Updated September 2, 2026 · By Sumbat.T

Words Per Minute Speaking: What the Research Actually Measured

There is a number that appears in almost every article about speaking speed, in most presentation guides, and in the AI summary Google shows above the search results: 150 words per minute. It is repeated so consistently that it has the texture of a fact.

It is not exactly wrong. But it is not a measurement either, and the research that did measure conversational speech carefully arrived somewhere quite different. This piece sets out where each number in this topic actually comes from, which ones have a study and a sample size behind them, and which ones are folklore that has been copied between websites for long enough to look authoritative.

The practical reason to care is the gap between speaking and typing. Whatever figure you accept for speech, it is several times faster than anyone types, and that difference is the whole basis of dictation software. It is also routinely overstated, so the second half of this piece is about what the gap is really worth once composition and correction are accounted for.

What people actually measured: 196 wpm, and 236 without the pauses

The most careful measurement of conversational speaking rate in English comes from a 2006 Interspeech paper by Jiahong Yuan, Mark Liberman and Christopher Cieri at the University of Pennsylvania and the Linguistic Data Consortium. They used the version of the English Switchboard corpus corrected and aligned at ICSI, comprising 2,438 conversations, which means every word has a timestamp rather than a transcript sitting loosely over an audio file.

That alignment is what makes the study useful, because it lets you calculate speaking rate two ways and see how much the definition matters. In the authors' own words: adding up the number of words spoken by both participants and dividing by the total elapsed time of the conversations gives “an overall average rate of 196 words per minute”. Using the alignments to exclude silences and non-speech noises gives “an average net speaking rate of 236 WPM”.

The distance between those two figures is the interesting part. About 40 words a minute, roughly a fifth of the total, is silence: breathing, thinking, waiting for the other person, the small gaps that make speech sound like speech rather than a recitation. Both numbers describe the same recordings. They differ only in whether the pauses count.

The same paper found two smaller effects worth knowing. Men spoke slightly faster than women, by about 4 to 5 words per minute or roughly 2%, which is a much smaller difference than the stereotype suggests. Older speakers were slower. And the topic of conversation moved the average meaningfully: across the assigned discussion topics in the corpus, average speaking rate ranged from 152 wpm to nearly 170 wpm, so what you are talking about changes how fast you talk about it.

Where the 150 wpm figure comes from, and what it is good for

The 150 wpm number is real, in the sense that it genuinely appears where people say it does. The National Center for Voice and Speech, a respectable research organisation, states in a tutorial page that the average rate of speech for English speakers in the United States is about 150 words per minute. That is the source almost everyone is ultimately citing, whether or not they know it.

What that page does not include is a sample size, a methodology, a corpus, or a citation to a study. It is a rule of thumb offered in an educational context, and it is a perfectly sensible one. It is simply not the same kind of object as a corpus measurement of 2,438 aligned conversations, and the two get quoted interchangeably as though they were.

There is also a good reason 150 feels right to people, and it is worth saying plainly rather than treating the figure as simply wrong. Presentations are slower than conversations. When you speak to an audience you pause for emphasis, you slow down for the parts that matter, and you leave room for people to follow. Somewhere around 130 to 150 wpm is a comfortable pace for a prepared talk. The number is good advice for the situation it came from. The error is generalising it from the podium to the telephone.

One further set of numbers deserves scepticism. Search for speaking rate and you will find confident tables giving separate figures for audiobook narration, podcasting, broadcast news and auctioneering. Those tables contradict each other from site to site, and none of the ones in wide circulation cite a study. They appear to be folklore that propagates between blogs. The one solidly measured figure in that territory is reading aloud at 183 wpm, from a 2019 meta-analysis by Marc Brysbaert covering 77 studies and 5,965 participants, which is a reasonable anchor for any kind of scripted delivery.

The other half of the comparison: typing at 51.6 wpm

The typing side of this comparison is on much firmer ground, because somebody ran the largest study of the question ever attempted. In 2018, Vivek Dhakal, Anna Feit, Per Ola Kristensson and Antti Oulasvirta published “Observations on Typing from 136 Million Keystrokes” at CHI, based on 168,000 volunteers from more than 200 countries.

The headline figure is a mean of 51.56 words per minute, with a standard deviation of 20.20. The range ran from 4 wpm to 158 wpm, and the fastest participants reached around 120 wpm. As with speech, the spread is the story: the average describes almost nobody, and the difference between a slow typist and a fast one is far larger than the difference between a slow speaker and a fast one.

Two findings from that study are worth repeating because they contradict what most people assume. The first is that formal training barely mattered: participants who had taken a typing course performed very similarly to those who never had, despite the trained group using more fingers. The second is what did separate the fast group, which was rollover, overlapping one keystroke with the next. Fast typists were not moving their fingers faster so much as running them in parallel.

