How much faster is dictation than typing, after corrections

Almost every page that quotes a speed multiple for dictation is quoting the same study, usually without naming it. It is Sherry Ruan, Jacob Wobbrock, Kenny Liou, Andrew Ng and James Landay, published in the Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies in December 2017. Thirty-two participants, sixteen working in English and sixteen in Mandarin. For English the result was 152.86 words per minute by speech against 52.24 by keyboard, which is where 2.93x comes from. The preprint was published under the title "Speech Is 3x Faster than Typing", and that title has done more work in the world than the paper has.

Three details about the study change what the number means, and I have not found a single page on the first page of results that mentions any of them.

The keyboard was an iPhone. Specifically the built-in iOS keyboard on an iPhone 6 Plus, at 52.24 words per minute. The study never tested a desktop keyboard, so the ratio is speech against thumb-typing on a phone. If you are sitting at a Mac with a full keyboard, the comparison you are reading about was never run.

The task was copying, not writing. Participants transcribed 120 prewritten phrases averaging 28.3 characters, with no punctuation and almost no capitalization. The authors are direct about the limits of that: they chose transcription because composition introduces confounds such as thinking time and word choice, they note that using transcription to draw conclusions about general text entry "might be suspect", and they describe their results throughout as upper-bound performance.

And the counter-number is inside the same paper. Ruan and colleagues' own literature review cites earlier work by Karat and colleagues in which users composing text by speech managed 7.8 words per minute, and transcribing by speech managed 13.6, against 32.5 words per minute for a keyboard-and-mouse method. The paper that gave the internet "3x faster" contains a citation in which speech composition was four times slower. Recognizers have improved out of all recognition since 1999, so I would not carry that 7.8 forward as a current figure. The point is the size of the gap between copying and composing, which nobody has closed by making the recognizer better, because the recognizer was never the bottleneck.

Most articles on this topic build their ratio from two figures: people speak at about 150 words per minute, and people type at about 40. Both are folklore. I went looking for the primary source behind each and neither has one.

The 150 is attributed almost everywhere to the National Center for Voice and Speech, with no paper, dataset or method named, and the pages it was taken from are no longer live. It is also contradicted by the best actual measurement of conversational English: Jiahong Yuan, Mark Liberman and Christopher Cieri, presented at Interspeech 2006, aligned 2,438 telephone conversations and found 196 words per minute counting silence and 236 with silence removed. If 150 were a measured conversational rate it would be far too low.

The 40 is worse, because there is a good study available and nobody cites it. Vivek Dhakal, Anna Feit, Per Ola Kristensson and Antti Oulasvirta collected 136 million keystrokes from 168,000 people for CHI 2018 and measured a mean of 51.56 words per minute, with the fastest tenth above 78. But that sample was recruited through a typing-test website, had a mean age of 24.5, and 72% had taken a typing course. It is a room full of people who enjoy typing tests, transcribing sentences they were given. So it is an upper bound too, in the same way the speech figure is, and the correct move is to use both as upper bounds rather than pretending one is a hard floor and the other a hard ceiling.

This matters more than it looks. A ratio built from two unsourced numbers pointing in convenient directions is not evidence, and the direction of the error is not random. Take the folklore pair and you get 3.75x. Take the two best measurements of the same quantities and you get something closer to 3x before you have subtracted a single correction.

This is the part the ranking pages skip, and it is the whole argument. Every raw words-per-minute figure prices your cleanup pass at zero. Three separate measurements say it is not zero.

The Stanford study gives the floor. Before any correction, initial speech transcription ran at 178.92 words per minute; after users fixed their errors it settled at 152.86. The authors state the ratio as 1.17, so correction took about 15% of raw throughput. Hold onto how favorable those conditions were: a quiet lab, five-word phrases, no punctuation to speak, and a recognizer being asked the easiest question in the field.

Heike Koester's 2003 study gives the ceiling. She watched 23 experienced speech-recognition users, all with at least six months of practice, complete real word-processing tasks on video. Correcting recognition errors took 37% of total task time. Each individual correction ran from 11 to 42 seconds and averaged 23. Net text entry came out at 16.7 words per minute. And the finding that should give any confident writer pause: comparing the 18 users who did both conditions, speech helped 11 of them and made the other 7 slower.

