y10ab1/guitar_effect_removal

Demo page and evaluation code for "Distortion Recovery: A Two-Stage Method for Guitar Effect Removal"

Guitar Effect Removal Evaluation

This repository is for reproducibility of the experiments conducted on the public dataset EGDB, as presented on our demo page. Our work, titled "Distortion Recovery: A Two-Stage Method for Guitar Effect Removal", was published at DAFx 2024. The paper introduces a novel two-stage method to recover clean guitar signals from distorted wet guitar signals. This research was conducted in collaboration with Positive Grid. The paper is now available on arXiv: https://arxiv.org/abs/2407.16639.

This repository contains code for evaluating the recovery of clean guitar tones from wet guitar tones using various metrics. The script calculates metrics such as SISDR, ESR, and MRSTFTLoss, and also computes the Frechet Audio Distance (FAD) score.

Installation

To install pytorch, run the following command:

pip install torch==2.1.0 torchvision==0.16.0 torchaudio==2.1.0 --index-url https://download.pytorch.org/whl/cu118

To install the other required packages, run the following command:

pip install -r requirements.txt

Data

To evaluate the additional experiments on the EGDB dataset, download the dry siganl and recovered results from this Google Drive.

Usage

To evaluate the recovery of clean guitar tones from wet guitar tones, run the following command:

python test.py --dry_folder /path/to/dry/folder --recovered_folder /path/to/recovered/folder