Typing Speed Statistics 2026: Average WPM, Phone vs Voice

Typing speed statistics show that most people write far slower than they talk. Volunteers typing on their own phones averaged 36.17 words per minute in an online test taken by 37,370 people (Palin et al., MobileHCI 2019, Oct 2019). Researchers estimate that people composing their own text stop being slowed by the input method only once it reaches about 67 words per minute (Kristensson and Vertanen, MobileHCI 2014, Sep 2014). And in a Stanford-led lab study, speaking produced English text 2.93 times faster than the iPhone keyboard, 153 words per minute against 52 (Ruan et al., IMWUT 2017, Dec 2017).

The figures below cover desktop and phone typing speed, what counts as a good WPM at work, speaking rates, voice versus keyboard input, error rates for typists and speech recognizers, the hours people spend typing, and the strain it puts on hands and wrists. Every number links to the study, survey or agency that produced it and carries the date of its data. The largest typing datasets come from volunteers who took online tests, so read them as what keen typists do, not as the whole population. For how people use voice assistants and voice search, see the voice search statistics hub.

Key Takeaways

Volunteers who took an online typing test averaged 51.56 words per minute on a physical keyboard, across 168,960 people and 136.9 million keystrokes (Dhakal et al., CHI 2018, Apr 2018).
The fastest 10% of keyboard typists type above about 78 WPM, and the slowest 10% below about 26 WPM (Dhakal et al., CHI 2018, Apr 2018).
Self-taught typists matched touch typists in a 30-person lab study: 58.9 WPM against 57.8 WPM, with no significant difference (Feit et al., CHI 2016, May 2016).
US federal clerk-typist jobs require 40 words per minute, on a 5-minute sample with three or fewer errors (US Office of Personnel Management, May 2022).
The 2026 world champion in keyboard text production typed 837.5 characters per minute, about 168 WPM, with a 0.088% error rate (Intersteno, Jul 2026).
Phone typists who rely on autocorrect average 43.4 WPM, against 34.8 WPM with no typing aids and 32.8 WPM with word prediction only (Palin et al., MobileHCI 2019, Oct 2019).
91% of US adults own a smartphone, up from 35% in 2011 (Pew Research Center, Nov 2025).
Recorded American phone conversations average 196 words per minute when both callers' words are counted against total call time, and 164 words per minute measured per speaker turn (Yuan et al., Interspeech 2006, Sep 2006).
Professional transcribers get 5.9% of words wrong on a standard phone-conversation benchmark, and 11.3% when friends and family talk freely (Xiong et al., Microsoft Research, Feb 2017).
Five commercial speech recognizers averaged a word error rate of 0.35 for Black speakers and 0.19 for white speakers (Koenecke et al., PNAS, Apr 2020).
Raw speech-recognition clinical notes had an error rate of 7.4%, falling to 0.4% after transcriptionist editing and 0.3% in the notes physicians signed (Zhou et al., JAMA Network Open, Jul 2018).
The average worker receives 117 emails a day, most of them skimmed in under 60 seconds (Microsoft Work Trend Index, Jun 2025).
Family physicians spend 5.9 hours of an 11.4-hour workday in the electronic health record, 86 minutes of it after clinic hours (Arndt et al., Annals of Family Medicine, Sep 2017).
About 4.8 million US workers (3.1%) had carpal tunnel syndrome in the past year, and clinicians attributed 67.1% of those cases to work (Luckhaupt et al., American Journal of Industrial Medicine, Jun 2013).

What is the average typing speed on a keyboard?

Most people never mastered touch typing, and it barely matters. The spread between fast and slow typists is far wider than the gap between trained and self-taught ones, and even a fast office typist works well below the speed of ordinary speech.

Typing speed statistics: distribution of desktop typing speed in words per minute across 168,960 online typing-test volunteers, peaking near 40 WPM with a long tail past 120 WPM, 2018 — Figure 1: Most volunteers cluster between 30 and 70 WPM, and the long right tail is where the rare 100+ WPM typists sit, which is why the average says little about any one typist. Source: Dhakal et al., CHI 2018, Observations on Typing from 136 Million Keystrokes, Apr 2018. Chart published by the source, captured 2 Oct 2026.

What is the average typing speed in words per minute?

The average in the largest open study hides a very wide range. Fast typists differ from slow ones less in how they move their fingers than in how much they overlap keypresses.

The fastest tenth of typists averaged 89.56 WPM and the slowest tenth 20.91 WPM, a gap of more than four times, and the quickest typists in the dataset passed 120 WPM (Dhakal et al., CHI 2018, Apr 2018).
The average gap between two keypresses was 238.66 milliseconds: about 120 ms for fast typists and over 480 ms for slow typists (Dhakal et al., CHI 2018, Apr 2018).
One in four keystrokes was typed with rollover, pressing the next key before releasing the last one. Most fast typists used it for 40% to 70% of keypresses, and rollover correlated with speed at r = 0.73 (Dhakal et al., CHI 2018, Apr 2018).
The volunteers were young and keen: mean age 24.5, 68% from the US, and 72% had taken a typing course (Dhakal et al., CHI 2018, Apr 2018).

Lab studies with motion capture and older typing research show the same wide range, and show that age alone does not slow a skilled typist down.

