The pace bench

It is almost certainly edited. Across 4,798,631 transcript words the bench finds 359 hesitation sounds (um, uh, er, erm, ah, hmm, mm), 0.75 per 10,000 words, and only 214 of the 2,299 talks contain any at all. Unscripted speech is not that clean, so the transcripts have been tidied for reading. So your phone transcript will show ums that a TED transcript never would, and comparing your um count with TED's is comparing a recording with an edited text. The bench shows TED's number for hesitations and does not rank you against it.

The running time is the whole video. It includes applause, laughter, the walk on and any clip the speaker played. Words per minute here is pace over the whole talk, slightly lower than the pace of the speech itself. Seventy-four percent of the talks carry at least one laughter cue and ninety-five percent at least one applause cue. Your own recording of an opening has none of that dead air, so it reads slightly fast against TED, not slow.

Talks. The Kaggle dataset "TED Talks" (rounakbanik/ted-talks) joins 2,550 talk records with 2,467 transcripts; after removing three duplicate transcripts, 2,464 talks have both. The bench keeps the 2,299 with one speaker, at least 300 transcript words and fewer than three music or singing cues, which removes performances. They were filmed between 1984 and 2017. TED's own usage policy licenses talks under CC BY-NC-ND 4.0 and does not permit edited transcripts, which is why only counts are shipped and no line of any transcript appears on this page.

Rules. Curly apostrophes become straight. A parenthesised cue is counted and then removed before counting words. A word is a run of letters, digits or apostrophes. A sentence ends at a run of full stops, question marks or exclamation marks, and trailing words with no end mark count as one more. Pause marks are full stops, commas, semicolons, colons, question and exclamation marks. Fillers: um, uh, er, erm, ah, hmm, mm (hesitations) and you know, I mean, sort of, kind of, basically, literally (phrases), matched as whole words. "Like", "so", "actually" and "right" are not counted, because a word list cannot tell the filler from the meaning.

The Python reference and the browser engine apply the same rules and agree on every count for all 2,464 transcripts, and on every bench number to six decimals.

Rung 1 runs at 171.3 words a minute, faster than 79.5% of the talks, with six sentences averaging 52.3 words each: longer than 99.8% of TED talks' average sentence. It is not the speed of a nervous mouth so much as the shape of an unplanned one: clauses joined with "and" because the speaker is composing while talking. Cutting to one idea (rung 2) drops 75 words and brings the pace to 136.6 and the sentences to 15.9 words, the TED median, without asking Noor to slow down at all.

Rung 3 has 28 sentences averaging 7.5 words and 18.0 pause marks per 100 words, more than 95.8% of the talks. Short sentences and marked pauses are what "Six days. For a refund." looks like on paper. Its pace, 126.6, is slower than three quarters of TED talks, and that is not a problem: a deliberate opening is meant to be slower than the middle of a talk. The bench reports where you sit; it does not grade a pace as good or bad, because there is no measured pace that is correct.

Rung 4 shows one phrase filler. It is "so you know where it sits": a real "you know", meaning "so that you are informed", not a verbal tic. A word list cannot tell the two apart, which is why the bench lists every match for you to judge instead of printing a score, and why "like", "so", "actually" and "right" are not counted at all. Rung 1's ten phrase fillers are the real thing: "sort of give you an update", "basically", "kind of grab". Every hesitation in rung 1 is an um or an uh that a phone transcript kept and a TED transcript would have removed.

Same slot, 150 more words. Pace goes from 126.6 to 216.6 words a minute: past the real 99th percentile of 205.7 and faster than 99.8% of the talks. Sentences go from 7.5 words to 60.2, because there is no time to stop, so the full stops become "and". Pause marks fall from 18.0 to 6.4 per 100 words, fewer than all but 0.1% of the talks. Pause marks are where a transcript gives a listener somewhere to catch up; the faster version removes almost all of them.

Whether the room understood. The widget prints "Comprehension: not measured" on every reading, because a count of words cannot know. It would be easy to claim that 216 words a minute "loses the room"; the bench cannot show that, and neither can a pace chart. What it shows is that the crammed version sits outside what 99 in 100 real speakers did, with almost no full stops. Whether that costs understanding is a question for research on listeners, which is a different kind of evidence from a transcript.

In a 1970 study by Thomas Sticht, 280 Army inductees listened to three short recorded passages (36 to 56 seconds each) at a normal 165 words a minute, or machine-compressed to 206, 275 or 300. Comprehension at 206 words a minute did not differ significantly from its slowed-back-to-normal control; at 275 and 300 it was significantly worse. The same paper repeats a broader earlier review as finding little loss up to about 275, which sits awkwardly with its own result at 275.

Three reasons not to stretch it to a talk. The speech was compressed by a machine, which shortens every pause in proportion; a speaker who hurries drops pauses altogether, which is exactly what rung 4's 6.4 pause marks show. The passages lasted under a minute, not ten. And the listeners were tested on purpose, not sitting at an all-hands after lunch. Rung 4's 216.6 sits between where Sticht found no significant drop and where he found one. The honest summary is that the bench cannot tell you whether the room kept up, and neither can this study.

Across the real talks, word count and running time are tightly linked (correlation 0.922): longer talks have more words, at a fairly steady pace. Press the break button and each talk's words are divided by some other talk's minutes. The median barely moves (149.1 on the first shuffle) but the middle half spreads from 137 to 168 out to 98 to 230, the correlation falls to -0.020, and 49.3% of talks land outside the real 1st to 99th percentile range, against 2.0% before. Two more shuffles give 49.8% and 50.5%. The tight band is a fact about speakers, not about the arithmetic, and this is what the bench looks like when there is nothing to measure.