A 2020 CHI study by Foley and colleagues isolates the mechanism cleanly. Trials that contained no correction averaged 141.1 words per minute. Trials that contained one averaged 54.6. A single correction episode roughly halves your throughput for that stretch, because you stop producing, switch task, locate the error, fix it, and restart. The cost is not the keystrokes, it is the context switch.

So the honest range for correction overhead runs from 15% in the friendliest lab conditions ever measured to 37% among experienced users doing real work. Where you land is set almost entirely by your error rate, which is the subject of its own page: word error rate explained covers how the figure is calculated and why a benchmark number is not the number you will get at your desk.

Nobody publishes the calculation, so here it is. Net throughput is your raw speaking rate divided by one plus your correction overhead:

net wpm = raw wpm / (1 + correction seconds per minute of speech / 60)

Work an example with numbers you can defend. Say you dictate at 150 words per minute and your recognizer has a 5% word error rate, so 150 words of speech leaves you roughly 7 or 8 errors. At Koester's average of 23 seconds per fix that is nearly three minutes of correction per minute of speech, which is catastrophic, and it tells you immediately that 23 seconds is the wrong constant for modern conditions. Her users were correcting by voice command against an 85% accurate recognizer in 2003. Correcting a word by hand in a modern editor is a few seconds, not 23.

So take 4 seconds per fix, which is a plausible cost for noticing a wrong word and retyping it. Seven errors at 4 seconds is 28 seconds of correction per minute of speech, giving 150 / (1 + 0.47), or about 102 net words per minute. Against a 51.56 wpm typist that is 2x, not 3x. Push the error rate to 10% and you get roughly 74 net wpm, or 1.4x. Push it to 20% and dictation loses to a fast typist entirely.

The break-even is the useful output here. On these assumptions, dictation stops beating a median typist somewhere around a 30% word error rate, and stops beating a top-decile typist at around 12%. That is why accuracy is a speed feature rather than a comfort feature, and why the gap between a good model and a mediocre one shows up in your calendar rather than in your transcript.

Two honest caveats about this arithmetic. The 4 seconds is my estimate, not a measured constant, and it is the number the whole calculation swings on. And the model assumes you fix errors as you go; if you dictate the whole draft and clean up afterwards the total may be lower, because you avoid the context switch that Foley's data suggests is the expensive part.

Even if correction were free, you would not get 3x, because typing is not the only thing you do when you write. John Gould and Stephen Boies published a study in Science in 1978 comparing writing, dictating and speaking letters. Their central finding has aged better than anything else in this literature: planning takes about two-thirds of composition time, and it takes that share regardless of which method you use. They also found that novices learned to dictate at the speed and quality of their writing within a few hours, and that experienced dictators were only 0 to 25% faster than novices, which is a strong hint that practice is not the missing variable either.

Put that together with an input speedup and the result is deflating. Suppose a writing session is 100 minutes, and following Gould, 67 of those are thinking and 33 are getting words down. Make the getting-words-down part three times faster and it takes 11 minutes instead of 33. Total session: 78 minutes. That is 1.28x, from a 3x input improvement. Apply the more realistic 2x from the section above and you get 84 minutes, or 1.19x.

This single frame reconciles everything else on this page. It explains why the lab measures 3x and the clinic measures nothing much. It explains why the writers who try dictation and report it life-changing are usually the ones with the words already in their head, and why the ones who bounce off it are usually composing as they go. And it explains why short messages show a bigger multiple than long-form writing does: a text message is nearly pure transcription, so it is close to the lab condition, while an essay is mostly planning.

Clinical documentation is the one field that has measured this at scale on real work, because hospitals have both the budget and the motive. The results are much less flattering than the consumer literature, and the pages quoting 3x do not mention them.

Suzanne Blackley and colleagues ran the cleanest test in the International Journal of Medical Informatics in 2020. Ten physicians, each documenting simulated encounters twice, once dictated and once typed, in randomized order. Their finding: documentation time was similar between methods, with dictation slightly ahead, and their conclusion states plainly that whether dictation is objectively faster than typing remains unclear. The dictated notes were substantially longer, 320.6 words against 180.8, and were rated higher quality. Interestingly the typed notes carried more uncorrected errors, 2.9 per note against 1.5, mostly misspellings. Ten physicians is a small study and I would not hang a decision on it alone, but it is a direct like-for-like comparison and it did not find the multiple.