In a motion-capture lab study, 30 everyday typists aged 20 to 55 ranged from 34 to 79 WPM (Feit et al., CHI 2016, May 2016).
Typing random letter strings instead of real sentences cut typists' speed by about half, for touch and self-taught typists alike (Feit et al., CHI 2016, May 2016).
In two studies of typists who ranged from 17 to 104 net WPM and from 19 to 72 years old, older typists were slower at tapping and choice reaction time but not slower at typing, because they looked further ahead in the text (Salthouse, Journal of Experimental Psychology: General, Sep 1984).

What is a good WPM for work?

Federal office jobs set a low bar, and court reporting sets one several times higher. Court reporters write spoken words on a stenotype machine, not a standard keyboard.

Federal data transcriber jobs set the bar lower than clerk-typist roles, at 20 WPM for GS-2 duties and 25 WPM for GS-3 and GS-4 duties (US Office of Personnel Management, May 2022).
Court reporters earning the entry Registered Professional Reporter credential must pass five-minute stenotype tests at 180 WPM (literary), 200 WPM (jury charge) and 225 WPM (testimony), each with 95% accuracy (NCRA, Jan 2025).
The next tier, Registered Merit Reporter, raises those tests to 200, 240 and 260 WPM (NCRA, Jan 2025).
Certified Realtime Reporters must write live testimony at 200 WPM with 96% accuracy and no chance to edit before submitting (NCRA, Jan 2025).
NCRA's National Speed Contest dictates five-minute legs at 220 WPM (literary), 230 WPM (legal) and 280 WPM (testimony), and only transcripts scoring 95% or better count (NCRA, Mar 2026).

At the top of keyboard typing, the world championship sets both the entry bar and the ceiling.

For a valid result in the senior world-championship category, typists copy a text for 30 minutes and must average at least 360 characters per minute (72 WPM) with no more than 0.25% errors (Intersteno, Liverpool 2026 competition regulations, Jan 2026).
Competitors on chord keyboards went faster still: the 2026 chord-keyboard winner reached 1,022.13 characters per minute, about 204 WPM (Intersteno, Jul 2026).

Does touch typing make you faster?

Only a little. A typing course buys a small average gain; the clearer difference is how often each group looks down at the keys.

Typists who had taken a typing course averaged 54.35 WPM against 49.00 WPM for those who had not, a small difference (Cohen's d = 0.27) (Dhakal et al., CHI 2018, Apr 2018).
47.6% of volunteers said they type with 9 or 10 fingers. More fingers went with more speed (r = 0.38), and the fastest tenth used 8.4 fingers on average against 5.3 for the slowest (Dhakal et al., CHI 2018, Apr 2018).
Self-taught typists spent 41% of their typing time looking at the keyboard, about double the 20% for touch typists (Feit et al., CHI 2016, May 2016).
A Vanderbilt study found that typists who use the standard finger-to-key mapping were faster and more accurate than nonstandard typists, especially when the letters were removed from the keys or the keyboard was covered (Logan et al., Journal of Experimental Psychology: Human Perception and Performance, Dec 2016).

What is the average typing speed on a phone?

Phones are where most people write now, and they are measurably slower than a physical keyboard. Autocorrect is the one keyboard aid that goes with faster typing; picking words from the suggestion bar does not.

Distribution of mobile typing speeds in words per minute among 37,370 volunteers, peaking near 30 WPM and thinning out past 80 WPM, with the US subsample overlaid — Figure 2: Most phone typists land between 20 and 50 WPM, and almost nobody reaches 80; the US subsample matched to the population by gender, age and phone OS (red) follows the full sample (grey) closely. Source: Palin et al., Aalto University, How do People Type on Mobile Devices? (MobileHCI 2019), Figure 2, Oct 2019. Chart published by the source, captured 2 Oct 2026.

How fast do people type with their thumbs?

Somewhere in the 30s of WPM for most people, a little lower when typing is logged on people's own phones over weeks. Two thumbs beat one thumb or a single finger in every study that measured posture.

Three in four mobile typists in the 37,370-person study typed slower than 43.98 WPM, and the fastest passed 80 WPM (Palin et al., MobileHCI 2019, Oct 2019).
A US subsample of 1,475 participants, matched to the US population by gender, age and phone operating system, averaged 35.99 WPM (Palin et al., MobileHCI 2019, Oct 2019).
Typing on a phone is about 15 WPM slower than typing on a physical keyboard when both are measured with the same test (Palin et al., MobileHCI 2019, Oct 2019).
74% of participants said they type with two thumbs, the fastest posture at 38.0 WPM, against 30.2 WPM with the right thumb alone and 25.0 WPM with the left index finger (Palin et al., MobileHCI 2019, Oct 2019).
Teenagers aged 10 to 19 typed fastest on phones at 39.6 WPM, while people aged 50 to 59 averaged 26.3 WPM (Palin et al., MobileHCI 2019, Oct 2019).

Field studies are small, but they log real typing on people's own phones rather than a test, and they land in the same range.

Logged on their own phones over three weeks of everyday typing, 30 participants averaged 32.1 WPM (Buschek et al., CHI 2018, Apr 2018).
In the same field study, two thumbs accounted for 74.5% of posture use and ran at 36.8 WPM, against 25.3 WPM for the right thumb alone (Buschek et al., CHI 2018, Apr 2018).
Across a ten-session lab study and a four-week study on participants' own phones, people entered text at 28 to 39 WPM, with 1.0% to 3.6% of characters still wrong after correction (Reyal, Zhai and Kristensson, CHI 2015, Apr 2015).

