The Pauses That Weren’t His

By Noam Shemla ·

By Noam Shemla · Speeches That Lasted

TL;DR We measured Steve Jobs’s introduction of the iPhone at MacWorld on 9 January 2007 with the same instrument our members record into. This time we had real audio instead of a scraped transcript with an estimated runtime. At 146 words a minute it scores a perfect 100 on pace, the only entry in this series to clear the band. 511 of his 9,913 words are filler (5.1 percent, the highest rate this series has measured), and the real pause data surfaced something the text-only version couldn’t have caught. Three of the recording’s longest silences come from somewhere else entirely: the exact points where Schmidt, Yang and Sigman’s own remarks were cut out of this corpus.

Every entry before this one in the series failed the pace band on the slow side, and for a real reason: the band was tuned for a three-minute practice clip, and a historic oration runs long, often with an audience that interrupts. This is the first entry that doesn’t fail that way, and this time the most interesting thing real audio turned up sits in the pause list.

One keynote, four voices, and the three we cut

Every other entry in this series measures a single continuous speech from one primary source. This one is built from one, longer recording with three other people’s voices inside it, and getting Jobs’s own numbers right meant cutting his real remarks out of the full keynote instead of treating the whole event as one speech.

This corpus covers only Jobs’s own words, drawn from the iPhone-introduction portion of the roughly two-hour keynote, from “This is the day I’ve been looking forward to” through the closing Wayne Gretzky line. It excludes on-stage remarks from three guests who speak during this stretch: Google’s Eric Schmidt, Yahoo’s Jerry Yang and Cingular’s Stan Sigman. A fair comparison to Obama, Lincoln, Churchill and King needs one person’s own words, and a transcript of the whole event would not give us that.

An earlier version of this entry estimated where those guest segments fell from a third-party published breakdown of the keynote. With the real recording and its timed captions in hand, those boundaries are now found in the recording itself: Jobs speaks 0:10 to 51:00, hands off to Schmidt, resumes 53:36 to 54:09, hands off to Yang, resumes 56:46 to 68:37, hands off to Sigman, and closes 74:30 to 79:05. Total measured speaking time: 67 minutes 50 seconds, against a 79-minute 10-second full recording.

This speech is also still in copyright (Apple’s), so, as with the other three entries in this batch, we work from structure and short attributed fragments instead of the transcript in full.

The one entry that doesn’t fail on length

That’s an interesting result, and the limits of what it says matter. It doesn’t say Jobs is a more skilled speaker than the other four people in this series. It says a live demo, where the speaker talks almost continuously through his own actions (“I just push it right here, and boom, I’m in the phone”) instead of pausing for a crowd to finish reacting to a rhetorical line, produces an elapsed-time pace that looks, to our instrument, like an ordinary two-minute practice recording. The pace band was testing for continuous, unbroken talking all along, and a demo turns out to be closer to that than a speech interrupted by applause thirty times.

A pause list that caught its own mistake

  • 78 pauses ≥0.5s. From real word-level timing gaps.
  • 3 are edit artifacts. The 3 longest (352.7s, 156.8s and 155.9s) sit exactly at the Schmidt/Yang/Sigman cut points.
  • 75 real pauses remain. Longest: 17.9 seconds, mid-demo.
  • 24 hesitant (<1.5s recovery). Against 45 deliberate ones.

The pause data, and the correction it needed before it meant anything.

We have left the mistake in instead of quietly fixing it after the fact. The audio fed to the pipeline is the full keynote recording, because that’s the only file that exists, but the transcript is Jobs-only, with the guest segments removed. The result is that the three biggest gaps between his timed words are not him pausing for six minutes; they’re the six minutes Eric Schmidt was talking. A pause detector has no way to know the difference between a speaker holding silence and a corpus that skipped a chapter. Once those three are set aside, 75 real pauses remain, the longest running 17.9 seconds. That one is itself almost certainly a slide-advance or a demo moment instead of Jobs standing in silence, but at least it is plausible, and no guest speaker in disguise.

Of those 75, our pause classifier calls 24 “hesitant”: under the 1.5-second mark that separates a held beat from a stumble recovering. That’s a specific number to weigh against Jobs’s reputation as an unusually poised presenter: even with the artifacts removed, roughly a third of his measurable pauses read, by this threshold, as recovery rather than intention. The instrument can’t tell you which ones from the pause length alone. “No, actually here it is” could be a bit landing exactly as planned, or the mic pickup on a real stumble finding the pause detector by the same route. What it can tell you is that the deliberate-sounding pauses (45 of them) outnumber the hesitant-sounding ones roughly two to one, which is a more specific and more interesting claim than “he pauses a lot” or "he never fumbles."

