If you record Zoom or Google Meet calls locally, you've probably noticed an annoying mismatch: you only care about the spoken words, yet you're stuck dragging around a 900MB .mp4 file full of 1080p screen-share frames.
Most online transcription tools make you upload the entire video container over your connection, wait 10 minutes for a progress bar, and then throw away 95% of the bytes on the server just to run ASR on the audio track.
When you're building or using developer tools for meeting notes, there are really two separate problems here:
- Social / compliance friction: Bot-based tools (Otter, Fireflies) join the live call as a participant. On client architecture reviews or NDA calls, people hate seeing an uninvited bot in the participant list.
-
Payload bloat: Post-call recording avoids the bot, but raw
.mp4/.movfiles from screen recordings are absurdly heavy to move around.
The cleanest fix I've seen—and ended up adopting for my own call notes—is doing demuxing client-side before the network request ever fires. Instead of sending a 1GB video container, the browser reads the file locally, strips the video stream, and extracts a compressed audio track (~25–35MB for a 1-hour call).
Here's roughly what that cuts out compared to a standard server-side pipeline:
Traditional: 1.0 GB MP4 ──(slow upload)──▶ Server ffmpeg ──▶ ~30 MB audio ──▶ Whisper
Client-side: 1.0 GB MP4 ──(local demux)──▶ ~30 MB audio ──(fast upload)──▶ Whisper
I've been using this browser-based meeting transcription flow when I don't feel like spinning up a local faster-whisper environment on a travel laptop. It handles the local track stripping automatically, runs Whisper large-v3-turbo, groups speaker turns so you can rename SPK_0 / SPK_1 afterward, and wipes files after 24 hours.
A couple of honest caveats from using this on engineering standups:
- Multi-speaker overlap still hurts: If four devs talk over each other during a heated bug triage, speaker turn grouping will merge or mislabel turns regardless of how fast the upload was.
-
Strict air-gapped work: For internal security audits or medical/legal audio that cannot leave your VPC or machine at all, browser uploads aren't the answer—you still want a local
mlx-whisper/faster-whisperor self-hosted Docker setup.
Still, for day-to-day standups and client syncs, moving audio extraction to the browser removes almost all the waiting without forcing a bot into the meeting.
Top comments (0)