Do autocorrect and word prediction speed you up?

Autocorrect does; word prediction mostly does not. Suggestion bars rarely offer the right word, people seldom tap them, and autocorrect users pay for their speed with more trips to the delete key.

13.9% of mobile typists used no autocorrect, prediction or gesture typing at all. Among those who did, 8% of words were autocorrected, 10% picked from the prediction bar and 22% swiped (Palin et al., MobileHCI 2019, Oct 2019).
On Google's Gboard, users tapped about 2.35% of the next-word predictions shown to them in live traffic (Hard et al., Google, Feb 2019).
In live Gboard traffic, the predictions shown in the three suggestion slots matched the word the user typed next 13.75% of the time with Google's federated neural model, up from 11.05% with the older n-gram model (Hard et al., Google, Feb 2019).
In everyday use, people picked one word suggestion for every 63 keystrokes, a median of 9.1 picks per person per day (Buschek et al., CHI 2018, Apr 2018).
Words entered from the suggestion bar went in at 75.6 WPM, an estimate of how fast people could type if the bar always offered the right word (Buschek et al., CHI 2018, Apr 2018).
8.9% of all keystrokes on the phone were presses of the delete key, and autocorrect users hit delete nearly twice as often as non-users (11% against 6% of keystrokes) (Buschek et al., CHI 2018, Apr 2018).

In one four-week study gesture typing beat tapping, but in the 37,370-person test gesture users were no faster, and few people swipe day to day.

Over four weeks on their own phones, participants reached 39.1 WPM by swiping words on a gesture keyboard, against 31.1 WPM tapping (Reyal, Zhai and Kristensson, CHI 2015, Apr 2015).
In the 37,370-person online test, typists who entered words only by gesture averaged 32.2 WPM, slower than typists who used no typing aids at all, though the gesture-only group was small (Palin et al., MobileHCI 2019, Oct 2019).
Gestures made up less than 4.5% of the text entered in a three-week field study of everyday phone typing (Buschek et al., CHI 2018, Apr 2018).

How much do people type on their phones?

A lot, and for many people the phone keyboard is the only one they have. Most of what gets typed on phones goes into chat apps.

16% of US adults are smartphone-only internet users, with a smartphone but no home broadband (Pew Research Center, Nov 2025).
27% of US adults aged 18 to 29 depend on a smartphone for internet access, against 11% of those aged 30 to 49 (Pew Research Center, Nov 2025).
Smartphone ownership falls from 97% of US adults aged 18 to 29 to 78% of those 65 and older (Pew Research Center, Nov 2025).
Over three weeks of everyday phone use, the median participant typed 20,353 keystrokes, and individual totals ranged from 1,489 to 180,948 (Buschek et al., CHI 2018, Apr 2018).
66% of the text stored in Gboard's on-device training caches came from chat apps and 16% from social apps (Hard et al., Google, Feb 2019).
Messaging apps took 82.7% of all keystrokes logged in a three-week study of everyday phone typing (Buschek et al., CHI 2018, Apr 2018).
Google's Gboard keyboard had more than 1 billion installs and supported more than 600 language varieties as of 2019 (Hard et al., Google, Feb 2019).

How fast do people speak compared with how fast they type?

People talk several times faster than they type, and the gap holds across languages with very different keyboards. The cost of voice input is the time spent fixing its errors, and in some clinical record systems that cost erased the gain.

Mean words per minute on an iPhone by input method, keyboard versus speech: English rose from 52 to 153 WPM and Mandarin from 43 to 123 WPM in a 2017 lab study — Figure 3: Speaking beat the iPhone keyboard by close to three to one in both English and Mandarin. Source: Ruan et al., IMWUT 2017, Comparing Speech and Keyboard Text Entry for Short Messages in Two Languages on Touchscreen Phones, Dec 2017. Chart published by the source, captured 2 Oct 2026.

What is the average speaking speed in words per minute?

There is no single speaking speed. Conversation, reading aloud and silent reading each run at a different rate, and the rate shifts with age, region and who you are talking to, so each figure below belongs to its own task.

Once pauses, silences and background noise are cut out, recorded American phone conversations run at a net 236 words per minute, from a low of 158 to a high of 312 (Yuan et al., Interspeech 2006, Sep 2006).
People talk about 10% faster with friends and family than with strangers: 214 words per minute in English CallHome calls between intimates against 193 in Fisher calls between strangers (Yuan et al., Interspeech 2006, Sep 2006).
Men speak slightly faster than women in conversation, but the gap is only about 4 to 5 words per minute, or 2% (Yuan et al., Interspeech 2006, Sep 2006).
In a study of 192 American English speakers aged from childhood to their early 90s, spontaneous speaking rate rose with age to a peak around 45 and slowed after that, with children and the oldest adults slowest (Jacewicz et al., JASA 2010, Aug 2010).
Region mattered more than sex: Wisconsin speakers talked at an estimated 5.21 syllables per second against 4.80 for North Carolina speakers, while men (5.09) edged out women (4.92) (Jacewicz et al., JASA 2010, Aug 2010).

Across languages, speakers trade speed for density, and reading aloud runs slower than conversation.

