This repository provides tools for training models for MMVC (RealTime-Many to Many Voice Conversion), an AI-powered real-time voice changer. It targets users who want to create custom voice models for voice conversion, enabling them to transform their voice into that of various characters or individuals.

How It Works

MMVC_Trainer leverages Google Colaboratory for accessible model training, abstracting away local environment dependencies. The core process involves preparing audio datasets (recorded speech and target voice samples) and corresponding text transcriptions. Users then utilize provided Jupyter notebooks to configure training parameters, initiate the training process using pre-trained models, and validate the resulting voice conversion models.

Quick Start & Requirements

• Install: Install by running the provided Colab notebooks.
• Prerequisites: Google Account, audio data (24000Hz, 16bit, 1ch), and text transcriptions. Pre-trained models are available via Hugging Face.
• Setup: Follow the step-by-step tutorials within the Colab notebooks.
• Links:
  • Colab Installation: https://github.com/isletennos/MMVC_Trainer
  • MMVC Client: https://github.com/isletennos/MMVC_Client
  • Community Discord: https://discord.gg/2MGysH3QpD

Highlighted Details

• Enables real-time voice conversion.
• Supports training custom voice models from user-provided audio.
• Utilizes Google Colab for accessible training environments.
• Includes tutorials for various voice conversion scenarios, such as transforming into "Zundamon."

Maintenance & Community

The project is active with a Discord community for support and discussion. The developer can be contacted via Pixiv Fanbox.

Licensing & Compatibility

Released under the MIT license, allowing for free use, distribution, modification, and commercial use. However, specific voice data used (e.g., Zundamon, Kyushu Sora) may have their own terms of use that must be adhered to. When using specific character voices, attribution is recommended.

Limitations & Caveats

The README mentions that Mac (Intel) performance can be slow, and while CPU operation is possible on recent hardware, GPU acceleration is generally preferred for optimal performance. Users must ensure their audio data meets the specified format (24000Hz, 16bit, 1ch).