Overview
ArionTalk is a voice AI agent you can embed on any website with a single HTML tag. It reads your page content, sees your images, and responds with voice — scrolling to and highlighting the exact content being discussed. Powered by Gemini Live or an on-device local engine.
Key Capabilities
- Page Understanding — Automatically extracts text, images, and structure from any webpage. The AI knows what’s on the page before you even ask.
- Interactive Highlights — As the AI discusses content, it scrolls to and highlights the exact section or image. Works on both engines.
- Natural Voice with Barge-in — Talk naturally, interrupt anytime. The AI stops, listens, and adapts.
- Offline Mode — The local engine runs the model on-device via Gemini Nano. No server and no API keys. Speech-to-text uses Chrome 139+ on-device recognition when available, with network fallback.
Packages
| Package | Description |
|---|---|
@ariontalk/core | Shared services, types, and page understanding — no engine implementation, no UI dependency |
@ariontalk/widget | Drop-in Web Component with a ready-made UI, bundling both engines |
@ariontalk/engine-local | On-device engine using the Chrome Prompt API (Gemini Nano) and Web Speech |
@ariontalk/engine-gemini | Cloud engine add-on using Gemini Live API for real-time voice conversations |
@ariontalk/plugin-silero-vad | Silero VAD plugin for AI-powered barge-in detection |
@ariontalk/token-server | Lightweight Hono server that issues ephemeral Gemini API tokens (not on npm — self-hosted) |
Engine Comparison
ArionTalk supports two engines. Choose the one that fits your use case:
| Feature | Local (On-Device) | Gemini Live (Cloud) |
|---|---|---|
| AI model | Gemini Nano (Chrome built-in) | Gemini 2.5 Flash (cloud) |
| Interactive highlights | Yes (via inline action tags) | Yes (via function calling) |
| Image understanding | No | Yes |
| Languages | 2 (en, es) | 12 |
| Voices | System TTS | 8 Gemini voices |
| Privacy | Page content and AI stay on-device; speech recognition on-device when available | Audio and page context sent to Google |
| Backend required | No | Yes (token server) |
| Offline support | Full when on-device STT is available, otherwise falls back to network STT | No |
| API key required | No | Yes (Gemini API key on server) |
Browser Support
| Engine | Browser | Notes |
|---|---|---|
| Gemini Live | Any modern browser | Requires WebSocket + microphone access |
| Local | Chrome 139+ | Requires Prompt API origin trial |
The widget automatically hides itself on unsupported browsers (unless force is set).