Across 17 languages and 170 native speakers reading the same texts, speech carried an average of 39.15 bits of information per second, whether the language is spoken fast with simple syllables or slowly with dense ones (Coupe et al., Science Advances 2019, Sep 2019).
The same 170 speakers read at an average of 6.63 syllables per second, and individual rates more than doubled from the slowest speaker (about 4.3 syllables per second) to the fastest (9.1) (Coupe et al., Science Advances 2019, Sep 2019).
Adults read English aloud at an average of 183 words per minute, based on 77 studies and 5,965 participants (Brysbaert, Journal of Memory and Language 2019, Dec 2019).
For contrast, silent reading runs at 238 words per minute for non-fiction and 260 for fiction, from a meta-analysis of 190 studies and 18,573 adult readers (Brysbaert, Journal of Memory and Language 2019, Dec 2019).

Is voice input faster than a keyboard?

On a phone, yes, and by a wide margin. Studies in virtual reality and on older mobile systems point the same way at lower absolute speeds.

The gap held in Mandarin Chinese: speech reached 123 words per minute against 43 on the Pinyin keyboard, 2.87 times faster (Ruan et al., IMWUT 2017, Dec 2017).
Before any correction, the recognizer's first transcript arrived 3.42 times faster than typing in English (about 179 words per minute) and 4.32 times faster in Mandarin, an estimate of what error-free speech would deliver (Ruan et al., IMWUT 2017, Dec 2017).
In virtual reality, speaking a sentence and fixing errors reached 28 words per minute at a 0.5% error rate, against 11 words per minute and 1.2% errors typing on a midair keyboard (Adhikary and Vertanen, IEEE TVCG 2021, May 2021).
When sentences held no unusual words such as proper names, speech with keyboard fixes rose to 36 words per minute with a 0.3% error rate (Adhikary and Vertanen, IEEE TVCG 2021, May 2021).
A decade earlier, mobile speech recognition managed only 18 words per minute seated indoors and 13 words per minute walking outdoors (Vertanen and Kristensson, IUI 2009, Feb 2009).

The Stanford-led figures come from seated university students transcribing short phrases, and the authors call them upper-bound performance. They still show the size of the gap that voice typing is built on.

Does fixing dictation errors cancel out the speed gain?

For short messages, no: correcting recognition errors cost speech some of its speed, but it stayed far ahead of the keyboard. In clinical documentation the evidence is mixed, with one set of trials finding speech slower and more error-prone and another finding similar time and fuller notes.

Fixing recognition errors cut the speed of speech input by a factor of 1.17 in English and 1.50 in Mandarin, measured as first-pass transcript speed divided by final entry speed (Ruan et al., IMWUT 2017, Dec 2017).
Users spent 8.5% of their speech-input time correcting errors, and for 86.0% of that correction time they reached for the keyboard rather than speaking again (Ruan et al., IMWUT 2017, Dec 2017).
In English, speech made fewer errors during entry than the keyboard (3.93% vs 4.72% corrected error rate) but left slightly more in the finished text (0.55% vs 0.35% uncorrected) (Ruan et al., IMWUT 2017, Dec 2017).

Inside electronic health records, results depend on the system and the task.

In a randomized test with 35 emergency department clinicians, documenting in an electronic health record by speech recognition took 18.11% longer than keyboard and mouse, and produced 138 non-typographical errors against 32 (Hodgson et al., JAMIA 2017, Jul 2017).
A replication with integration problems fixed found the same direction: complex documentation tasks took 224.4 seconds by speech against 191.9 by keyboard and mouse, 16.94% slower (Hodgson et al., Applied Clinical Informatics 2018, May 2018).
Physicians at Brigham and Women's Hospital who dictated notes wrote 320.6 words against 180.8 when typing, in about the same documentation time (Blackley et al., International Journal of Medical Informatics 2020, Sep 2020).
In the same study, typed notes carried more uncorrected errors than dictated ones (2.9 vs 1.5 per note, mostly minor misspellings), and dictated notes scored higher on quality (7.7 vs 6.6) (Blackley et al., International Journal of Medical Informatics 2020, Sep 2020).
In a survey sent to 1,731 clinicians who use speech recognition in Boston and Aurora, Colorado, 77.2% of respondents agreed it improves efficiency, yet 19.6% estimated that half or more of its errors were clinically significant (Goss et al., International Journal of Medical Informatics 2019, Oct 2019).

How accurate are typing and speech recognition?

Typists leave few errors in their final text, but they spend a lot of keystrokes getting there. Speech recognition matched professional transcribers on one benchmark years ago, and it still gets some speakers wrong far more often than others.

Average word error rate of five commercial speech recognition services (Apple, IBM, Google, Amazon, Microsoft) for Black and white speakers in recorded US interviews: every service makes more errors for Black speakers, from about 0.27 to 0.45 against 0.15 to 0.23 for white speakers, 2020 — Figure 4: All five commercial recognizers made nearly twice as many errors for Black speakers as for white speakers. Source: Koenecke et al., PNAS, Racial disparities in automated speech recognition, Apr 2020. Chart published by the source, captured 2 Oct 2026.

How many errors do typists make?

Typists fix most of their mistakes as they go, and phones let more slip through.

