GAEY

GAEY is a voice conversation AI that helps Chinese international students — and anyone learning American English — understand and start using authentic, everyday American English: current slang, idioms, and casual expressions. The method is simple: just talk to it like you would talk to a friend.

Lots of students have strong "textbook" English but still get lost when Americans talk casually — the slang, the references, the filler. GAEY is a low-pressure, friend-like partner to practice with: warm in tone, speaks real spoken English, without lots of profanity, without caricaturing any one accent, and without being aggressive — just the way friends actually talk.

GAEY is built on NEGA (this repository was forked from the NEGA project). NEGA is a solid ElevenLabs Conversational AI starter; GAEY keeps its clean technical foundation and swaps the persona for a friendly "teach-me-real-American-English" experience.

Docs: How to run GAEY · Roadmap (PLAN.md) · Recommended GAEY agent prompt

Features

• 🎙️ Real-time voice conversation with a friendly American "buddy".
• 💬 On-screen transcript — every line you and GAEY say, shown live so you can read along and catch the slang.
• 🐢🐇 Speech-speed control — a slider (0.7×–1.2×) to slow GAEY down or speed it up to a pace you can follow.

How it works

This web app contains no AI model. It is a thin client for ElevenLabs Conversational AI:

1. The browser asks the local /api/signed-url route for a short-lived signed URL (built from AGENT_ID + XI_API_KEY).
2. The client opens a realtime voice session (microphone + audio) to it.
3. The LLM (e.g. Gemini 2.5 Flash), speech-to-text, and text-to-speech all run on ElevenLabs' servers.

So the AI's persona, voice, model, and default speech speed are all configured on the ElevenLabs dashboard, not in this code. (Fixing "too much profanity / too heavy an accent" means editing the prompt and changing the voice on the dashboard — not editing code.)

Cost & hardware

• Running this app is free and open source — no paywall in the code.
• The only cost is ElevenLabs: there is a free tier (the setup screenshots show a free workspace with 10,000 credits), but voice conversations consume credits per minute, so heavy use needs a paid plan.
• No GPU required, ever. You never run a model locally — the model, speech-to-text, and text-to-speech all run on ElevenLabs' servers.
• You only need: a normal computer, Node.js 18+, a package manager, a modern browser with a microphone, and an internet connection.
• "Do I need a GPU to change the model / persona?" No. Changing the model is a dropdown on the dashboard; changing the persona is editing prompt text; changing voice/speed is a dashboard setting. None of it uses local compute.

Quick start

pnpm install
cp .env.example .env   # then fill in AGENT_ID and XI_API_KEY
pnpm dev               # open http://localhost:3000

You need your own ElevenLabs agent (Agent ID + API key) in .env first. See RUNNING.md for the full step-by-step setup.

Learn more

• Conversational AI Tutorial
• Conversational AI SDK