Narration in short steps

That average is the clearest signal that this is a different kind of speaking than the rest of the series. Every other entry here alternates long, built sentences with occasional short ones for effect. This one is short almost everywhere: narration broken into small, sequential steps (“I just push it right here,” “and boom, I’m in the phone,” “now I want to show you something else”), because that’s what talking through a live interface, one tap at a time, sounds like.

Openers that are habits, not devices

Compare these lanes to King’s. “I have a dream” and “let freedom ring from” are rhetorical devices, built and deployed once each, for effect. “And uh,” “and i,” “this is” and “so i” are the connective tissue of live, unscripted narration: the verbal equivalent of a hand gesture, used to bridge from one screen tap to the next dozens of times over. Our repetition figure cannot tell the difference between anaphora and a verbal tic; it counts sentence-opening recurrence, full stop. Here, that crudeness is itself informative. The real, word-timed transcript surfaces “and uh” as the single most repeated opening phrase in the entire series, a filler-word pairing leading a field that includes “we shall fight” and “I have a dream.”

The number that doesn’t flatter him

  • 9,913 words. Jobs’s own words only, guests excluded, real audio-timed count.
  • 67m 50s elapsed. Measured from real segment boundaries in the recording.
  • 1,072 sentences. Mean 9.4 words, the shortest of any entry in this series.
  • 511 filler words. 5.1 percent of everything he said, the highest rate in this series.
  • 68.5 clarity score. The lowest clarity score this series has measured.
  • 16.7% vocabulary diversity. The lowest in the series. Gettysburg scored 50.9%.

The full set of numbers, now including real acoustic measurement alongside the text.

Publishing numbers like these is why the series exists. “The presentation secrets of Steve Jobs” is a small publishing genre in its own right, and the pace score above would fit neatly inside it. The filler rate and the vocabulary-diversity score would not. 511 filler words (“so” alone accounts for 161 of them, “uh” for another 104) is a real filler rate over five percent, higher than any prepared political address in this series and higher than the text-only estimate this entry originally published. None of that makes the keynote a bad presentation. It was live and improvised, with the verbal roughness that comes with narrating your own actions in real time instead of reading a drafted text. The instrument doesn’t know the difference between a speech and a demo. It counts what’s there, now from the real recording instead of a transcript someone else cleaned up first.

  • 195 Hz mean pitch. ±69 Hz, not monotone by our threshold.
  • 0.11 avg volume. Range 0 to 0.72 across the full recording.

Pitch and volume, measured across the full 79m 10s recording (the tool members record with only keeps the first 4m 10s of any single clip).

What the recording still hides

Three more recordings, and the one where he is not selling

We later measured three more Jobs recordings: an unscripted 1997 answer from the floor, the 2001 iPod launch, and the 2005 Stanford commencement address. Three of the four are product events. One is not.

This runs against the received account of him, which is built almost entirely on the keynotes. The instrument agrees the keynotes are the more animated performances. It scores them dynamic and marks their energy high. It just does not rank them first. The recording where he stands still and talks about death is the one that comes out ahead.

Pace and clarity are separate questions

Pace and clarity are different axes, and this entry is the cleanest proof of that this series has produced. A confident, well-rehearsed live demo can post a perfect elapsed-time pace score while still running a filler rate five times higher than a prepared political address, because the two numbers measure different things. One measures whether you’re talking continuously enough for a listener to stay with you. The other measures whether the words you’re using, while you do, are doing real work. Separately, a pause list is only as reliable as the corpus it’s built from. If you ever splice or trim your own practice recordings before reviewing the report, check for exactly this: a gap in the footage reads to the instrument exactly like a gap in your speech.

The same audio pipeline our members record into: word-timed pace and pauses from timed captions, pitch and volume from the decoded recording.

Measured

  • Word count and pace against real elapsed time
  • Real pauses (0.5s+) from gaps between timed words
  • Pitch: mean 195 Hz, ±69 Hz, not monotone
  • Volume: average 0.11 across a 0 to 0.72 range
  • Filler and hedging counts and rates
  • Vocabulary diversity: 16.7 percent of words distinct
  • Positions of every sentence-opening phrase that repeats three times or more

Inferred, not measured

  • Word timings are spread evenly inside each caption cue, so the pause list is a floor, not a total
  • The transcript is a third-party auto-caption of a live event, not a primary source we produced
  • Guest remarks (Eric Schmidt, Jerry Yang, Stan Sigman) were identified by their introduction and hand-back timestamps and excluded from this corpus; 3 of the raw pause list’s longest entries are the resulting edit-boundary artifacts, and no silence of Jobs’s own

Sources