Volunteers who took an online typing test on a physical keyboard left an average of 1.167% of characters wrong in their final text, and 90% of them left fewer than 2.66% (Dhakal et al., CHI 2018, Apr 2018).
Getting to that clean text took an average of 2.29 corrections per sentence, 6.3% of all keypresses, and typists at the 99th percentile pressed Backspace or Delete 8.5 times per sentence. The result was 1.173 keystrokes for every character of finished text (Dhakal et al., CHI 2018, Apr 2018).
In a detailed analysis of 783 participants, substitution errors (hitting the wrong key, 1.65%) were more common than omissions (leaving a letter out, 0.8%) or insertions (adding an extra one, 0.67%) (Dhakal et al., CHI 2018, Apr 2018).
The slowest tenth of typists left 1.78% errors uncorrected, two and a half times the 0.71% left by the fastest tenth (Dhakal et al., CHI 2018, Apr 2018).

On phones, fixing a typo costs more effort, and more errors survive.

On phones, 37,370 volunteers left 2.34% of characters wrong on average, and 75% of them left less than 3.07% (Palin et al., MobileHCI 2019, Oct 2019).
Substitutions made up 55.6% of the errors phone typists left behind, omissions 33.3% and insertions 11.1% (Palin et al., MobileHCI 2019, Oct 2019).
Phone typists pressed Backspace 1.89 times per sentence, fewer than desktop typists, yet needed 1.18 keystrokes per character, about the same as on a physical keyboard (Palin et al., MobileHCI 2019, Oct 2019).

How accurate is speech-to-text in 2026?

On standard English benchmarks, the best recognizers now miss only a few words in a hundred, and fewer than one in a hundred on clean read audiobooks. Accuracy still drops on casual conversation and on languages with little training data. Error rates from different test sets are not on one scale, so each figure below names its test.

Microsoft's system edged past its own human-transcriber benchmark with error rates of 5.8% on Switchboard and 11.0% on CallHome, a result Microsoft called human parity in conversational speech recognition (Xiong et al., Microsoft Research, Feb 2017).
IBM hired its own transcribers and found the best one, after quality checks, reached 5.1% on the same Switchboard test and 6.8% on CallHome; IBM's system reached 5.5% and 10.3% (Saon et al., IBM Research, Mar 2017).
Word error rates on the Switchboard benchmark had fallen from 14% to 8.0% and then 6.6% in the few years before those parity claims (Saon et al., IBM Research, Mar 2017).
OpenAI trained its Whisper models on 680,000 hours of audio, and the zero-shot model made 55.2% fewer errors on average across other English test sets than a conventionally trained model with the same 2.7% error rate on LibriSpeech's clean audiobook test (Radford et al., OpenAI, Dec 2022).
Across languages, Whisper large-v2's word error rate on the FLEURS read-speech test ranged from 3.0% in Spanish to 156.5% in Sindhi, with Amharic at 140.3%, both languages with little training data (Radford et al., OpenAI, Dec 2022).

The freshest numbers come from Hugging Face, which runs its own evaluations of open and commercial systems.

On the Open ASR Leaderboard's ten default English test sets, which include meetings, earnings calls, podcasts, audiobooks, parliament speeches and Indian-accented conversation, the top-ranked system averaged a 4.38% word error rate, and the best score on LibriSpeech's clean audiobook test was 0.87% (Hugging Face Open ASR Leaderboard, Sep 2026).
As of March 2026, the best system on the leaderboard's multilingual track (ElevenLabs Scribe v2) averaged a 2.67% word error rate across German, French, Italian, Spanish and Portuguese; the leaderboard as a whole listed 86 systems from 26 organizations (Hugging Face Open ASR Leaderboard, paper, Mar 2026).
In 217 dictated clinical notes, 63.6% of the raw speech-recognition drafts contained at least one clinically significant error, against 7.8% of the notes physicians finally signed (Zhou et al., JAMA Network Open, Jul 2018).

Whose speech do recognizers get wrong most often?

Speakers who are underrepresented in training data. Black American speakers, non-native speakers, children and people with speech disabilities all see error rates well above average, though targeted data has cut some of those gaps by half or more.

Counting a transcript with more than half its words wrong as unusable, 23% of clips from Black speakers were unusable against 1.6% from white speakers (Koenecke et al., PNAS, Apr 2020).
The gap was widest for Black men, at an average error rate of 0.41, compared with 0.30 for Black women, 0.21 for white men and 0.17 for white women (Koenecke et al., PNAS, Apr 2020).
When Black and white speakers said the identical short phrase, every system still made about twice as many errors for Black speakers. For Microsoft's, the best performer, the error rate was 0.13 against 0.07, which points to pronunciation and prosody in the acoustic models, not vocabulary (Koenecke et al., PNAS, Apr 2020).
Whisper large-v2's error rate rose to 16.2% on CORAAL, a corpus of interviews with African American speakers, about six times its clean-audiobook score (Radford et al., OpenAI, Dec 2022).
A 2021 Dutch speech recognizer had a 26.1% average error rate on read speech from native Dutch speakers but 55.9% for non-native speakers, and native children (35.3%) fared worse than native teenagers (18.4%) (Feng et al., TU Delft, Apr 2021).

Speech disabilities show the largest gaps, and the clearest gains from targeted data.

