Verba
A browser-based transcription tool that records speech, sends it to Whisper for transcription, and streams the text back in near real time.
- Next.js
- TypeScript
- Whisper
- Supabase
- Built in
- 2 days
Overview
Verba records audio in the browser, runs it through OpenAI's Whisper large-v3-turbo model, and streams the transcript back a few seconds later.
What was built
A Next.js app with Supabase handling accounts, a browser-side voice-activity detector that only sends audio once it hears real speech, and a Hugging Face-hosted Whisper endpoint doing the actual transcription. It filters out the handful of stock phrases Whisper hallucinates on silence (things like "thank you" or "thanks") before they ever hit the transcript.
Performance
The first pass felt sluggish, so I went back and tuned the voice detector across two rounds: cut the minimum speech duration in half, from 1000ms to 500ms, cut the reprocessing debounce 60%, from 5000ms to 2000ms, and cut the volume threshold 75%, from 0.02 to 0.005, so quieter speech still gets picked up. It just feels a lot snappier now.