Emulation of Guitar Effects Using AI

Overview

This project aims to emulate guitar effects using deep learning, leveraging the WaveNet model to process audio data and generate guitar effects in real time. By training a neural network to learn the transformations applied by various guitar effects, this project seeks to recreate the nuanced sound modifications applied by traditional guitar pedals and audio processors.

Methodology and Key Technologies

TCN-Based Real-Time Guitar Effect Emulation

This repository contains the implementation and supplementary materials for a Temporal Convolutional Network (TCN)-based model designed to emulate guitar effects in near-real-time. By receiving a clean guitar input signal, the model generates output signals that closely approximate audio produced by professional hardware effect units. This approach integrates advanced preprocessing techniques and a parameter-efficient network architecture, enabling sub-millisecond inference rates and high-fidelity output quality.

Key Contributions

• Near-Real-Time Inference:
  Achieves average inference times in the range of 1–3 ms, thereby facilitating live performance and recording scenarios with minimal latency.

• Efficient TCN Architecture:
  Employs a TCN optimized for time-series audio analysis. Compared to conventional frameworks such as WaveNet, the TCN model is more parameter-efficient and amenable to parallelization, thereby reducing computational overhead and improving inference speed without compromising audio quality.

• Enhanced Data Preprocessing:
  Incorporates RMS normalization, cross-correlation-based synchronization, STFT-based spectral transformation, and overlapped segmentation. These steps collectively ensure stable input-output alignment and improved frequency-domain representation, contributing to superior signal-to-noise ratio (SNR) and more accurate replication of the target effects.

• Multi-Modal Loss Integration:
  Utilizes a combination of MSE, STFT, and perceptual (Mel-spectrogram) loss functions to account for both time- and frequency-domain characteristics. This strategy ensures that the model does not merely minimize numerical error but also enhances perceptual sound quality.

• Comprehensive Performance Evaluation:
  Evaluates the model using SNR, BFRE (Batch-Level Frequency Response Error), BTAS (Batch-Level Temporal Alignment Score), and PATS (Perceptual A/B Test Score). The proposed solution improves SNR from -3.58 dB in initial configurations to approximately +2.10 dB. In perceptual testing, the model achieves an ~82% preference rating relative to authentic hardware-processed audio signals.

Repository Structure

• dataset/: Placeholder directory for dataset handling scripts and instructions (no raw datasets included).
• docs/: Additional documentation, figures, and references detailing architectural design and experimental setups.
• README.md: Contains an overview of the project, including high-level methodology, performance indicators, and references.

Requirements and Setup

Prerequisites:

• Python 3.8+
• PyTorch >= 1.7
• torchaudio, numpy, scipy, librosa
• A GPU is recommended for both training and real-time inference.

Installation:

git clone https://github.com/YourUsername/TCN-Guitar-Effect-Emulation.git
cd TCN-Guitar-Effect-Emulation
pip install -r requirements.txt