Inspiration

After years of language apps, I could ace quizzes but froze when actually speaking. Traditional apps teach you to read and tap—but fluency requires talking.

The problem? No safe place to practice. Native speakers are intimidating. Tutors are expensive.

Then I found Google's Gemini Live API—real-time voice AI with near-zero latency.

What it does

Talkora is a voice-first language learning app with two modes:

Mode Function
Learn Talk with AI characters that adapt to your CEFR level
Translate Real-time bidirectional voice interpretation

Key Features: Real-time voice conversations Live pronunciation and grammar feedback Smart hints when you're stuck Unlockable personas based on skill level Custom character creation

How we built it

I used Google AI Studio to vibe code the initial prototype, leveraging its rapid prototyping capabilities with Gemini. From there, I manually refined the code—tweaking prompts, adjusting the UI, and fine-tuning the audio handling.

Layer Tech
Frontend React 19 + Tailwind CSS
AI Gemini Live API
Audio Web Audio API

Multi-model strategy: Voice conversations → gemini-2.5-flash-native-audio Hints and grammar → gemini-3-flash Avatar generation → imagen-3.0

Challenges we ran into

Audio sync — Preventing stuttering in real-time playback Interruptions — Handling rapid speaker switches in translator mode Feedback balance — Corrections needed to feel helpful, not annoying

Accomplishments that we're proud of

Real-time voice conversations with near-zero latency—it actually feels like talking to a person Seamless audio pipeline that handles browser-to-API format conversion without stuttering Adaptive difficulty that adjusts vocabulary and speed based on CEFR level Dual-mode functionality combining language learning and live translation in one app Persona unlock system that gamifies progression and keeps users engaged Smart hint system that rescues users from awkward silence without breaking immersion Built a fully functional prototype using vibe coding in Google AI Studio

What we learned

manual refinement where vibe code sometimes fail expectation Prompt engineering

What's next for Talkora

Progress check and maybe make this into a game

Built With

Share this project:

Updates