At larger scale, a 2025 cross-sectional study in JAMA Network Open looked at speech-recognition use across clinicians at a rural US health system and found each 1% rise in usage associated with a 0.25% rise in lines documented per hour. That is an association in one health system rather than a controlled speedup, and it should not be extrapolated to a headline multiple, but the direction and magnitude are consistent with everything above: a real gain, in the tens of percent, not a tripling.

The error side has its own good measurement. Li Zhou and colleagues examined 217 clinical notes from 144 clinicians across two health systems for JAMA Network Open in 2018 and counted errors at three stages. The raw speech-recognition draft had a 7.4% error rate. After a transcriptionist reviewed it, 0.4%. After the physician signed it, 0.3%. That is the clearest published picture of what cleanup buys: review removed roughly 94% of the errors. It is also the clearest picture of why draft speed and finished speed are two different measurements, since the draft in that study was never usable as it came out.

There is one number in the Stanford study that nobody quotes and that I think is the most practically useful thing in it. Of the total time in a dictation trial, only 39.2% was spent actually speaking. Another 31.5% was lost to user delay, meaning the participant had activated the microphone and was not yet talking, or had finished talking and had not yet pressed done.

Nearly a third of the interaction was dead air. And this was people reading a phrase off a card. In real writing, where you genuinely stop to think, the share is going to be higher, not lower.

That is a design problem, not a speech problem, and it is the one place where how the app is built changes your actual throughput. An always-on recognizer is listening through every one of those pauses. It pays for your hesitation twice, once in latency while it decides you have stopped, and once in whatever it transcribes from the room while you were thinking. Systems that solve this with a silence timeout, as Apple's built-in dictation does by stopping after 30 seconds of quiet, solve it by punishing exactly the pause you needed.

Push-to-talk inverts the arrangement. You hold a key while you talk and release it when you are done, so the boundary is set by your hand rather than inferred from your silence. Thinking costs nothing, because the microphone is not open while you do it. The trade is a keypress and the discipline of remembering it. I have written up the full comparison separately in push-to-talk versus always-on dictation, including where always-on genuinely wins.

This is also the honest reason the throughput literature is dated in one specific respect. Every speed figure quoted on this page was produced against a server-based recognizer, and Ruan and colleagues note in their own limitations that embedded systems have lower latency and better reliability. Eight years on, on-device Mac dictation is ordinary and the round-trip is gone, but nobody has redone the arithmetic for local inference. If anything the published figures now understate the input side, which makes the thinking-time ceiling above matter more rather than less.

The multiple is not a property of dictation. It is a property of the task you point it at, and the pattern across all the evidence above is consistent.

It is genuinely faster when:

• You already know what you want to say. Replies, notes, briefs, anything closer to transcription than composition. This is the lab condition, and the lab condition is real work for a lot of people.
• The output is narrative prose. Long unbroken stretches of ordinary sentences, where punctuation is predictable and structure is linear.
• Your hands are the constraint. If typing hurts, the comparison stops being about speed. More on that below.
• You are producing a rough first draft on purpose. Blackley's finding that dictated notes came out substantially longer is a feature if volume is what you want out of a first pass.

It is slower, or not worth it, when:

• The exact characters matter. Code, formulae, identifiers, anything where you would spend longer speaking the punctuation than typing it. The punctuation command list shows how much you have to say out loud to get a well-formed paragraph.
• The document is mostly structure. Templates, forms and tables are navigation and selection work, and dictation does not help with either.
• You are composing as you go. Gould's two-thirds is the whole reason. If you are finding the words while speaking, you are paying for dead air.
• You can be overheard, or the room is noisy. One costs you privacy, the other costs you accuracy, and accuracy costs you time.
• You are editing rather than writing. Restructuring existing text is reading and deciding, not producing.

Every page on this subject, this one included, is giving you an average from someone else's conditions. Your own figure takes about forty minutes to establish and it is the only one that should influence what you do. The trick is to measure finished words rather than produced words, because that is the quantity you actually get paid in.

