Conference PaperPDF Available

How do People Type on Mobile Devices?: Observations from a Study with 37,000 Volunteers


Abstract and Figures

This paper presents a large-scale dataset on mobile text entry collected via a web-based transcription task performed by 37,370 volunteers. The average typing speed was 36.2 WPM with 2.3% uncorrected errors. The scale of the data enables powerful statistical analyses on the correlation between typing performance and various factors, such as demographics, finger usage, and use of intelligent text entry techniques. We report effects of age and finger usage on performance that correspond to previous studies. We also find evidence of relationships between performance and use of intelligent text entry techniques: auto-correct usage correlates positively with entry rates, whereas word prediction usage has a negative correlation. To aid further work on modeling, machine learning and design improvements in mobile text entry, we make the code and dataset openly available.
Content may be subject to copyright.
How do People Type on Mobile Devices?
Observations from a Study with 37,000 Volunteers
Kseniia Palin
Aalto University, Finland
Anna Maria Feit
ETH Zurich, Switzerland
Sunjun Kim
Aalto University, Finland
Per Ola Kristensson
University of Cambridge, UK
Antti Oulasvirta
Aalto University, Finland
This paper presents a large-scale dataset on mobile text entry
collected via a web-based transcription task performed by
37,370 volunteers. The average typing speed was 36.2 WPM
with 2.3% uncorrected errors. The scale of the data enables
powerful statistical analyses on the correlation between typ-
ing performance and various factors, such as demographics,
nger usage, and use of intelligent text entry techniques. We
report eects of age and nger usage on performance that
correspond to previous studies. We also nd evidence of re-
lationships between performance and use of intelligent text
entry techniques: auto-correct usage correlates positively
with entry rates, whereas word prediction usage has a nega-
tive correlation. To aid further work on modeling, machine
learning and design improvements in mobile text entry, we
make the code and dataset openly available.
Human-centered computing
Empirical studies in
Mobile text entry; word prediction; auto-correct
ACM Reference Format:
Kseniia Palin, Anna Maria Feit, Sunjun Kim, Per Ola Kristensson,
and Antti Oulasvirta. 2019. How do People Type on Mobile De-
vices? Observations from a Study with 37,000 Volunteers. In 21st
International Conference on Human-Computer Interaction with Mo-
bile Devices and Services (MobileHCI ’19), October 1–4, 2019, Taipei,
Taiwan. ACM, New York, NY, USA, 12 pages.
Permission to make digital or hard copies of all or part of this work for
personal or classroom use is granted without fee provided that copies are not
made or distributed for prot or commercial advantage and that copies bear
this notice and the full citation on the rst page. Copyrights for components
of this work owned by others than ACM must be honored. Abstracting with
credit is permitted. To copy otherwise, or republish, to post on servers or to
redistribute to lists, requires prior specic permission and/or a fee. Request
permissions from
MobileHCI ’19, October 1–4, 2019, Taipei, Taiwan
©2019 Association for Computing Machinery.
ACM ISBN 978-1-4503-6825-4/19/10. . . $15.00
This paper contributes to eorts in understanding typing
performance with mobile devices, a central topic in recent
HCI research (e.g. [
]). Mobile devices are
extensively used for text input, in activities such as email,
internet browsing, texting, and social media [
]. However,
mobile typing is generally slower than typing on physical
keyboards [
]. Existent literature attributes this to a num-
ber of factors (see Related Work), including the use of virtual
instead of physical buttons, the use of fewer number of n-
gers, and the absence of training regimes like the ten-nger
touch typing system. At the same time, a large number of
intelligent text entry techniques exist, the eect of which is
poorly charted beyond prototype evaluations.
We present a new large-scale dataset and rst observa-
tions of correlates of typing performance. To improve text
entry techniques, it is important to understand their eects
beyond controlled laboratory studies. While most studies in
HCI have involved a relatively low number of participants
], and often focused on prototype evaluation, we report
here results from a large-scale dataset of over 37,370 volun-
teers. Large-scale analyses of mobile interaction are relatively
rare and mostly undertaken by commercial organizations
that may keep the datasets proprietary. Such analyses can
contribute to more comprehensive statistical analyses of a
larger number of interacting variables, and serve as training
data for machine learning models. However, self-selection
bias is a real threat to generalizability of results in online
studies with volunteers or paid workers. To this end, we
report on participant demographics and perform stratied
subsampling that allows for partial bias mitigation and better
estimation of population distribution.