The mobile figure comes from a companion study, “How do People Type on Mobile Devices?”, covering 37,370 volunteers. It found an average of 36.2 wpm with 2.3% uncorrected errors, with over 74% of people typing with both thumbs, which was significantly faster than any other technique. The fastest mobile typists were aged 10 to 19, which will surprise nobody who has watched a teenager text.

So how much faster is speaking? About three to four times

Put the two sides together and the arithmetic is straightforward. Against the 196 wpm conversational figure, speech runs roughly 3.8 times faster than the average physical keyboard and about 5.4 times faster than a phone keyboard. Against the more conservative 150 wpm convention it is closer to 2.9 times.

Those are back-of-envelope comparisons between separate studies, though, and there is a better number available. In 2016 a team from Stanford, Baidu and the University of Washington ran the comparison properly: same people, same task, both input methods. Sherry Ruan, Jacob Wobbrock, Kenny Liou, Andrew Ng and James Landay had 32 participants enter short messages on an iPhone 6 Plus, by voice and by keyboard, and measured both.

There is a detail in that study which almost every article citing it gets wrong, and it is worth stating clearly. The keyboard being beaten three-to-one was a smartphone touchscreen keyboard, not a physical one, and the task was transcribing short messages rather than composing original text. The paper is frequently cited under its original working title, “Speech Is 3x Faster than Typing”, which invites people to read it as a claim about desktop typing in general. It is not that. The finding is solid and the study is good; the scope is narrower than the headline suggests.

The gap, in practice

Type at the speed you talk

BlabbyAI turns speech into text in any Windows application on Ctrl+Space, or in any browser tab through the Chrome extension. Whisper large v3 turbo, no voice training, no per-minute billing. 60 credits a week free, no card.

Add BlabbyAI to Chrome

Why dictation does not feel three times faster in practice

Anyone who has actually used dictation for real work will have noticed that it does not deliver a clean 3x. The honest explanation has two parts, and understanding them is what separates people who get value out of dictation from people who try it for a week and stop.

The first is composition. Transcribing text that already exists is a different task from working out what to say, and the gap is enormous in both modalities. A 1999 IBM study by Karat and colleagues measured keyboard users at around 33 wpm when transcribing but only 19 wpm when composing. The bottleneck moves from your fingers to your head. If you are staring at a blank document deciding what you think, dictation cannot make you think faster, and the input rate stops being the limiting factor at all.

The second is correction. Every error you fix eats into the time you saved. This is where modern speech recognition has changed the arithmetic most, because the 1999 numbers came from an era when dictation meant training a voice profile for twenty minutes and correcting constantly. Today's models need no training at all. But the principle holds: the speed you get is entry speed minus cleanup, and cleanup depends on your microphone, your accent, and how much background noise you are working in.

A note on accuracy, because the marketed numbers are optimistic

Speech recognition accuracy is usually quoted as a word error rate, and the figure that circulates is around 2%. That number is real but it comes from LibriSpeech test-clean, a benchmark of clearly read audiobook passages recorded in good conditions. It is the easy case by design.

Across the mixed datasets of the Open ASR Leaderboard, OpenAI's Whisper large-v3 averages a word error rate of 7.44. On AMI, a corpus of real meeting recordings with overlapping speakers and room noise, the same model reaches 15.95. So a single accuracy percentage tells you almost nothing without knowing what it was measured on, and anyone quoting 99% accuracy is quoting the best case.

The good news for dictation specifically is that dictation is close to the easy case. One person, speaking deliberately, into a microphone near their mouth, in a room they chose. That is much closer to the audiobook benchmark than to the meeting-room one, which is why dictation accuracy in practice tends to be good and why accuracy is rarely the thing that makes people give up.

The finding that reframes the whole question: 39 bits per second

One more study is worth knowing about, because it suggests that words per minute is a slightly odd thing to measure in the first place.

In 2019, Christophe Coupé, Yoon Mi Oh, Dan Dediu and François Pellegrino published a paper in Science Advances examining 17 languages from 9 families, using 170 native adult speakers and around 240,000 syllables. They measured two things: how fast people produce syllables, and how much information each syllable carries.

Syllable rate varied enormously, from 4.3 to 9.1 syllables per second depending on the language and speaker. But the languages spoken quickly turned out to carry less information per syllable, and the slower ones carried more. The two effects cancelled. The rate of information actually transmitted landed at roughly 39 bits per second across all 17 languages, with a standard deviation of just 5.1.

Closing the gap: dictation on Windows and in the browser

If the numbers above make the case for talking instead of typing, the practical question is what actually types the words. BlabbyAI is built for exactly the case this article describes: not transcribing a recording after the fact, but replacing the keyboard while you work.

On Windows it is a desktop app with a global shortcut. Press Ctrl+Space in any application, speak, press it again, and the text appears in whatever field the cursor is in. That works in Word, in Slack, in a browser tab, in an IDE, in a hospital records system, anywhere Windows will accept typed characters. The transcription runs on Whisper large v3 turbo, which handles punctuation and casing without being told, and there is no voice profile to train before you start.

