TTS Radio - AI Voice Cloning System

5-Presenter AI Radio System using 748M parameter voice cloning models

NO FALLBACKS - This system uses AI models ONLY (no basic TTS)

---

🎯 Overview

This project creates a radio show with 5 different AI-generated presenters, each with unique personalities and voice characteristics. All voices are generated using state-of-the-art AI models:

• NeuTTS Air - 748M parameter Qwen2-based voice cloning model
• Bark - Advanced AI voice synthesis model

5 AI Presenters

---

⚡ Quick Start

1. Install Dependencies

pip install -r requirements.txt

2. Run Tests

# Verify no TTS fallbacks exist
python test_no_fallbacks.py

# Check NeuTTS Air setup
python test_neutts_setup.py

3. Generate Radio Show

# Recommended: NeuTTS Air (highest quality)
python neutts_voice_cloning.py

# Alternative: Simplified version
python simple_radio.py

# Alternative: Bark model
python voice_cloning_system.py

---

📋 Requirements

System Requirements

• Python: 3.8 or higher
• RAM: 8GB minimum (16GB recommended)
• Storage: 5GB for models

Required Python Packages

• PyTorch 2.0+
• Transformers 4.30+
• NumPy, SciPy
• soundfile
• requests, aiohttp

Optional Dependencies

• espeak - For NeuTTS phonemization (improves quality)

  # Windows
  winget install espeak
  # or
  choco install espeak
  

• Ollama - For AI-generated content (optional)

  • If not available, uses personality-based responses

---

🚀 Usage

Option 1: NeuTTS Air (Recommended)

Best quality voice cloning with 748M parameter AI model.

python neutts_voice_cloning.py

Features:

• Real voice cloning from 3-second samples
• 748M parameter Qwen2 architecture
• On-device processing (no cloud)
• Human-like prosody and timbre

Output: output/ai_voice_*.wav

Option 2: Simple Radio

Quick and easy radio generation.

python simple_radio.py

Features:

• Uses NeuTTS Air
• 5 presenters with different AI voices
• Ollama integration for AI responses
• Clean, simple code

Output: output/radio_*.wav

Option 3: Bark Voice Cloning

Alternative AI model for voice generation.

python voice_cloning_system.py

Features:

• Bark TTS voice cloning
• Multiple voice presets
• Fast generation
• No external dependencies

Output: output/ai_speech_*.wav

---

🧪 Testing

Test for TTS Fallbacks

Verifies that NO basic TTS libraries are used:

python test_no_fallbacks.py

Checks:

• ✅ No pyttsx3
• ✅ No gTTS
• ✅ No edge-tts
• ✅ Only AI models used
• ✅ requirements.txt clean

Test NeuTTS Setup

Verifies NeuTTS Air is properly configured:

python test_neutts_setup.py

Checks:

• ✅ NeuTTS Air imports
• ✅ Model files present
• ✅ Dependencies installed
• ✅ Sample audio files available

---

📁 Project Structure

tts-radio/
├── neutts_voice_cloning.py    # NeuTTS Air implementation (RECOMMENDED)
├── simple_radio.py             # Simplified NeuTTS Air version
├── working_radio.py            # Working radio with NeuTTS Air
├── voice_cloning_system.py    # Bark TTS implementation
├── generate_audio_demo.py      # Audio generation demo
├── ai_model_demo.py            # Model capabilities demo
├── test_neutts_setup.py        # Setup verification test
├── test_no_fallbacks.py        # TTS fallback detection test
├── requirements.txt            # Python dependencies
├── CHANGES.md                  # Changelog
├── README.md                   # This file
├── neutts-air/                 # NeuTTS Air model directory
│   └── samples/                # Voice reference samples
│       ├── dave.wav
│       └── jo.wav
├── output/                     # Generated audio files
├── presenters/                 # Presenter profiles
└── voice_samples/              # Voice samples

---

🎙️ How It Works

1. AI Content Generation

Each presenter gets AI-generated responses based on their personality:

# Example: Dave's personality
"Technical expert, analytical, concise"

# Generated response
"This is Dave. From a technical perspective, artificial intelligence 
represents a fascinating intersection of innovation and practical application."

