Guitar String Estimation on GuitarSet via YourMT3

Reproducible code for the DAGA 2026 paper: "Bestimmung gespielter Gitarrensaiten anhand von Audiofeatures basierend auf multipler Grundfrequenzschätzung" ("Estimation of Played Guitar Strings Using Audio Features Based on Multiple Fundamental Frequency Estimation") — Simon Büchner, Paul A. Bereuter, Alois Sontacchi · IEM Graz

Description

This project builds on YourMT3 to implement a machine learning pipeline for guitar string classification using the GuitarSet dataset. Given a guitar audio file, the pipeline detects notes via multi-pitch estimation and classifies which string each note was played on.

If you have an idea for an improvement, found a bug, or just want to talk about the implementation, feel free to open an issue!

---

Prerequisites

• Python 3.11
• Git LFS
• GPU recommended for YourMT3 inference (not strictly required)

---

Setup

1. Install Git LFS

git lfs install

2. Clone the repository

git clone https://github.com/SimonBuechner/GuitarStringEstimation.git
cd GuitarStringEstimation

If the checkpoint download fails due to Git LFS bandwidth limits, retrieve it from the original YourMT3 repository and place it at amt/logs/2024/ptf_all_cross_rebal5_mirst_xk2_edr005_attend_c_full_plus_b100/checkpoints/model.ckpt. A migration to Hugging Face is planned.

3. Create a Python environment

Using venv:

python3.11 -m venv .venv
source .venv/bin/activate  # Linux / macOS
.venv\Scripts\activate     # Windows

Install dependencies:

pip install -r requirements.txt

(Alternatively, conda works fine too.)

4. Download the dataset

Download the GuitarSet dataset and place it at: data/GuitarSet/

---

Project Structure

GuitarStringEstimation/
├── .venv/
├── amt/                      # YourMT3 code and checkpoints
│   ├── content/
│   ├── logs/2024/ptf_all_cross_rebal5_mirst_xk2_edr005_attend_c_full_plus_b100/
│   │   └── checkpoints/model.ckpt
│   └── src/
│       ├── model/
│       ├── utils/            # Utilities from YourMT3
│       └── ...
├── gse/                      # Guitar string estimation code
│   └── src/
│       ├── beta_distributions.py
│       ├── YMT3_inference.py # Calls YourMT3 for note detection
│       ├── svm_full.joblib   # Pretrained SVM for string classification
│       └── ...
├── data/
│   └── GuitarSet/
│       ├── annotation/*.jams
│       ├── audio_mono-mic/
│       └── ...
├── scripts/                  # Miscellaneous experiment scripts (see below)
├── .gitignore
├── .gitattributes
├── LICENSE
└── README.md

---

Usage

Run the following steps in order to reproduce the results from the paper.

1. Extract tracks and notes

python extract_GuitarSet.py

Creates a noteData/ directory containing serialized (pickle) note data from GuitarSet.

2. Run YourMT3 inference

python YMT3_inference.py

Runs the YourMT3 model on train and test subsets, extracts predicted notes, and matches them with ground truth annotations.

⚠️ All subsequent steps operate only on these matched notes.

3. Train the SVM classifier

python train_classifier.py

Extracts features from the note data and trains an SVM string classifier. This step is computationally expensive but only needs to be run once.

A pretrained model (svm_full.joblib) is already provided so you can skip this step for evaluation.

4. Evaluate the classifier

python evaluate_classifier.py

Runs string classification and evaluation. Includes options to reproduce the permutation importance experiments from the paper.

5. Evaluate inharmonicity coefficient estimation

python beta_distributions.py

Collects inharmonicity coefficients (β) for all strings, normalizes them to open-string equivalents, and generates plots for selected subsets.

---

Scripts

The scripts/ directory contains miscellaneous scripts used for additional experiments.

⚠️ These are not guaranteed to work out of the box. They were primarily used to evaluate other state-of-the-art approaches (e.g., Hjerrild et al., 2019) and are provided for reference only.

---

Notes & Tips

• Feature computation can take a significant amount of time and uses multiprocessing.
• Ensure the GuitarSet directory structure matches what is shown above; mismatches will cause script failures.
• GPU is recommended for YourMT3 inference but not strictly required.

---

Citation

If you use this code, please cite:

@inproceedings{buechner2026guitarstring,
  title     = {Bestimmung gespielter Gitarrensaiten anhand von Audiofeatures basierend auf multipler Grundfrequenzsch{\"a}tzung},
  author    = {B{\"u}chner, Simon and Bereuter, Paul A. and Sontacchi, Alois},
  year      = {2026},
}

---

Based On

• YourMT3 — Multi-pitch music transcription transformer
• GuitarSet — Guitar recordings with detailed annotations