Disclosure: published by Relay; author Raymond Zhu is a founding member. This is a research summary, not a hands-on test, and it makes no speed claim about Relay. Studies and figures were checked against the primary sources linked below on October 29, 2026. Sources: Ruan et al. (2017), PACM IMWUT, Ruan et al. full paper (PDF), Dhakal et al. (CHI 2018), Yuan, Liberman and Cieri (Interspeech 2006), Relay for macOS.
The short answer: faster to enter, not always faster to finish.
Speaking is usually quicker than typing at getting words in. Whether it is quicker at finishing a piece of work depends on how much you then fix, rearrange and check. The research below is good evidence for the first part and much weaker evidence for the second.
This article looks only at the speed question and how much weight the numbers can carry.
What the best-known study actually tested.
The most cited comparison is Ruan, Wobbrock, Liou, Ng and Landay, “Comparing Speech and Keyboard Text Entry for Short Messages in Two Languages on Touchscreen Phones”, published in PACM IMWUT in December 2017. Forty-eight participants entered short phrases on an iPhone 6 Plus in a lab, using either a speech recognizer (Baidu’s Deep Speech 2) or the built-in iOS keyboard.
In English, speech entry ran at 153 words per minute against 52 for the keyboard, about 2.93 times faster. In Mandarin, it was 123 against 43, about 2.87 times faster. The full paper describes the method in detail.
An earlier 2016 preprint, titled “Speech Is 3x Faster than Typing for English and Mandarin Text Entry on Mobile Devices”, reported slightly different figures. The peer-reviewed 2017 version is the one to cite.
Sources: Ruan et al. (2017), 2016 preprint.
Errors: fewer along the way, a few more left behind.
Speed alone would be misleading if the text came out worse, so the study separated two kinds of error. The corrected error rate counts mistakes made and fixed during entry: 5.30% for speech against 11.22% for the keyboard in English. The uncorrected error rate counts what was left in the final text: 1.30% for speech against 0.79% for the keyboard.
Read together, speech produced fewer slips to fix while entering text, but the errors it did make were slightly more likely to survive. That matches a common experience: a misheard word looks plausible, so it is easier to miss than a typo.
Sources: Ruan et al. full paper (PDF).
What about a desktop keyboard?
The 52 words per minute in Ruan et al. is typing on a phone. For computers, a separate large study by Dhakal, Feit, Kristensson and Oulasvirta at CHI 2018 analyzed online typing tests from 168,000 volunteers and found an average of 52 words per minute, with wide variation between people.
It is tempting to set that desktop average against 153 words per minute for speech and call it settled. That would not be a measured comparison: the people, tasks and test conditions differ, and neither test includes the time spent deciding what to say. What the two studies do suggest is that many people type on a computer well below the pace at which they can speak.
Sources: Dhakal et al. (CHI 2018), EurekAlert summary.
How fast people speak, and why that isn’t a dictation speed.
Conversation is quick. Yuan, Liberman and Cieri (Interspeech 2006) measured 2,438 recorded telephone conversations and found an average of 196 words per minute over total elapsed time, ranging from 111 to 291 across conversations, with both speakers’ words combined. Measured turn by turn it was 164; excluding silences, 236.
Those figures describe people chatting, not people composing. When you dictate, you pause to decide what comes next, and that thinking time is real time. A useful way to see it: speech removes much of the mechanical cost of entering words, but not the cost of knowing what you mean.
Sources: Yuan, Liberman and Cieri (2006).
What none of these studies measure.
Editing. Short phrases need little restructuring. In a report or a long email, moving paragraphs around and fixing a misheard name can eat into the time saved during entry.
Setting. Labs are quiet.
Task type. Code, tables and one-word fixes are often still quicker by keyboard. Many people end up mixing the two: voice for drafting, keyboard for fine control. Prompts to AI tools are one place where saying more helps; how to dictate better AI prompts covers that.
How to answer the question for your own work.
Pick three tasks you do every week, such as a reply, a short update and a longer note. Do each one twice, once typed and once dictated with whatever tool you use, and time it from the first word to the moment it is ready to send, including every correction. Use the same kind of task rather than the same text, so you aren’t just retyping something you’ve memorized.
If dictation wins only when you skip proofreading, it hasn’t won. If it wins on drafts but loses on edits, that is still a useful answer: use each input for the part it does well.
Our position is a belief, not a measurement: we believe speaking carries ideas at higher fidelity and throughput than typing, which makes it the path of least resistance for putting intentions into a computer. The studies above are why that belief is reasonable for entry speed. Whether it holds for your finished work is something you can check in a week, with Relay for macOS or any other tool. To start with the tools you already have, see how to use dictation on a Mac.
Sources: Relay for macOS.