Guitab

Guitab turns a recording of a solo guitar into a tab you can open in Guitar Pro, TuxGuitar, MuseScore or any online gp5 player. It is made for steel-string fingerstyle playing, works on any recording you can play (a phone video is fine), and runs on a laptop: a 4-minute song takes well under a minute.

You get a .gp5 file with the notes on the right strings and frets, laid out on a grid that follows the beats of the recording, with an optional capo. A local web app lets you record or load a song, transcribe it and play the tab back, and a browser version does the same with nothing to install.

What to expect. On clean recordings about 85% of the notes come out right, with the correct string and fret. Dense 3–4 note chords, quiet inner voices and phone recordings in a reverberant room are much harder. Treat the result as a first draft to correct by ear, not as a finished tab.

What's new in 1.2 (2026-10-31): optional chord names above the tab (--chords, or the Chord names option in the local app and the browser), from a small chord model (0.5 MB) that reads the transcribed notes: about 3 in 4 shown chords are right on real fingerstyle songs and 9 in 10 on strummed accompaniment, and chords the model is unsure of are left out. A Chords switch over the tab shows or hides them. All versions are in the changelog.

Results

Test set	Metric (onset ±50 ms)	E15 (public data)	E19 (+ custom dataset)	E25 (default)	E24 (tiny)
GuitarSet test (player 05, steel string)	string + fret F1 / pitch F1	0.836 / 0.866	0.843 / 0.866	0.838 / 0.861	0.833 / 0.856
GuitarSet test, phone/room-degraded copy	string + fret F1 / pitch F1	0.650 / 0.735	0.640 / 0.721	0.685 / 0.757	0.655 / 0.737
GAPS test (classical guitar, unseen players and rooms)	pitch F1	0.827	0.813	0.830	0.825
Real fingerstyle recordings (19 held-out songs of the custom dataset)	string + fret F1	0.548	0.617	0.612	0.601
EGDB electric guitar test (25 clips, clean DI, never seen in training)	string + fret F1 / pitch F1	0.636 / 0.833	0.712 / 0.850	0.721 / 0.849	0.691 / 0.830
EGDB electric guitar test, through two amps (clean and crunch)	string + fret F1 / pitch F1	0.613 / 0.784	0.631 / 0.787	0.652 / 0.788	0.622 / 0.781

• E25 (default) is a 6M-parameter model, a quarter of E19's size, trained to imitate the average of E19 and the 66M-parameter E18 (distillation). It is within a point of E19 on clean steel-string audio, clearly better on phone recordings and classical guitar, and 1.5x faster.
• E24 (tiny) follows the same recipe at half the size: 2.9M parameters, an 11 MB file, 2.3x faster than E19 and about one point behind it on GuitarSet.
• E19 is the strongest model trained on the labels alone. Besides the public datasets it used a custom dataset of real fingerstyle recordings with aligned tabs, which is not distributed.
• E15 is trained on public datasets only, and anyone can reproduce it with the commands in section 4.

Electric guitar. The models never saw an electric guitar during training, but they also work on clean or lightly overdriven ones. On the EGDB test set (Chen et al., ICASSP 2022) the notes are found almost as well as on acoustic recordings, while the string choice is less reliable, so expect more fingering corrections. Heavy distortion was not tested.

The whole research history, with every experiment, setting and result (including the approaches that did not work), is in EXPERIMENTS.md.

1. Install

From the repository root, with Python 3.11, on macOS, Linux or Windows:

uv venv .venv --python 3.11
uv pip install --python .venv/bin/python -r requirements.txt   # exact tested versions
uv pip install --python .venv/bin/python -e . --no-deps          # the `guitab` command

pip install -e ".[train]" works too; requirements.txt pins the environment the results were produced with. On Linux the default PyTorch wheel includes CUDA; for another CUDA version or a CPU-only build, install torch first as described on pytorch.org. On Windows, the commands below use .venv\Scripts\ instead of .venv/bin/.

Every command takes --device auto|cuda|mps|cpu. The default, auto, uses an NVIDIA GPU if present, else the Apple GPU, else the CPU.

ffmpeg is optional. WAV, FLAC, OGG, MP3 and AIFF files are read without it. The guitab command needs ffmpeg only for other formats, such as videos or M4A files (brew install ffmpeg on macOS, apt install ffmpeg on Debian or Ubuntu). The web app decodes those in the browser and needs no ffmpeg. Training needs ffmpeg for its audio augmentation and the GAPS download.

Model weights

The weights are on Hugging Face at danmor/guitab:

File	Model	Size
guitab_e25.safetensors	E25, the default	24 MB
guitab_e24.safetensors	E24, the tiny model	11 MB
guitab_e19.safetensors	E19	91 MB
tablm.safetensors	the tab language model behind --music_lm	13 MB
beat_this_small0_int8.onnx	the Beat This! beat tracker (small model) with int8 weights, shared with the browser version	3.6 MB
beat_this_final0_int8.onnx	the larger Beat This! model, used only if you select it (below)	21 MB
chord_model.safetensors	the chord model behind the chord names (EXPERIMENTS.md, C1)	0.5 MB

여기까지가 README 앞부분입니다. 나머지는 원문에서 읽을 수 있습니다. README 전문 보기 ↗