VoxAgent Neural is a high-performance Speech-to-Text platform designed for the Edge. Unlike cloud wrappers, it runs a quantized Neural Engine locally (CPU-Optimized), delivering sovereign, low-latency transcription via a persistent Agentic WebSocket connection.
Initialize the Neural Link:
# 1. Start Backend (The Brain)
cd backend && python main.py
# 2. Start Frontend (The Interface)
cd frontend && npm run devSetup Guide: See GETTING_STARTED.md for Python venv setup.
Audio In -> LiveKit (UDP) -> Quantized Model -> Text Out (TCP)
The Control Plane separating Signal Transport (LiveKit) from Inference (Python)
Solving Head-of-Line Blocking by using UDP for Audio and TCP for Text
Deep Dive: See ARCHITECTURE.md for the "Turnaround Time" logic.
- ⚡ CPU-Optimized: Runs
faster-whisper(INT8) on standard consumer hardware. No GPU needed. - 📡 Hybrid Transport: Combines WebRTC (UDP) for audio stability and WebSocket (TCP) for text reliability.
- 🕵️ Stealth UI: Minimalist "Agentic" interface with zero-distraction animations.
- ⏱️ Latency Awareness: Real-time "Turnaround Time" (TAT) metrics warn users if they speak too fast.
| Document | Description |
|---|---|
| System Architecture | Hybrid Control Plane & Quantization. |
| Getting Started | Python/Node Setup & LiveKit Config. |
| Failure Scenarios | Handling CPU Saturation & Network Jitter. |
| Interview Q&A | "Why Faster-Whisper?" and "UDP vs TCP". |
| Component | Technology | Role |
|---|---|---|
| Brain | FastAPI (Python) | Inference Engine. |
| Model | Faster-Whisper | Quantized STT. |
| Transport | LiveKit | WebRTC Signaling. |
| Interface | React + Vite | Agent UI. |
Harshan Aiyappa
Senior Full-Stack Hybrid Engineer
GitHub Profile
This project is licensed under the MIT License - see the LICENSE file for details.