English Teacher Agent: real-time feedback while you speak

I’ve been building english-teacher-agent, a small project born from a simple observation about my own English lessons.

The motivation

When I turned on the AI analytics feature in my English learning app, I found out that in a 25-minute class I was speaking for 21 minutes, while the teacher spoke for only 4. That’s 84% of the time being me.

If your main goal is practicing speaking, the teacher’s role in that interaction is structurally replaceable by AI — as long as it’s natural enough to keep the conversation flowing. I already do two daily 15-minute sessions with ChatGPT Voice Mode, and that works well. So the goal of this project isn’t to build yet another conversational agent. The real value to add is real-time feedback while I speak.

Demo

What it does

It’s a web app (and potentially a Google Meet plugin) with four core features:

• Live transcription — your words appear on screen as you say them, in real time
• Inline corrections — grammar mistakes are underlined in red with a correction tooltip, a few seconds after they happen
• Ghost text — after 3 seconds of silence, a grayed-out suggestion appears showing how to continue the sentence
• Filler word tracker — words like “uh”, “like”, or “you know” are highlighted in amber, with a per-session counter for each one

Stack

The backend bot is built with pipecat in Python, and the frontend is a Node.js app. Locally, it’s just:

# backend
cd server
uv run bot.py

# frontend
cd client
npm run dev

Code is up at github.com/quezadasubiabre/english-teacher-agent-public.