings: riffs, chord changes, bends, and triplets, each with a hand-checked answer key.

Two recent results, so you have a sense of where things stand. On my reference recordings, Binturong R.E.D finds the right notes about 87% of the time, measured as an F-score, which counts both missed notes and extra ones. And the change that weighs the whole take before placing bars raised rhythm accuracy on those same recordings from 91% to 97%.

Those are my own recordings, so they flatter the engine a little. That’s why the public datasets matter, and why I’ll keep reporting both.

What’s still hard

• Heavy distortion and dense chords blur the notes together.
• Tab is a guess whenever the same note could be played in several places. Binturong R.E.D’s guesses are playable, but not always what you did.
• Fast playing packs notes closer together than the engine can always pull apart.

Next, I’m working on finding where bars start more reliably, using the strongest beats in the playing. I’ll write that up when it lands.

GuitarSet is by Q. Xi, R. Bittner, J. Pauwels, X. Ye, and J. P. Bello (ISMIR 2018). Guitar-TECHS is by H. Pedroza, W. Abreu, R. M. Corey, and I. R. Roman (ICASSP 2025). Both are licensed CC BY 4.0.