We provide listening examples of our paper “Distortion Recovery: A Two-Stage Method for Guitar Effect Removal”, published at DAFx 2024.

Abstract

Removing audio effects from electric guitar tracks significantly increases the flexibility in post-production, providing a more adaptable environment for sound editing. An audio distortion recovery model not only improves the guitar’s clarity but also opens up new opportunities for creative adjustments in mixing and mastering. Previous efforts to tackle this challenge have largely focused on synthetic distortions that are too simplistic to accurately capture the complexities encountered in real-world situations. Moreover, the performance of existing methods leave room for improvement.

In this paper we introduces a novel two-stage methodology to efficiently eliminate audio effects from electric guitar tracks, initially purifying the audio signal in the Mel-spectrogram domain, and subsequently employing a neural vocoder to reconstruct the pristine original guitar sound from the processed Mel-spectrogram. Our evaluation utilizes a comprehensive combination of VST plugins to assess the effectiveness of our proposed methodology. Experimental results demonstrate that our approach outperforms existing ones, exhibiting superior performance through both subjective and objective metrics.

Contents:

Notes:

Comparison of different models

Wet Dry Ours HifiGAN denoiser Demucs V3 DCUnet

Additional experiments on EGDB

We randomly rendered dry (direct input) guitar tracks from the EGDB dataset using BIAS FX2 and fine-tuned both our model and Demucs V3 on this rendered dataset.

The EGDB dataset consists of 240 tracks (~1.5 hours) and was randomly split by tracks into training, validation, and test sets, following an 8/1/1 split. Our model was then compared with Demucs V3. During both the training and testing phases, each track was randomly segmented into 4-second clips.

The segmented clips for testing, including wet/dry signal, results of our model and Demucs V3, are available here. The test code for evaluating the following metrics is available at this GitHub repository.

Objective evaluation metrics

Model FAD ↓ ESR ↓ SISDR ↑ MR-STFT ↓
Demucs V3 0.545 1.033 6.478 1.931
Ours 0.270 1.574 30.241 1.493

Audio samples

Wet Dry Ours Demucs V3