In the Speech Accessibility Project's 2025 challenge, run on more than 400 hours of speech from over 500 people with speech disabilities such as Parkinson's disease and ALS, the winning team cut the word error rate to 8.11% from 17.82% for an off-the-shelf Whisper large-v2 baseline (Speech Accessibility Project, Interspeech 2025, Jul 2025).
For speakers with ALS, the best challenge model lowered the average speaker-level error rate from 16.29% to 5.05%, and for speakers with Parkinson's disease from 18.47% to 9.61% (Speech Accessibility Project, Interspeech 2025, Jul 2025).
A consumer speech recognizer's average word error rate was 13.6% for people with moderate stuttering and 49.2% for severe stuttering; Apple researchers' changes cut utterance cut-offs by 79.1% and lowered the error rate from 25.4% to 9.9% for participants with moderate to severe stuttering (Lea et al., Apple, CHI 2023, Apr 2023).

How much of the workday goes to typing?

Office workers and doctors spend hours a day writing. Microsoft 365 telemetry counts the messages coming in, and EHR event logs time how long doctors spend on notes, often after hours.

Microsoft 365 telemetry: the average employee receives 117 emails and 153 Teams messages each workday, 2025 — Figure 5: The average Microsoft 365 user receives 117 emails and 153 Teams messages each workday. These count incoming messages, not time spent typing replies. Source: Microsoft Work Trend Index, Breaking down the infinite workday, Jun 2025. Chart published by the source, captured 2 Oct 2026.

How many emails and messages do workers handle a day?

Hundreds of incoming items, by Microsoft's own telemetry, and the flow continues after hours.

The average worker also receives 153 Microsoft Teams messages per weekday, and messages per person rose 6% year over year worldwide (Microsoft Work Trend Index, Jun 2025).
Among the 20% of Microsoft 365 users who receive the most pings, a meeting, email or chat interrupts them every 2 minutes during core hours, or 275 times across a 24-hour day (Microsoft Work Trend Index, Jun 2025).
Workers now receive an average of 58 chat messages outside the 9-to-5 workday, and after-hours chats are up 15% year over year (Microsoft Work Trend Index, Jun 2025).
By 10 pm, 29% of workers who are still active have gone back into their email (Microsoft Work Trend Index, Jun 2025).
In 2023 data, the average Microsoft 365 user spent 57% of their app time communicating in meetings, email and chat, and 43% creating documents, spreadsheets and presentations (Microsoft Work Trend Index, May 2023).
The heaviest email users, the top 25%, spend 8.8 hours a week reading and writing email (Microsoft Work Trend Index, May 2023).

Writing is also what people most often hand to AI tools at work.

Writing is the most common work task people bring to ChatGPT: 40% of work-related messages in June 2025, and about two-thirds of writing requests ask it to edit or rework text the user already wrote (Chatterji et al., NBER Working Paper 34255, Sep 2025).
Among ChatGPT users in management and business jobs, writing makes up 52% of work-related messages; among users in other professional jobs, such as education and health care, it is 49% (Chatterji et al., NBER Working Paper 34255, Sep 2025).

How many hours do doctors spend typing notes?

Hours a day, and documentation is the largest single slice of it. Each study below measures a different thing (total EHR time, active time, time per visit, total work), so the numbers are not interchangeable.

Documentation is the single biggest EHR task for family physicians, at 84 minutes a day (23.7% of EHR time), and clerical work as a whole (documentation, order entry, billing and coding, security) takes 44.2% (Arndt et al., Annals of Family Medicine, Sep 2017).
Over the clinic day, physicians spent 27.0% of their time face to face with patients and 49.2% on EHR and desk work, or nearly 2 hours of screen and desk work for every hour of patient time (Sinsky et al., Annals of Internal Medicine, Sep 2016).
Across about 100 million outpatient visits and 155,000 US physicians, doctors spent an average of 16 minutes 14 seconds in the EHR per patient encounter, with documentation taking 24% of that time (Overhage and McCallie, Annals of Internal Medicine, Jan 2020).
US clinicians spend 90.2 minutes a day actively using the EHR against 59.1 minutes for clinicians outside the US, and 40.7 of those US minutes go to notes (Holmgren et al., JAMA Internal Medicine, Dec 2020).
A full-time primary care physician works a median of 61.8 hours a week caring for a patient panel, by EHR log and administrative data from 406 physicians at Mass General Brigham (Rotenstein et al., Annals of Internal Medicine, Oct 2025).

Ambient AI scribes, which draft notes from the visit conversation, are now being tested in randomized trials. The time savings so far are modest and vary by tool.

Before a 2025 University of Wisconsin trial, 83% of the 66 clinicians enrolled wrote their notes by manual typing, usually mixed with templates (85%) and speech-recognition software (59%) (Afshar et al., NEJM AI, Nov 2025).
In the same randomized trial, clinicians using an ambient AI scribe spent 0.36 fewer hours a day on notes, and 38% of the 71,487 notes written during the study were drafted with the tool (Afshar et al., NEJM AI, Nov 2025).
In a randomized trial of 238 UCLA outpatient physicians, one ambient AI scribe (Nabla) cut time spent writing each note by 9.5% against usual care, while the other (Microsoft DAX Copilot) made no significant difference (Lukac et al., NEJM AI, Nov 2025).
In an uncontrolled before-and-after study across six US health systems, the share of 263 clinicians reporting burnout was 51.9% before and 38.8% after 30 days with an ambient AI scribe (Olson et al., JAMA Network Open, Oct 2025).

How do students write on computers?