The shortcut panel. Ctrl+Space by default, and you can bind additional keys to specific modes.

The second product is a Chrome extension, and it is not a lesser version of the first. It is the right answer for a locked-down work laptop where you cannot install desktop software, for a Chromebook, and for anyone on macOS. It does the same job inside the browser, which for a lot of people is where most of their typing happens anyway: Gmail, Google Docs, Notion, Linear, a CRM, a support queue.

The feature that matters most for the composition problem described above is custom modes. A mode is a plain-English instruction applied to what you said, so instead of transcribing you verbatim it can clean up filler, restructure a rambling thought into bullet points, or format the same spoken sentence as a formal email in one mode and a short Slack message in another. That is aimed squarely at the gap between speaking quickly and producing usable text, which is the real bottleneck rather than the raw word rate.

A custom mode is written as an instruction, not programmed as a macro.

On pricing, the free tier is 60 credits a week, which is roughly 2,000 words of transcription and enough to find out whether dictation suits how you work. Unlimited use is $8.49 a month. There is no per-minute billing, because this is a dictation tool rather than a transcription service.

What else is available, and what it costs

Dictation software is a reasonably crowded category now, and the options divide fairly cleanly by platform and by price. The free built-in option is genuinely worth trying first, because if plain transcription is all you need then you already have it.

Tool	Price	Platform	Notes
BlabbyAI	Free tier, then $8.49/mo	Windows app + Chrome extension	Whisper large v3 turbo, custom modes, Ctrl+Space
Windows Win+H	Free, built in	Windows 11 and 10	Unlimited, but no vocabulary or formatting control
Wispr Flow	Free tier, then $15/mo	macOS, Windows, mobile	Free tier caps at 2,000 words a week
Superwhisper	Free tier, Pro $8.49/mo	macOS, Windows, iOS	Can run models locally
Dragon Professional v16	Quote only, no published price	Windows	Sold through a sales conversation

Windows dictation on Win+H is free, unlimited and already installed, and for straight transcription it is decent. What it does not give you is any control over the output: no custom vocabulary for the terms you use all day, no formatting rules, and no way to transform what you said rather than simply transcribe it. That is the line between the built-in tool and a paid one, and it is worth testing before you spend anything.

Windows dictation on Win+H. Free and unlimited, but the words arrive raw.

How to measure your own rate, and what to do with it

Every number in this article is an average over other people. Your own rate is more useful, and it takes two minutes to find.

Speak for one minute about something you actually need to write, not a prepared passage, and count the words. Then type the same content for one minute and count those. The ratio between them is your personal version of the comparison this whole article is about, and it will be more informative than 196 against 51.6, because it includes your accent, your keyboard, your subject matter and your habits.

• Measure spontaneous speech, not reading. Reading aloud has its own benchmark, 183 wpm, and it is faster and steadier than working out a thought while you speak.
• Count a full minute. Speech is bursty, and a ten-second sample will flatter or punish you depending on where the pauses fall.
• Do it on real work. The rate that matters is the one you hit on the actual emails, notes and documents you produce, not on a passage chosen to be easy to say.

If the gap you find is large and most of what you write is content you already know you want to say, dictation will pay for itself quickly. If you spend most of your writing time deciding what you think, the gain will be smaller and it will come from somewhere else: getting a messy first version out of your head faster, and tidying it afterwards.

Common questions

Sources

• Yuan, J., Liberman, M. and Cieri, C. (2006). Towards an Integrated Understanding of Speaking Rate in Conversation. Interspeech 2006. The 196 and 236 wpm figures, from 2,438 word-aligned Switchboard conversations.
• Dhakal, V., Feit, A., Kristensson, P.O. and Oulasvirta, A. (2018). Observations on Typing from 136 Million Keystrokes. CHI 2018. 168,000 participants, mean 51.56 wpm.
• Palin, K., Feit, A., Kim, S., Kristensson, P.O. and Oulasvirta, A. (2019). How do People Type on Mobile Devices? MobileHCI 2019. 37,370 participants, mean 36.2 wpm.
• Ruan, S., Wobbrock, J.O., Liou, K., Ng, A. and Landay, J. (2016). Comparing Speech and Keyboard Text Entry for Short Messages in Two Languages on Touchscreen Phones. The controlled 2.93x speech-versus-keyboard comparison.
• Coupé, C., Oh, Y.M., Dediu, D. and Pellegrino, F. (2019). Different languages, similar encoding efficiency. Science Advances. 17 languages, ~39 bits per second.
• Brysbaert, M. (2019). How many words do we read per minute? A review and meta-analysis of reading rate. Journal of Memory and Language. Oral reading at 183 wpm.
• National Center for Voice and Speech, Voice Qualities tutorial. The source of the 150 wpm convention.
• OpenAI, Whisper large-v3 model card. Word error rate figures.

For more on turning the speaking-speed advantage into working software, dictation software is the category overview, Windows speech to text covers the desktop side, and the Chrome extension covers dictation in the browser.