How Accurate Are Transcription Timestamps? We Measured 1,373 of Them
When a transcript says someone started talking at 12:04, how close is that to when they actually did? We checked 1,373 human-marked utterance starts in 69 minutes of real meeting audio.
Word-level timestamps: a median 130 ms off
Segment-level timestamps, same engine: 250 ms
Whisper large-v3-turbo, segment-level: 460 ms
One segment timestamp in ten is off by about a second and a half
Plus when half a second doesn't matter, when it does, and three things to check before you trust a tool's timestamps.
Listen with the word-synced transcript…
The skinny
The skinny isn't ready yet — notes appear once the transcript is processed.