In this work, we rst present the data collection method
and describe the dataset. We then report on distributions of
common metrics of typing performance, including words
per minute (WPM), error rate, and keystrokes per character
(KSPC). To better understand mobile typing behavior, we
present observations on the impact of demographic factors
and typing styles on performance. In particular, we report
on previously underexamined relation with intelligent text
entry techniques (ITE), such as autocompletion, gestural
text entry, and word prediction. Our key ndings are: (1)
when compared to small-scale studies of mobile typing, per-
formance in practice seems to be higher than previously
reported; ca. 36.2 WPM on average in our study; (2) dif-
ferences in age, experience with English, and typing style
impact performance, whereas prior training of touch typ-
ing for desktop keyboards does not; and (3) intelligent text
entry methods have a varied eect: autocorrection appears
to be positively associated with performance while word
prediction negatively.
A large body of research has emerged seeking to understand
factors aecting performance and to nd techniques to assist
typing on mobile devices. In this section, we briey review
some main earlier results.
Factors Aecting Mobile Typing Performance
Many studies of typing performance on virtual keyboards
have been conducted to evaluate a new input technique or
to collect training data for models. Some studies exist that
investigate typing behavior or background factors impacting
typing performance. Generally, it is known that typing with
one-nger is slower than using two thumbs. Azenkot and
Zhai reported speeds of 50.03, 36.34, and 33.78 WPM when
entering text with two thumbs, one thumb or the index n-
ger, respectively [
]. The superiority of two-thumb input is
attributed to frequent switching between the sides of the dis-
play, which allows for preparatory movements that decrease
inter-key intervals [6, 26, 28, 33].
Errors in mobile typing are costly in comparison to phys-
ical keyboards. The lack of tactile feedback makes it hard
to recognize pointing errors even when focusing on the vir-
tual keyboard. Thus, users need to shift attention from the
keyboard to the input eld to detect the mistakes and back
to the keyboard to correct them. The more often one looks
at the input eld, the more quickly one can detect an error.
However, typing will be slower as attention is needed to
guide ngers on the display and editing the input eld is
often cumbersome. Recent work found that this may cause
adjustments in a speed-accuracy trade-o. Users may for
example slow down to minimize the risk of errors [
]. Cog-
nitive and motor impairments, such as dyslexia, tremor, or
memory dysfunction, and various eects of aging, can have
a detrimental eect on typing performance. Users adjust
their sensorimotor strategies to nd a suitable compromise
between speed and accuracy and reliance on intelligent text
entry techniques such as the word prediction list [32].
Two recent studies report mobile typing behavior in situ.
Buschek et al. [
] conducted a study of 30 people using a
customized keyboard over three weeks, which sampled and
logged the typed text in a privacy-preserving mode. They
reported an average typing speed of 32.1 WPM. 74% of typing
was done using two thumbs, only 12.7% with the right thumb
and less than 3% for all other hand postures. Twenty-seven
participants used the word prediction feature, on average for
about 1.6% of the entered words. Sixteen used the autocor-
rection features. Only 0.63% of keystrokes were performed
in the landscape mode. Komninos et al. [
] conducted a
eld study (N = 12) using a customized keyboard over 28
days. They reported on average ca. 34 keystrokes per typing
session, with 1.98 uncorrected words. If mistakes were no-
ticed, most were corrected by using 1-5 backspace keystrokes.
However, the generalisability of these observations is lim-
ited by the size and composition of the samples (e.g., low
number of participants sampled from technical elds of a
single country [7]).
Intelligent Text Entry Methods
ITE methods use statistical language models to exploit the
redundancies inherent in natural languages to improve text
entry. Such improvements can be channeled to the user in
various ways. For example, an ITE method can autocorrect
previous typing, complete on-going input or predict the next
word for the user (see [21] for a brief review).
Numerous ITE methods have been presented in the litera-
ture and are implemented in commercial keyboards Many
aim at improving input accuracy, and thus speed, for ex-
ample by correcting touch points [
], resizing key tar-
gets [
], creating personalized touch models [
taking into account individual hand postures and nger us-
age [
], or by adapting to walking speed [
Statistical decoding to auto-correct users’ typing has been
demonstrated to be quite powerful, such as in the context of
smart watch typing [39].
