Speaking vs typing speed.

Where the 3x number comes from

One paper: Ruan, Wobbrock, Liou, Ng and Landay, posted as a preprint in 2016 and published in 2017, comparing speech and keyboard entry for short messages on touchscreen phones. Their result, in their own words, is that "English input rate was 2.93 times faster (153 vs. 52 WPM)" and Mandarin "2.87 times faster (123 vs. 43 WPM)".

Three things about the method, none of which survive the retelling. The keyboard was an iPhone 6 Plus running the built-in iOS keyboard, so the baseline is thumb-typing rather than a desk. The authors were split across Stanford, the University of Washington, and industry, so calling it a Stanford study is loose. And the speech system was Baidu’s Deep Speech 2, with the paper disclosing that two of its authors worked for Baidu when the work was done. None of that makes the result wrong. It does mean the honest version of the claim is that speaking beat thumb-typing by about three times, in an evaluation of one engine by a team that included two of its makers.

The finding that complicates it

Vendors like to extend the 3x number to writing generally. The obvious objection is that the study measured transcription, copying a message that already existed, and most of what you write you are making up as you go.

Somebody tested that. Foley, Casiez and Vogel, at CHI 2020, ran the comparison for composition as well as transcription. Speech was faster than typing at both, so the headline direction holds. But the margin was smaller when composing than when transcribing, and speech had the lower error rate during composition while not during transcription. Ruan’s own paper flags the mechanism, citing Shneiderman’s argument that composing by speech competes for the same cognitive resources you need to work out what to say.

So: faster, reliably, but less dramatically than the number on the poster. Quote 3x for writing prose and you are quoting a transcription result.

The typing baselines

The figures everyone cites for typing come from Karat and colleagues in 1999, who measured average computer users at 32.5 words per minute transcribing and 19.0 composing, as summarised on Wikipedia. In their sample, in other words, composing ran at about 60 percent of copying speed, and that ratio is the more useful half of the study.

Handle the rest of it carefully, though. The same study measured speech at 7.8 words per minute for composition and 13.6 for transcription, figures reported in Ruan’s related work, because that was what recognition could do in 1999. Pairing its typing figures against modern speech rates compares a 27-year-old keyboard measurement with today’s microphones, which is defensible for the keyboard half and worth stating out loud. For context on the range: professional typists sit at 43 to 80 words per minute, some roles want 80 to 95, advanced typists exceed 120, and phone typing was measured at 36.2 in 2019.

What speech runs at

Comfortable speech sits near 150 words per minute. Audiobooks are recommended at 150 to 160, and slide presentations get delivered slower, around 100 to 125, because a room needs the pauses. The fastest speaking policy debaters have been clocked from 350 to over 500.

Those are delivery rates, not text-entry throughput. Dictation costs you the pauses where you think, plus whatever cleanup the recognizer leaves behind, which is why the measured 153 in the phone study is the more honest number to reason with than the 150 an audiobook is recommended to hit.

The ceiling, and the exception

The typing record Guinness carried through its late-1970s and early-1980s editions, before dropping it, belonged to Barbara Blackburn, whose entry read that she "can maintain 150 wpm for 50 min (37,500 key strokes) and attains a speed of 170 wpm using the Dvorak Simplified Keyboard". She claimed a brief peak of 212 in 1985. Guinness dropped the category, saying records on electric and computer-driven typewriters "cannot be compared with any accuracy", and the same article notes her records were accepted without an official test after an advocate lobbied on her behalf.

Treat it as folklore with a citation rather than a measurement, then notice what it implies anyway: the most celebrated sustained typing on a conventional keyboard runs at roughly the pace of ordinary conversation. Exceptional typists do beat it. The same Wikipedia article records sustained speeds over 300 words per minute on short tests, and Blackburn’s own listed 170 is above conversational pace. The real ceiling belongs to a different machine: Guinness records 360 words per minute at 97.23 percent accuracy on a stenotype, more than twice comfortable speech, on a chorded keyboard that takes months to learn. What none of that changes is the ordinary case, where a good typist composing prose is working somewhere well below the rate they could talk at.

Where typing still wins

Speed is one axis. The phone study found speech made fewer errors during entry, 5.30 percent against 11.22 percent corrected, but left slightly more behind in the final text, 1.30 percent against 0.79 percent uncorrected. Faster input, marginally more proofreading.

Plenty of text also isn’t speech-shaped. Code resists dictation, with its brackets and camel case and exact identifiers. So does anything symbol-dense. Short edits are quicker with your hands already on the keys, and open-plan offices rule speech out socially rather than technically. Anyone telling you to dictate everything is selling something.

What it means for prompts

A growing share of what developers write is instructions to an AI, and that text is composition rather than copying: describing intent, or correcting a plan you have already been given. Foley’s result says the speed advantage there is real but narrower than the poster number, which is the honest thing to expect.

The better argument for dictating prompts isn’t the multiplier anyway. It’s that people say more than they type. A spoken instruction tends to carry the context and the caveat that a typed one drops out of impatience, and agents work better with the longer version. That case is made properly on the voice for vibe coding page, and if you want tools rather than research, the dictation comparison has them.

Common questions