Expodecaprio/OmniVoice-Studio

OmniVoice Studio 🎙️

An enhanced, streamlined, and modular version of OmniVoice, featuring persistent .pt voice prompt serialization, clean English-only UI, custom offline caching, and optimized generation controls.

---

🌟 Interactive Demo & Model Access

🎙️ OmniVoice Studio (Custom GitHub Repo)

OmniVoice Studio: State-of-the-Art Zero-Shot Voice Cloning Explore our streamlined open-source repository on GitHub or run the interactive live Space on Hugging Face. ⭐ View Custom Repo on GitHub  |  🚀 Live Interactive Space  |  📦 Pre-trained Weights: k2-fsa/OmniVoice

---

⚖️ CRITICAL ETHICAL USE & LEGAL DISCLAIMER

🛑 STRICT PROHIBITION ON UNAUTHORIZED VOICE CLONING

Do NOT use this software, model weights, or generated audio for any purpose involving cloning the voice of any individual without their explicit, verifiable, prior written consent.

By downloading, cloning, or operating this project, you strictly agree to the following terms:

1. No Unauthorized Cloning: You will not impersonate any person (living or deceased), public figure, celebrity, or private individual without valid legal consent.
2. No Deceptive Content: You will not generate defamatory, misleading, fraudulent, non-consensual sexual, or politically manipulative speech (Deepfakes).
3. Non-Commercial Restriction: The pre-trained neural network weights (k2-fsa/OmniVoice) are licensed strictly under Creative Commons Attribution-NonCommercial 4.0 International (CC-BY-NC 4.0). Commercial deployment, monetization, or paid API integration of the model weights is forbidden without explicit authorization from the original copyright holders.

The creators and maintainers of this repository assume no liability for misuse, illegal activities, or copyright/personality rights violations arising from the operation of this software.

---

🏆 Attribution & Acknowledgements

This project is built upon the foundational work of OmniVoice by the Xiaomi AI Lab Next-gen Kaldi Team. We give our utmost gratitude and full credit to the original authors and maintainers:

• Original Open-Source Project: k2-fsa/OmniVoice on GitHub
• Original Authors & Research Team: Han Zhu (zhuhan), Next-gen Kaldi team (k2-fsa), and contributors.
• Original License: Apache License 2.0 (Codebase) & CC-BY-NC 4.0 (Model Checkpoints).

All core neural network architectures, flow-matching mechanisms, and multi-lingual acoustic modeling belong to their pioneering research.

---

✨ Key Enhancements & Features

This repository introduces several critical engineering upgrades to make local deployment cleaner, faster, and more reusable:

1. 💾 Voice Prompt Serialization (VoiceClonePrompt.save & .load)

Instead of re-extracting speaker embeddings and processing reference audio on every single synthesis request, this enhanced version allows:

• Saving (.pt): Extract the acoustic prompt (VoiceClonePrompt) once from any reference audio and save it to a lightweight .pt file.
• Instant Loading: Directly upload or select a pre-saved .pt prompt file in the UI to skip reference audio processing entirely.
• Automatic Export: Every time you generate speech, the extracted prompt is automatically exported as a downloadable artifact in saved_prompts/.

2. 🎨 Clean, Streamlined English-Only UI

• Removed redundant dual-language Chinese/English clutter from labels and buttons for a clean, modern aesthetic.
• Removed the experimental "Voice Design" tab to focus 100% on High-Fidelity Voice Cloning.
• Optimized Defaults: Default generation speed is tuned to 0.9 for more natural, deliberate pacing and clearer articulation across complex sentences.

3. 🛡️ Offline Caching & Storage Management

• Automatic redirect of Hugging Face cache (HF_HOME, HF_HUB_CACHE) and PyTorch cache (TORCH_HOME) to a local models_cache/ directory when running locally on Windows/Linux.
• Prevents C:\ drive exhaustion when downloading the ~3.36 GB model weights.
• Pre-enabled HF_HUB_OFFLINE=1 once weights are downloaded locally to eliminate startup latency and network timeouts.

---

🛠️ Installation & Setup

Prerequisites

• Windows 10/11 or Linux with NVIDIA GPU (CUDA 11.8+ recommended).
• Python 3.9, 3.10, or 3.11.

1. Clone the Repository

git clone https://github.com/Fazzbro/OmniVoice-Studio.git
cd OmniVoice-Studio

2. Create & Activate Virtual Environment

python -m venv .venv
# Windows PowerShell:
.\.venv\Scripts\Activate.ps1
# Linux / macOS:
source .venv/bin/activate

3. Install Dependencies

pip install --upgrade pip
pip install -r requirements.txt

---

🚀 Running the Studio locally

Simply launch the entry point script:

python "Omni Voice.py"

1. On the first launch, the script will automatically download the k2-fsa/OmniVoice checkpoint (~3.36 GB) and store it safely in the models_cache/ folder.
2. Once loaded, open your web browser to the Gradio interface:

   http://127.0.0.1:7860
   

---

📂 Project Structure

OmniVoice-Studio/
├── app.py                 # Hugging Face Spaces entry point
├── Omni Voice.py          # Local runner with D:\ drive offline cache config & Gradio server
├── requirements.txt       # Clean, exact Python dependencies
├── LICENSE                # Apache 2.0 License with Xiaomi AI Lab attribution
├── README.md              # Project documentation, Spaces YAML, and legal disclaimer
└── omnivoice/             # Enhanced bundled package (self-contained)
    ├── __init__.py        # Exports OmniVoice, VoiceClonePrompt, generation configs
    ├── models/
    │   └── omnivoice.py   # Core model + VoiceClonePrompt serialization (.save/.load)
    ├── cli/
    │   └── demo.py        # Streamlined English-only Gradio UI (0.9 speed default)
    └── ...                # Core utility and generation modules

---

📜 License Summary

• Source Code: Distributed under the Apache License 2.0. You are free to modify and redistribute the code provided proper credit and copyright notices are maintained.
• Pre-trained Model Weights (k2-fsa/OmniVoice): Distributed under CC-BY-NC 4.0. Non-commercial use only.