US students now take the national writing test on a computer, and the stronger writers use the editing tools more. Students whose teachers had them draft on computers also scored higher.

When US students wrote on computers in NAEP's first computer-based writing test, 24% of 8th and 12th graders performed at the Proficient level (National Center for Education Statistics, NAEP Writing 2011, Sep 2012).
Among eighth graders in the top quarter of NAEP writing scores, 41% pressed the backspace key more than 500 times during the test, against 5% of those in the bottom quarter (National Center for Education Statistics, NAEP Writing 2011, Sep 2012).
57% of top-quarter eighth graders right-clicked to open spell-check 1 to 10 times during the test, against 31% of bottom-quarter writers (National Center for Education Statistics, NAEP Writing 2011, Sep 2012).
44% of eighth graders had teachers who said they very often or always asked students to draft and revise on computers, and those students scored higher than students whose teachers asked less often (National Center for Education Statistics, NAEP Writing 2011, Sep 2012).
About two-thirds of eighth graders spent more than 15 minutes on a typical school day writing a paragraph or more for English class, on paper or on computer: 40% spent 15 to 30 minutes, 21% spent 30 to 60 minutes and 4% spent more than an hour (National Center for Education Statistics, NAEP Writing 2011, Sep 2012).
56% of US twelfth graders said they always or almost always use a computer to make changes to their writing, and those who did had higher writing scores (National Center for Education Statistics, NAEP Writing 2011, Sep 2012).

Can heavy typing cause pain or injury?

Hand, wrist and neck problems are common among workers, and a carpal tunnel case keeps someone off work far longer than the typical musculoskeletal case. Keyboard work is one documented risk factor among several, alongside force, posture, age and sex; the studies below show associations, not proof that typing alone causes injury.

Table comparing the share of work-related carpal tunnel syndrome cases in each US occupational group with that group's share of the workforce, 2010 National Health Interview Survey. Production (ratio 2.52) and office and administrative support (ratio 1.66) are the most overrepresented groups — Figure 6: Desk-bound office and administrative support jobs show up among work-related carpal tunnel cases far more often than their share of the workforce predicts, second only to factory production work. Source: Luckhaupt et al., American Journal of Industrial Medicine, Prevalence and Work-Relatedness of Carpal Tunnel Syndrome in the Working Population, Table IV, Jun 2013. Screenshot of the source page, captured 2 Oct 2026.

How common is carpal tunnel syndrome among workers?

Common enough to touch millions of US workers each year, and costly when it happens. The national prevalence estimate comes from the 2010 National Health Interview Survey, still the most recent one published; case counts come from employer reports.

6.7% of current or recent US workers said a clinician had diagnosed them with carpal tunnel syndrome at some point in their lives (Luckhaupt et al., American Journal of Industrial Medicine, Jun 2013).
Women workers were more than twice as likely as men to have current carpal tunnel syndrome (4.5% vs 1.9%), and workers aged 45 to 64 about five times as likely as those aged 18 to 29 (4.7% vs 0.9%) (Luckhaupt et al., American Journal of Industrial Medicine, Jun 2013).
Office and administrative support workers made up 13.4% of the workforce but 22.2% of work-related carpal tunnel cases, 1.66 times their share; among occupational groups, only production workers were more overrepresented (Luckhaupt et al., American Journal of Industrial Medicine, Jun 2013).
US private employers reported 5,140 carpal tunnel cases that caused days away from work across 2023 and 2024. The median case kept a worker off for 28 days, twice the 14-day median for all musculoskeletal disorders (US Bureau of Labor Statistics, SOII Table MSD1, Jan 2026).
Musculoskeletal disorders accounted for 484,620 private-sector cases with days away from work in 2023 and 2024 combined, an annualized rate of 22.9 cases per 10,000 full-time workers (US Bureau of Labor Statistics, SOII Table MSD1, Jan 2026).
In the BLS occupation table, which counts all carpal tunnel cases with days away from work rather than only those Table MSD1 classes as musculoskeletal disorders, office and administrative support jobs accounted for 500 of 5,210 cases in 2023 and 2024, about 1 in 10. The same jobs made up 3.3% of all private-sector days-away cases (60,930 of 1,834,600, our calculation from Table R9) (US Bureau of Labor Statistics, SOII Table R9, Jan 2026).
Repetitive microtasks, such as repeated small hand motions, caused 32,110 private-sector injuries and illnesses with days away from work across 2023 and 2024 (US Bureau of Labor Statistics, SOII Table R4, Jan 2026).

Great Britain's regulator counts the same problem by workers affected and days lost.

In Great Britain, 511,000 workers had a work-related musculoskeletal disorder in 2024/25, and in 211,000 of them the upper limbs or neck were mainly affected (Health and Safety Executive, Nov 2025).
Upper-limb and neck disorders accounted for 34% of the 7.1 million working days lost to work-related musculoskeletal disorders in Great Britain in 2024/25, an average of 11.4 days per case (Health and Safety Executive, Nov 2025).

How many computer users develop hand, wrist or neck pain?

More than half of new computer-heavy hires in the best-known prospective study reported symptoms within a year. Most symptoms were not diagnosed disorders, and the links to keyboard design and hours of keying are associations.

