Is dictation faster than typing?

What changed was the error rate, not the voice

People spoke at the same rate in 1999 as in 2017. What moved is how often the machine got a word wrong, and therefore how much of the session was spent going back. Karat’s participants were not slow talkers — they were correctors. Once recognition became reliable enough that correction stopped dominating the session, the same activity flipped from twice as slow to three times as fast.

This is why “is dictation faster” has never been a question about speech. It is a question about the recogniser you happen to be using, on the language you happen to be speaking. A tool that transcribes your accent poorly puts you back in 1999, whatever the marketing page says.

Where the gap narrows

Both figures above come from transcription: the sentence was handed to the participant, who only had to get it into the device. That is a fair model of replying to a message, and a poor model of writing something you have not worked out yet.

Foley, Casiez and Vogel added a composition task and found speech still ahead, but by a much smaller margin — and with no advantage at all on perceived workload, except for physical demand (CHI 2020, doi:10.1145/3313831.3376861).

Gould and Boies had explained why, forty years earlier: planning takes roughly two thirds of composition time regardless of the output mode (Science, 201(4361), 1145–1147, doi:10.1126/science.201.4361.1145). Speaking removes the typing bottleneck. It does not remove the thinking one, and on a hard paragraph the thinking is most of the job.

What this does not prove

• Not “three times more productive.” Ruan measured a text entry rate on a short transcription task. Productivity is a different quantity, and it was not measured.
• Nothing about a physical keyboard. Both modern measurements used a touchscreen. The desktop comparison, with a current recogniser, has not been published — so every multiplier you read for desktop work, ours included, is an extrapolation.
• Nothing about long or technical text. The task was short messages. Names, code, jargon and acronyms are exactly where recognisers still lose, which is why a personal dictionary matters more than a headline speed figure.
• Nothing about accuracy rankings. We publish none, and we would distrust anyone’s: in this category every accuracy claim comes from a benchmark its own vendor designed. Ruan’s study is worth one caveat of the same kind — the engine tested was Baidu’s, and one co-author was its chief scientist at the time. The venue is independent; the engine choice was not neutral.
• Nothing about memory. Saying words aloud does help you remember them — but so does typing them, which is also a form of production (Forrin, MacLeod & Ozubko, Memory & Cognition, 2012, doi:10.3758/s13421-012-0210-8). “Dictating helps you retain more than typing” has never been tested. We are not going to be the first to claim it.