Content Sources:

• Ollama AI (if available) - Dynamic AI responses
• Personality-based (fallback) - Pre-defined personality responses

2. AI Voice Cloning

Voice is generated using AI models:

NeuTTS Air Process:

1. Load 748M parameter model
2. Encode reference audio (3-second sample)
3. Generate speech with cloned voice
4. Save as high-quality WAV file

Bark Process:

1. Preload Bark models
2. Select voice preset (v2/en_speaker_X)
3. Generate audio with AI
4. Save as WAV file

3. Output

Audio files are saved to output/ directory:

• ai_voice_dave.wav
• ai_voice_jo.wav
• ai_voice_alex.wav
• ai_voice_sarah.wav
• ai_voice_mike.wav

---

⚙️ Configuration

Using Custom Voice Samples

For NeuTTS Air, add your own voice samples:

1. Record 3+ seconds of clear speech
2. Save as WAV file in neutts-air/samples/
3. Create matching text file with transcript
4. Update presenter configuration:

{
    "name": "CustomName",
    "ref_audio": "neutts-air/samples/your_voice.wav",
    "ref_text": "neutts-air/samples/your_voice.txt"
}

Using Ollama for AI Responses

1. Install Ollama: https://ollama.ai
2. Pull a model:

   ollama pull llama2
   

3. Run Ollama server:

   ollama serve
   

4. Scripts will automatically detect and use Ollama

---

🔍 Verification

Verify No TTS Fallbacks

python test_no_fallbacks.py

Expected Output:

[SUCCESS] All tests passed!
The codebase uses AI models only - NO TTS fallbacks!

Verify AI Models Work

python test_neutts_setup.py

Expected Output:

[OK] NeuTTS Air 748M parameter AI model is ready
[OK] All Python dependencies are installed
SUCCESS: NeuTTS Air is properly set up!

---

🐛 Troubleshooting

Issue: "ModuleNotFoundError: No module named 'neuttsair'"

Solution:

# Ensure neutts-air directory exists
# Install dependencies
pip install torch transformers

Issue: "espeak not found"

Solution:

# Windows
winget install espeak

# Or download from:
# https://espeak.sourceforge.net/download.html

Issue: "Ollama connection failed"

This is OK! The system will use personality-based responses instead.

To enable Ollama:

ollama pull llama2
ollama serve

Issue: "CUDA out of memory"

Solution: Use CPU instead (already default):

NeuTTSAir(
    backbone_device="cpu",  # Use CPU
    codec_device="cpu"
)

---

📊 Performance

NeuTTS Air

• Quality: ⭐⭐⭐⭐⭐ (Highest)
• Speed: ~10-30 seconds per segment (CPU)
• Memory: ~4GB RAM
• Voice Quality: Near-human

Bark

• Quality: ⭐⭐⭐⭐☆
• Speed: ~5-15 seconds per segment (CPU)
• Memory: ~2GB RAM
• Voice Quality: Very good

---

🚫 What This System Does NOT Use

• ❌ pyttsx3 (basic TTS)
• ❌ gTTS (Google TTS)
• ❌ edge-tts (Microsoft Edge TTS)
• ❌ Any cloud-based TTS services
• ❌ Any non-AI fallback methods

This system uses AI models ONLY!

---

📝 License

This project uses:

• NeuTTS Air - Check neuphonic licensing
• Bark - MIT License
• PyTorch - BSD License

---

🤝 Contributing

To contribute:

1. Ensure no TTS fallbacks are added
2. Run python test_no_fallbacks.py before committing
3. All speech generation must use AI models
4. Update tests if adding new features

---

📚 Additional Resources

• NeuTTS Air Documentation
• Bark TTS
• PyTorch
• Ollama

---

✅ Status

• Audited: ✅ All files checked
• Tested: ✅ Comprehensive test suite
• No Fallbacks: ✅ 100% AI models
• Working: ✅ Fully functional

Last Updated: 2025-10-05