My Live Translate

The problem

Attending a lecture, meeting, or event in a language you are still learning is exhausting. The cognitive load of real-time translation leaves little bandwidth for actual comprehension. Existing tools are either asynchronous (upload a recording after the fact) or require specialized hardware.

What was built

My Live Translate is a real-time audio transcription and translation pipeline. It captures audio input from a microphone, transcribes it in near-real-time, and outputs a translated version of the spoken content — live, with minimal latency. The backend is built on FastAPI (Python) for the audio processing and translation pipeline; the frontend is lightweight JavaScript for display and audio capture. Storage uses SQLite for session history.

My role

Solo builder. This was a personal tool that became a small project. The architecture prioritizes low latency over perfect accuracy — for live use, a slightly imperfect translation that arrives in two seconds is more useful than a perfect one that arrives in eight.

What I learned

Real-time audio processing has hard constraints that batch processing does not. Every architectural decision — buffer size, transcription model choice, network protocol — has a direct effect on the latency experienced by the user. Building for real-time taught me to think in milliseconds rather than seconds.