Why I built it
I built it specifically for public speeches and university club events where recording is allowed. Everyone has a certain attention span, and the world is a big place with a lot of accents in it. At first, any unfamiliar accent can feel like an alien language, but recording and transcribing solves that, and I get captions.
I still use it for speeches on topics I'm less into, when I don't want to miss the logic or fall behind.
Problem
At a public speech or a club event, a fast speaker or a heavy accent makes it easy to miss key points, and audio from the back of the room is often too quiet for speech recognition.
Approach
Audio is captured with PyAudio on a background thread, and an automatic gain stage boosts distant voices up to 8x. A local faster-whisper model transcribes it live, with timestamps, without needing a powerful GPU.
When the session ends, the transcript is saved as Markdown and summarized into key points, definitions and action items with Gemini, with a local fallback when offline. It can also transcribe a saved recording.
Results
Light enough to run through a long talk on a laptop. Building it meant handling live audio without blocking and tuning the gain so quiet voices come through without clipping.