1. Pick one real task you do often. Not a typing test. A genuine email, note or section of a document, of the kind you write weekly.
2. Do it typed, timed, to finished quality. Stop the clock when you would actually send it. Count the words.
3. Do a comparable one dictated, timed the same way. Include the cleanup pass in the time. If you would not send it as it stands, the clock is still running.
4. Divide finished words by total minutes for each. That ratio is your number, and it already contains your error rate, your thinking time, your accent, your room and your subject matter, none of which any study can give you.
5. Repeat once a week for three weeks. Dictation has a learning curve that typing does not, so a single trial will understate it. Gould's finding suggests the curve is shorter than people expect, but it is not zero.

If your ratio comes out near 1.0, you are probably composing rather than transcribing, and the honest conclusion is that dictation is not your bottleneck. If it comes out above 1.5, you have found real time and the only remaining question is your error rate. And if the two numbers are close but dictating leaves you less tired at the end of the day, that is a legitimate reason to switch that has nothing to do with speed.

A good share of the people who move to dictation are not chasing speed at all. They are trying to keep working with wrists, shoulders or a neck that have stopped cooperating, and for them a ratio of 1.0 would be a fine outcome. The evidence here is real, but it is more qualified than the marketing suggests, and it cuts both ways.

Birgit Juul-Kristensen and colleagues measured muscle activity directly, publishing in Ergonomics in 2004. Speech recognition reduced static muscle activity in the forearm, neck and to some extent the shoulder, with longer pauses in muscle activity in the forearm and shoulder. But it increased static activity in the voice muscles, and their own recommendation was to use speech recognition as a supplementary tool alongside a keyboard and mouse rather than as a replacement.

Two other findings belong here for balance. Elsbeth de Korte and Peter van Lingen, in Applied Ergonomics in 2006, found improved wrist, forearm, upper arm, shoulder and neck postures after a six-week training period, and also that productivity decreased for most subjects. And Olson and colleagues documented five patients with existing upper-limb strain injuries who developed muscle tension dysphonia shortly after taking up speech recognition, all of whom had normal voices in ordinary speech. Five retrospective cases is not a rate, and I am not going to present it as one, but "you can hurt your voice this way" is a real thing that exists and is absent from every enthusiastic article on the subject.

I should be equally clear about what is missing: I could not find a single randomized trial of dictation as a treatment or prevention for repetitive strain injury. The case rests on the mechanical evidence above, which is genuine, rather than on outcome trials, which do not appear to have been run. The dictation for RSI page goes into the practical setup in more detail.

I make Parakeety, a one-person Mac dictation app, so I had an obvious commercial reason to write "dictation is 3x faster" and stop there. I have argued the opposite, because the 3x claim does not survive contact with the paper it comes from, and because a reader who buys a dictation app expecting to triple their output will be disappointed by a tool that genuinely helps them. Every figure above is attributed to a named study with a date, and where I could not find a primary source, as with the ubiquitous 150 and 40 words per minute, I have said so instead of repeating it.

Three things I want to be straight about. The net-throughput arithmetic in this piece is a model, not a measurement, and it turns on a 4-second-per-correction estimate that is mine rather than published. Several of the studies are old, and 2003-era recognizers were far worse than current ones, which means the correction overhead I quote as a ceiling is almost certainly too pessimistic today. And I have not run a controlled trial of my own app against typing, so there is no Parakeety number in that table, because there is no study behind it. The self-test above is what I would do in place of trusting anyone's marketing, including mine.

If the arithmetic above persuaded you of anything, it should be that the two levers on your real speed are your error rate and your dead air, not your speaking rate. Parakeety is built around exactly those two. It runs NVIDIA's Parakeet TDT 0.6B v3 on the Apple Neural Engine, which posts a 6.34% word error rate on the Open ASR Leaderboard, and that number is a benchmark on clean read audio rather than a promise about your desk, your accent or your room. And it is push-to-talk by design: hold the section key while you speak, release when you have finished, so every pause to think costs you nothing and nothing is captured while you are silent. Everything runs on your Mac, so there is no cloud round-trip in the loop either. It needs Apple Silicon and macOS 14 or later. There is a free 7-day trial with no card, which is long enough to run the self-test three times and get your own number before you decide.

Try Parakeety free →