Word-level transcription timestamps landed a median 130 milliseconds from when a person actually started speaking; segment-level ones from the same engine 250 ms, and Whisper's 460 ms. That is what we measured on 69 minutes of real meeting audio, and it is what this first episode of The Transcript Lab is about.
Free
The Transcript Lab is a short podcast from TranscribeNext about what transcription actually does, measured rather than assumed. Every number in the episode comes from data we collected ourselves and publish in full.
Listen and read along
The player below runs the episode next to its transcript, produced by TranscribeNext from this exact MP3. The highlighted word is the one the engine timed at that moment — the same word-level timestamps the episode is about. Click any line to jump to it.
The Transcript Lab · Episode 1
How accurate are transcription timestamps? · 4:46
0:00 / 4:46
The transcript is what TranscribeNext returned for this exact MP3, with speaker identification on and the two speaker labels renamed to the hosts. One correction: the first half of Chris’s line at 2:38 (“If you’re making captions,”) had been attributed to Laura. Everything else, including numbers written as digits, is unedited output. Line and word timings are measured — this run returned per-word boundaries, so the highlight follows the words as the engine timed them.
What the episode covers
How the test was set up: two meetings from the AMI Meeting Corpus, 69 minutes recorded on a single table microphone, 1,373 human-marked utterance starts
Word-level timestamps: median error 130 ms, about three in four within a quarter of a second
Segment-level timestamps: 250 ms from the same engine, 460 ms from Whisper large-v3-turbo
The bad cases: one segment timestamp in ten is off by roughly a second and a half
When half a second does not matter (finding a quote) and when it does (captions, word highlighting, cutting video)
The full method, the results table and a CSV of every measured onset are in the guide to transcription with timestamps.