Among 632 newly hired workers who used a computer at least 15 hours a week, hand or arm symptoms appeared at a rate of 39 cases per 100 person-years and diagnosed hand or arm disorders at 21 per 100 person-years. The most common disorder was de Quervain's tendonitis (Gerr et al., American Journal of Industrial Medicine, Apr 2002).
Neck and shoulder problems were more common still in the same cohort: 58 symptom cases and 35 diagnosed disorders per 100 person-years (Gerr et al., American Journal of Industrial Medicine, Apr 2002).
More than 50% of the new computer users reported musculoskeletal symptoms in their first year on the job, and 46% of neck or shoulder symptoms and 32% of hand or arm symptoms began in the first month (Gerr et al., American Journal of Industrial Medicine, Apr 2002).
In the same cohort, more hours of keying per week were associated with more hand and arm symptoms and disorders, as were keyboards whose J key sat more than 3.5 cm above the desk and keys needing more than 48 g of force to press (Marcus et al., American Journal of Industrial Medicine, Apr 2002).
About three in five EU workers reported musculoskeletal complaints in 2015; backache (43%) and muscular pain in the upper limbs (41%) were the most common (EU-OSHA, Nov 2019).
In 2015, 61% of EU-28 workers were exposed to repetitive hand or arm movements for at least a quarter of their working time, and 58% worked with computers, laptops or smartphones for that long, up from 47% in 2000, according to EU-OSHA's analysis of the European Working Conditions Survey (EU-OSHA, Nov 2019).
In a five-year study of 7,092 Swedish adults aged 20 to 24, those who texted more were twice as likely to report new hand or finger symptoms after one year (odds ratio 2.0); the researchers concluded the effects were mostly short-term (Gustafsson et al., Applied Ergonomics, Jan 2017).

How many people have trouble using their hands?

Millions of US adults, most of them older. These are prevalence figures only: none of the studies below measured typing, and none of them says a writing tool changes a condition.

5.6% of US adults, about 13.4 million people, had difficulty using their fingers to grasp small objects such as a glass or pencil, and 0.5% (1.1 million) could not do it at all, in 2014 survey data (US Census Bureau, Americans With Disabilities: 2014, Nov 2018).
Difficulty grasping was twice as common among adults 65 and older (9.9%) as among adults aged 18 to 64 (4.6%) (US Census Bureau, Americans With Disabilities: 2014, Nov 2018).
Of the 8.9 million US adults aged 18 to 64 with difficulty grasping, 33.0% were employed in 2014 (US Census Bureau, Americans With Disabilities: 2014, Nov 2018).
21.2% of US adults, about 53.2 million people, reported diagnosed arthritis in 2019 to 2021, and nearly half of them (48.3%) were 65 or older (CDC, Morbidity and Mortality Weekly Report, Oct 2023).
After adjusting for age, arthritis was more common among women (20.9%) than men (16.3%), and adults aged 45 and older made up 88.3% of US adults with arthritis (CDC, Morbidity and Mortality Weekly Report, Oct 2023).
About 1.71 billion people worldwide had a musculoskeletal disorder in 2019, the largest group among conditions that can benefit from rehabilitation, in an analysis of Global Burden of Disease 2019 data (Cieza et al., The Lancet, Dec 2020).

Glossary

WPM: words per minute. Typing studies count five characters, including spaces, as one word.
CPM: characters per minute, the unit used in typing championships.
IKI: inter-key interval, the time between two keypresses.
Rollover: pressing the next key before releasing the previous one.
KSPC: keystrokes per character, how many keys are pressed for each character of finished text.
Corrected and uncorrected error rate: errors made and then fixed during typing, versus errors left in the final text.
WER: word error rate, the share of words a transcript gets wrong (substituted, deleted or inserted). It can pass 100% because inserted words count as errors.
ASR: automatic speech recognition.
Switchboard, CallHome, LibriSpeech, FLEURS, CORAAL: standard speech test sets: phone calls between strangers, calls between family and friends, read audiobooks, read speech in many languages, and interviews with African American speakers.
EHR: electronic health record.
AI scribe: software that listens to a clinical visit and drafts the note.
MSD: musculoskeletal disorder, an injury or disorder of muscles, nerves, tendons, joints or cartilage.
CTS: carpal tunnel syndrome, compression of the median nerve at the wrist.
SOII: the BLS Survey of Occupational Injuries and Illnesses.
NHIS: the National Health Interview Survey, run by the CDC's National Center for Health Statistics.
NAEP: the National Assessment of Educational Progress, the US national school assessment.
ALS: amyotrophic lateral sclerosis.
BLS, CDC, NCES, NCRA, OPM, HSE, EU-OSHA: US Bureau of Labor Statistics, Centers for Disease Control and Prevention, National Center for Education Statistics, National Court Reporters Association, US Office of Personnel Management, Great Britain's Health and Safety Executive, European Agency for Safety and Health at Work.

Methodology and Sources

Every figure on this page comes from the organization that ran the study, survey or test: peer-reviewed papers, government agencies, regulators, professional bodies and the companies that measured their own systems. Aggregators, typing-test vendor benchmarks and other statistics roundups were excluded, and per-job WPM claims without a primary source were dropped. Each URL was fetched and the number found on the page before it was kept. Foundational behavioral studies are cited with their study year because typing and speaking rates change slowly, while speech-recognition accuracy, workplace messaging and AI-scribe trials use the newest data available. The page is checked and refreshed monthly.

149 Typing Speed Statistics October 2026