🎙️ NeuTTS Air - Voice Cloning TTS

A Gradio-based web interface for NeuTTS Air, enabling instant voice cloning and natural-sounding text-to-speech synthesis.

Features

• Instant Voice Cloning: Clone any voice with just 3-15 seconds of reference audio
• Auto-Transcription: Automatic transcription of reference audio using Whisper
• Multiple Model Options: Choose between Q4 (faster) or Q8 (higher quality) models
• Long Text Support: Automatic text chunking with smooth crossfading for longer passages
• Seed Control: Reproducible generation with optional seed parameter
• User-Friendly Interface: Clean Gradio web UI with real-time status updates

Requirements

• Python 3.8+
• CUDA-capable GPU (recommended for best performance)
• espeak-ng (for phonemization)

Installation

1. Install espeak-ng

Windows:

# Using Chocolatey
choco install espeak

# Using Scoop
scoop install espeak

macOS:

brew install espeak-ng

Linux:

sudo apt-get install espeak-ng

2. Install Python Dependencies

pip install -r requirements.txt

Usage

Starting the Application

python app.py

The Gradio interface will launch in your browser at http://localhost:7860

Using the Interface

1. Model Setup (first time only):

   • Select a model (Q4 for faster, Q8 for higher quality)
   • Click "Download" to download the model
   • Click "Initialize" to load the model into memory

2. Generate Speech:

   • Enter the text you want to synthesize
   • Upload a reference audio file (3-15 seconds, clean audio recommended)
   • The reference text will auto-transcribe, or you can provide it manually
   • (Optional) Set a seed for reproducible results
   • Click "Generate Speech"

3. Model Management:

   • Use "Unload" to free up memory and switch between models

Model Options

• Q4 (Faster, Lower Quality): neuphonic/neutts-air-q4-gguf

  • Faster inference
  • Lower memory usage
  • Good for testing and quick iterations

• Q8 (Slower, Higher Quality): neuphonic/neutts-air-q8-gguf

  • Higher quality output
  • More memory intensive
  • Best for final production

Tips for Best Results

• Reference Audio: Use 3-15 seconds of clean audio with minimal background noise
• Reference Text: Should match the audio exactly (or leave empty for auto-transcription)
• Text Length: The system automatically handles long texts by chunking and crossfading
• Seed: Use the same seed for consistent results across generations

Project Structure

.
├── app.py                 # Main Gradio application
├── neuttsair/
│   ├── __init__.py
│   └── neutts.py         # Core NeuTTS Air implementation
├── requirements.txt       # Python dependencies
└── README.md             # This file

Dependencies

• gradio - Web interface
• torch - Deep learning framework
• openai-whisper - Audio transcription
• soundfile - Audio I/O
• librosa - Audio processing
• neucodec - Neural audio codec
• phonemizer - Text to phoneme conversion
• llama-cpp-python - GGUF model support
• transformers - HuggingFace models
• resemble-perth - Audio processing utilities

License

See LICENSE file for details.

Acknowledgments

• Built on NeuTTS Air by Neuphonic
• Uses Whisper for transcription
• Powered by Gradio for the web interface