Zeyta AI Web Playground

A fully self-contained, locally-hosted web application providing comprehensive AI capabilities including Text-to-Speech, Speech-to-Text, Large Language Model chat, Autonomous Agent Mode, and a Visual Pipeline Builder.

✨ Features

🗣️ Text-to-Speech (TTS)

• High-quality speech synthesis using ChatterboxTTS
• RTX 50-Series Optimization: FP16 precision, CUDA Graphs, and embedding caching for lightning-fast generation
• Voice cloning with reference audio support
• Adjustable parameters (temperature, exaggeration, CFG weight)

🎤 Speech-to-Text (STT)

• Accurate transcription using Faster-Whisper
• Support for audio file upload
• Live microphone recording and transcription
• Multiple language support
• Configurable model sizes (tiny to large-v3-turbo)

💬 AI Chat & Web Search

• Context-aware conversations powered by Ollama
• Web Search Plugin: Real-time internet browsing via DuckDuckGo to answer current events
• Calculator Plugin: Solves math problems accurately
• Vision Plugin: Analyzes screenshots and images
• Chat history management

🤖 Autonomous Agent Mode

• Screen Analysis: The AI can "see" your screen and understand UI elements
• Auto-Execution: Performs multi-step tasks (clicking, typing, navigating) autonomously
• Task Planning: Breaks down complex goals into actionable steps
• Permission System: Granular control over what the agent can do (mouse, keyboard, file system)

🧩 Visual Pipeline Builder

• Drag-and-drop interface to create custom AI workflows
• Connect blocks: Mic → STT → LLM → TTS → Audio Output
• Mobile-Friendly: Responsive design that adapts layout (vertical on PC, horizontal on mobile)
• Real-time execution logging

⚙️ System Management

• Smart Self-Healing Setup: Automatically detects missing environments, creates virtual environments, and resolves dependency conflicts (e.g., fixing broken pip installs).
• System Integrity Checks: Verifies critical folders and files on every startup.
• Real-time system status monitoring (CPU, RAM, GPU VRAM)
• GPU/CPU detection with automatic fallback
• Model configuration display

🚀 Quick Start

⚠️ Prerequisite: Download Models

Cloning the repository is not enough! You must ensure the AI models are available:

1. Install Ollama: Download from ollama.com.
2. Pull a Chat Model: Run ollama pull llama3.1 (or mistral, gemma, etc.) in your terminal (or anything that can run on your system).
3. First Run Downloads: The app will download TTS and STT models on the first launch automatically. Ensure you have a stable internet connection.

Option 1: Simple Run (Recommended)

# Run the application as it will install dependencies automatically (may not be enough - Double check!)
.\start.ps1
# OR
.\start.bat

Option 2: Manual Installation

# Install dependencies first
pip install -r requirements.txt

# Run the application
python web_app.py

Then open your browser to: https://localhost:5000

Note: You will see a "Not Secure" warning because the app uses a self-signed certificate. This is required for microphone access. Click "Advanced" -> "Proceed to localhost (unsafe)" to continue.

📋 Requirements

System Requirements

• Python: 3.11 (Voice cloning TTS uses 3.11 strictly)
• RAM: 16GB minimum (32GB recommended for Agent Mode or smarter models)
• GPU: NVIDIA GPU with CUDA support (RTX 3050 or better recommended; RTX 50-series optimized)
• Storage: 20GB+ free space for models

Python Dependencies

All dependencies are automatically installed on first run:

• Flask (web framework)
• PyTorch & TorchAudio (deep learning)
• ChatterboxTTS (text-to-speech)
• Faster-Whisper (speech-to-text)
• Ollama (LLM backend)
• DuckDuckGo Search (web browsing)
• PyAutoGUI & Pillow (screen interaction)
• SoundDevice & SoundFile (audio processing)

📁 Project Structure

├── web_app.py          # Main Flask application with auto-dependency management
├── agent.py            # Autonomous agent logic (screen analysis, task execution)
├── tts_optimizer.py    # RTX 50-series specific optimizations
├── requirements.txt    # Python dependencies
├── README.md           # This file
├── start.ps1           # PowerShell startup script (with integrity checks)
├── start.bat           # Batch startup script (with integrity checks)
├── setup/
│   ├── smart_setup.py  # Self-healing dependency installer
│   └── ...
├── templates/          # HTML templates
│   ├── index.html      # Home page
│   ├── tts.html        # Text-to-Speech page
│   ├── stt.html        # Speech-to-Text page
│   ├── chat.html       # AI Chat page
│   ├── agent.html      # Agent Mode page
│   ├── pipeline.html   # Visual Pipeline Builder
│   └── ...
├── static/             # Static assets (CSS, JS)
├── uploads/            # Uploaded files (auto-created)
└── outputs/            # Generated audio files (auto-created)

🎯 Usage Guide

1. Home Page

• View system status
• Quick navigation to all features
• Real-time model status updates

2. Text-to-Speech

1. Click "Initialize TTS" to load the model
2. Enter text in the text box
3. (Optional) Upload reference audio for voice cloning
4. Click "Generate Speech"

3. Speech-to-Text

1. Click "Initialize STT" and choose model size (Turbo recommended)
2. Upload a file or use "Live Recording" to transcribe speech

4. AI Chat

1. Click "Initialize LLM" to connect to Ollama
2. Enable plugins like Web Search or Vision
3. Ask questions about current events or math problems
4. The AI will use tools to provide accurate answers

5. Agent Mode (WIP)

1. Click "Initialize Agent"
2. Select an Ollama model (e.g., LLaVA or Qwen3-VL for vision)
3. Type a task: "Open Notepad and write a poem about AI"
4. Watch as the agent takes control of your mouse and keyboard to complete the task
5. Emergency Stop: Click "Cancel Task" at any time

6. Pipeline Builder

1. Drag blocks from the sidebar (Mic, STT, LLM, TTS, Speaker)
2. Connect them to form a chain
3. Click "Run Pipeline" to execute the flow step-by-step
4. Great for testing custom interactions without coding

🔧 Configuration

Model Settings

TTS Configuration

• Device: Auto-detected (GPU if available, else CPU)
• RTX 50-Series: Automatically enables FP16 and CUDA Graphs if detected

STT Configuration

• Model Size: tiny, base, small, medium, large-v3, large-v3-turbo
• Compute Type: Auto, float16 (GPU), int8 (CPU)

LLM Configuration

• Provider: Ollama (local), OpenAI, Anthropic
• Model: Select from installed Ollama models

🔒 Security Notes

This application is designed for LOCAL USE ONLY

• Agent Mode: Grants the AI control over your mouse and keyboard. Use with caution and monitor execution.
• Web Server: Runs on all network interfaces (0.0.0.0) by default for local network access (so you can use it on other devices).
• No Authentication: Do not expose to the public internet.

📜 License

Copyright © 2026 Zitacron. All rights reserved.

---

Made with ❤️ for easy AI interaction