How these numbers were produced
A speaking rate needs two things measured on the same delivery: how many words were said, and how long it took to say them. Most published rates give you neither, or give you a word count from one source and a duration from another. Recorded conference talks solve it: the talk pages publish a complete English transcript in which every caption cue carries a millisecond offset into the recording.
Every talk in this dataset comes from ted.com, whose public talk pages embed exactly that — the full English transcript with per-cue timings, alongside the recording. We name the source outright because the whole point of this page is that you can go and check it: each row in the table below links to the transcript it was measured from. We are not affiliated with TED, and nothing here is endorsed by them; the talks are simply the best public material we could find where both numbers describe the same delivery.
That yields an exact interval. The words of cue i are spoken between cue
i’s timestamp and cue i+1’s, so across the window from
the first cue containing speech to the instant the last one ends, the speaker delivered
exactly the words in between. Both edges of that window are observed timestamps, not
estimates, and the rate is simply those words divided by that interval. In
37 of 39 talks a trailing non-speech cue —
a bracketed stage direction such as (Applause) or (Laughter), or
an on-screen text card — closes the window on a real instant; in the remaining
2 there is no cue after the final sentence, so the window closes at the
last cue start and that sentence’s words are left out; we do not guess at them.
Bracketed stage directions — (Laughter), (Applause),
(Music) — are removed from the word count, since nobody spoke them, but
the seconds they occupy stay in the denominator. A speaker waiting out a laugh is still
using their slot. These are therefore delivered-on-stage rates with real pauses
in them, which is the version that answers "will my script fit".
What we found
The median of the 39 talks is 165.7 words per minute. The middle half falls between 149.0 and 175.4; the full set runs from 101.8 to 210.3. To make that spread concrete: a 1,000-word script would take 9:48 at the slowest rate in the set and 4:46 at the fastest. Same script, and more than double the running time.
Set against the 125–150 wpm band that circulates as the standard speaking rate, 10 of 39 talks fall inside it, 28 are faster than 150, and 1 is slower than 125. We are not claiming the band is wrong about whatever it was originally measuring — we cannot, because we have not read a source for it. We are saying it does not describe this sample, and this sample is the one we can show you.
The extremes are worth naming. The slowest, Joe Smith’s How to use a paper towel at 101.8 wpm, is a physical demonstration: long stretches of the recording are the speaker demonstrating in silence. The fastest, Stephen Ritz’s A teacher growing green in the South Bronx at 210.3 wpm, is sustained high-energy delivery. Neither is an error; both are what the range is for.
The one judgement call, and what it is worth
Some transcript lines open with a Label: prefix. Sometimes that marks a second
voice — a host’s question, an audience response. Sometimes it is the speaker’s
own rhetorical scaffolding: Autonomy:, First:, or the
speaker’s own name. Nothing in the transcript distinguishes the two reliably, and an
early version of our script deleted all of them as "another voice" — which removed real
delivered words and moved one talk’s rate by 15.2 wpm.
We did not make the call silently: every word is counted, and the script publishes the alternative alongside it. 16 of the 39 talks contain at least one labelled line. Recomputing the whole set with those lines removed gives a median of 165.7 wpm — identical to the published figure. The ambiguity moves individual talks (the largest shift is 15.2 wpm, on Joe Smith’s How to use a paper towel) but not the number this site actually uses. Both columns are in the table below so you can see it for yourself.
Known limits of this sample
- Genre. These are rehearsed, coached, strictly time-limited conference talks. That is the closest public material to a prepared speech, and it is still not the same as a nervous classroom presentation or an unscripted meeting.
- Filler words. The transcripts are lightly cleaned by their publisher, so hesitations are largely absent and the true acoustic rate of these deliveries was somewhat higher. For matching a written script — which also has no "um" in it — the cleaned count is the right basis.
- Language. English only. Rates are not portable across languages, because languages differ in how much meaning a word carries.
- Selection. The candidate list is well-known talks across a range of lengths and speakers, not a random draw from all recorded talks. Popular talks may be better delivered than typical ones.
- Sample size. 39 talks is enough for a median and a range, not for confident claims about subgroups. We do not slice it by speaker, topic or year, because at this n those slices would be noise.
- Dropped candidates. 1 candidate was dropped by an automated
guard rather than by hand:
nick_vujicic_overcoming_hopelessness: HTTP 404. The guards check that the audio and transcript are English, that a single speaker is credited, and that the measured window exceeds a minute.
Every talk we measured
Sorted slowest to fastest. Window is read from the transcript's own timestamps; words is counted from its text by the rule above; wpm is their quotient. Ex-label is the sensitivity variant described above — identical to the published rate wherever a talk has no labelled lines. Published length is the whole recording including titles and applause, shown so you can see it is always longer than our measured window.