ASR-Workshop-Demo
The Problem
In real-time captioning, transcripts need to be correct the first time. ASR gets you speed but not accuracy. Manual transcription gets you accuracy but not speed. I wanted to find a way to mesh both—use ASR as a foundation while allowing for real-time manual correction.
What I Built
A PyQt6 desktop app with live audio capture, streaming transcription via Deepgram’s API, and real-time waveform visualization. The idea was to create an interface where ASR and human input could coexist.
What I Learned
This approach didn’t solve the problem I was after. The fundamental tension between speed and accuracy needs a different solution than what I landed on here. But it was a good learning experiment—I got comfortable with async audio streaming, WebSocket APIs, and Qt’s signal-based architecture.
Sometimes you build something to find out it’s the wrong direction. That’s still useful.