Transcribe WebM

WebM is what a browser writes. Screen and microphone recordings made with the MediaRecorder API — Loom, Chrome extensions, in-page recorders, our own extension — all produce it, and so does downloading from a site that serves WebM.

The numbers here say the same: 3,010 files, 8% of uploads, median 8 minutes 26 seconds at 129 kbps. That bitrate is audio-shaped, not video-shaped — most WebM reaching us is a browser capture of speech rather than a downloaded film.

Both variants upload as they are, along with .weba for the audio-only form.

These are medians measured on our production cluster, expressed as time per hour of audio.

Against the WebM files uploaded here: the median, about 8 and a half minutes, comes back in roughly 12 seconds. Nine in ten are under 49 minutes.

This is the one WebM-specific problem worth knowing about. MediaRecorder writes files without a duration field — the header says the recording is of unknown or infinite length, because the browser does not know when you will press stop.

Players then show 0:00 or a nonsensical figure, and some tools refuse the file outright. It does not affect transcription here: the duration is measured from the stream itself rather than trusted from the header, so the transcript and its timestamps come out correct even when your player disagrees.

If you need the file to behave elsewhere, remuxing it fixes the header without re-encoding: ffmpeg -i in.webm -c copy out.webm.

Duration, comfortably. At 129 kbps a 5 GB file would be over 80 hours, so the 8-hour cap arrives far sooner; the largest WebM processed here was 848 MB.

• Long screen recordings are the common case. A three-hour webinar is well within limits at these bitrates.
• Opus is the usual audio codec, and it holds up well. Speech stays intelligible at bitrates where MP3 would smear, so browser captures transcribe cleanly.