Inspiration
I always found it hard to practice Korean with a real person. There are plenty of apps for flashcards, vocabulary, and grammar. What most of them miss is the immersive part: how you actually use Korean with someone in a live conversation.
In real life, people speak fast. They use words you have not seen before. You cannot pause them after every sentence and ask what each word means. So you freeze, fall back on something generic, and leave the moment without really learning.
Talkaroo started from that gap. We wanted practice that feels like talking to a person, with room to slow down, understand unfamiliar words, and craft a reply that sounds like you.
What it does
Talkaroo is Korean conversation practice for anyone trying to use the language in everyday situations.
Ordering coffee or saying your name is easy. What is hard is the interaction. Someone asks how your day was, and suddenly the word you studied yesterday is gone. Talkaroo puts you in that moment. You hear a partner’s voice, reply out loud, and practice the kind of daily chat you would actually have, not a line from a textbook.
The partner stays in character across real scenes: daily chat, café, restaurant, and directions. When you need help, coaching appears under the line. You can look up words the partner used, get reply ideas, and polish what you said into something more natural. The conversation keeps moving. Learning stays on tap.
That means you get time to think. Instead of blurting out something generic under pressure, you can build an answer that is personal and contextual, the kind of reply that helps a real relationship grow.
How we built it
Talkaroo is a Next.js and TypeScript app with React, Tailwind CSS, and Supabase for auth and session data. It is deployed on Vercel.
The speaking partner runs on Vertex AI Gemini Live with native audio. Microphone input is captured in the browser, resampled to PCM, and sent over a WebSocket bridge to the Live API. Audio and transcripts come back the same way and play through the browser. For snappier captions, we also use the browser’s Korean speech recognition and merge it with Live transcription.
Coaching is a second Gemini model, called only when you tap Understand or Polish. It analyzes a single line and returns structured help without interrupting the voice partner. Scenarios and prompts keep the partner in character, while the Learning HUD owns vocabulary, replies, and polish. After a session, a short recap highlights a win, a focus phrase, and a next mission.
Challenges we ran into
The hardest part was the voice channel itself: browser to Gemini Live and back again, with a natural turn taking feel.
Keeping that path stable on Vercel meant pinning the Live session to one instance over WebSocket. Sample rates, mic gating while the partner speaks, and echo from speakers all fought accurate recognition. Transcription was another struggle. Live finals often arrive late, interim captions jump, and Korean speech needs careful filtering so junk text does not land in the chat.
We also built this under a tight window of about two days, which forced hard choices about what had to work end to end versus what could wait.
Accomplishments that we're proud of
We shipped a full, deployed product people can open and speak into. Live voice partner, on tap coaching, scenarios, auth, and session recap all work together as one experience.
We are proud that it is more than a chat wrapper. The partner stays in the scene, and the Learning HUD sits under each line so you can learn without breaking the flow. That dual channel design is the heart of Talkaroo, and it is ready for real users today.
What we learned
We learned how voice actually works in the browser: capture, resampling, playback timing, and the fragile path between client and a realtime model.
We also learned that conversation practice needs two jobs kept separate. One model should stay in character and keep talking. Another should teach only when asked. That separation is what makes the flow feel immersive instead of like a lesson that keeps interrupting itself.
Under time pressure, we learned to ship the thinnest complete loop first: speak, hear, understand, polish, reflect.
What's next for Talkaroo
Next we want a smoother chat experience and a UI that feels closer to an intelligent assistant at your side: clear breakdowns of what the partner said and what you said, stronger corrections, and smarter next line suggestions.
We may refine levels further, but that is optional. The bigger bet is partnerships with language platforms that already teach grammar and textbooks well. Talkaroo can bring the missing piece: live voice practice that turns study into conversation.
Log in or sign up for Devpost to join the conversation.