AI-Generated Voice Cloning Using GANs

This repository contains the Mini Project Report (MPR) notebook for "AI-Generated Voice Cloning Using GANs", focused on developing a system that synthesizes human-like voices using Generative Adversarial Networks (GANs). The project utilizes the Mozilla Common Voice dataset to train and evaluate the generative model on diverse speech samples.

🎯 Project Objective

To build an AI system capable of cloning human voices by training a GAN architecture on real-world multilingual voice data. The aim is to replicate the natural tone, pitch, and speaking style of a given speaker through synthesized speech samples.

🧠 Key Components

• Generative Adversarial Network (GAN) Architecture:

  • Custom generator and discriminator models tailored for raw audio signal generation.
  • Feature extraction techniques applied before feeding audio into GAN.

• Voice Preprocessing & Feature Engineering:

  • Audio normalization, silence trimming, and spectrogram generation.
  • Conversion to Mel spectrograms for stable GAN training.

• Training Loop:

  • Balanced generator-discriminator training cycle.
  • Loss functions customized for speech signal characteristics.

• Voice Cloning Evaluation:

  • Comparison of real vs. synthesized audio using waveform visualization and audio output.
  • Metric evaluation (e.g., Spectral Convergence, Signal-to-Noise Ratio).

📦 Dataset

This project uses the Mozilla Common Voice Dataset:

• Open-source, multilingual dataset of speech samples.
• Provides thousands of validated clips across multiple speakers, languages, and accents.
• Used for both training and evaluation phases.

🗂️ File Structure

voice-cloning-gan/
├── AAI MPR.ipynb      # Main project notebook
└── README.md          # Project documentation

⚙️ Setup Instructions

1. Clone the repository:

   git clone https://github.com/yourusername/voice-cloning-gan.git
   cd voice-cloning-gan

2. Install dependencies:

   pip install -r requirements.txt

3. Download dataset:

   • Visit Common Voice and download your preferred language version.
   • Extract and place the audio files inside a data/ directory (if using externally).

4. Run the notebook:

   jupyter notebook "AAI MPR.ipynb"

🧪 Results Preview

• Synthesized voice samples generated after each training epoch.
• Visual comparison between input voice spectrograms and generated outputs.
• Evaluation through waveform plots and perceptual listening.

📌 Status

✅ Core GAN architecture implemented
🚧 Currently testing across multiple speakers and languages
📊 Future improvements include attention layers and multi-speaker conditioning