However, the ecacy of word prediction is unclear for
mainstream mobile text entry, For example, the user has to
switch attention from the keyboard and the typed text to the
word prediction list. Usefulness is therefore determined in a
complex interplay of many factors, including the eciency
of the used text input method, the experience of the user, the
accuracy of the prediction and other factors. Accordingly,
results reported in the literature have been mixed [
In particular, for typing on mobile keyboards, a recent study
showed that the use of word prediction methods can be
detrimental to performance [29].
Gesture keyboard entry (originally called SHARK or Shape
writing) [
], where users continuously draw from
one letter of a word to another, permits the use of gestures
that are argued to evolve with repetition into fast-to-execute
open-loop motor programs. When assessed outside the lab,
people performed almost 10 WPM faster after practice than
using tapping-based input [
]. While gesture keyboard en-
try is a mainstream text input method, some research indi-
cates that it does not experience frequent use in practice [
Many of these techniques have been tested in controlled
laboratory evaluations with small sample sizes. The gener-
alizability of these benets to a broader population is less
understood. Our dataset presented gives insights into the
use of ITE methods across a broad population. Moreover, it
can serve as training data to improve these methods and to
develop new techniques.
Methods for Studying Mobile Typing
Our methodology closely follows prior work on large-scale
online studies of desktop typing [
]. To assess a user’s typ-
ing speed, we use transcription typing, a common task to
study motor performance that excludes cognitive aspects
related to the process of text generation. Using an online
typing test allows us to reach a larger and more diverse set
of people [
] than would be possible in a lab study. In com-
parison to desktop settings, mobile typing studies have also
been frequently conducted outside the lab (e.g. [
This often allows observation of more realistic behavior of
users and thus yields dierent insights in comparison to lab
studies (e.g. as discussed in [
]). Using an online platform
allows us to reach a much larger number of participants
in-situ. Similar approaches have been used for example to
collect large amount of training data for creating touch mod-
els [
]. In comparison to most prior work discussed above,
we do not require users to install a dedicated app. Instead,
we oer an online typing test for users to assess their typing
performance using any keyboard they are comfortable with
and any ITE method they are used to. This also allows us to
reach an even larger and more diverse sample of participants
than previous in-situ studies.
Data were collected in a web-based transcription task hosted
on a university server. A web-based method, as opposed to
a laboratory- or app-based data collection, permits a larger
sample and broader coverage of dierent mobile devices, but
comes with the caveats of self-selection and compromised
logging accuracy (see below). Still, the typing test setting
imposes a more controlled environment than an in-the-wild
study. Our test supports the main mobile operating systems
and browsers and was available globally on the Internet in
a collaboration with a Web company oering typing test-
ing and training. In the design of the software, we directly
built on work by Dhakal et al. [
] who studied transcription
typing on physical keyboards: (1) we used the same phrase
set representative of the English language; (2) we updated
performance feedback only after users committed a phrase;
Table 1: Summary of demographics and typing-related back-
ground factors in the full sample and the U.S. subsample af-
ter pre-processing. SD shown in brackets.
Full sample U.S. subsample
Factor Result Remark Result Remark
Gender 65/31 % f/m 4% n/a 51/49 % f/m
Age 24.1 (8.8) 75% 28 yrs. 25.7 (12.3) 75% 32 yrs.
Countries 163 47% U.S. 1 100% U.S.
En native speakers 68% 88.2%
Typing course 31.4% Desktop 40.1% Desktop
H/day typing 6.5 (6.2 ) on mobile 5.6 (5.9 ) on mobile
Detected OS 51/49 % Android / iOS 53/47% Android / iOS
and (3) we chose well-understood performance metrics cov-
ering speed and errors. However, we needed to adapt the
software to support mobile devices, making it responsive to
dierent screen sizes, changing the logging and updating
the database structure. Also, we added questions regarding
the keyboards used, people’s typing posture, and the use of
ITE methods. In our analysis, we perform subsampling to
mitigate the self-selection bias.
Our participants volunteered via a public website
for train-
ing and testing of typing skills. HTML requests to the site
that originated from devices detected as mobile (screen width
), were redirected to our test. The data were col-
lected between September 2018 and January 2019.
Table 1 summarizes the demographic background of the
37,370 voluntary participants left in the database after pre-
processing (see below). Similar to the general user-base of
the company hosting the webpage (see [
]), the test was
completed by more females than males, the majority of which
came from the U.S. and were mostly experienced in typing
in English (native - 56%, always - 21%, usually - 12%). The
majority reported entering text using two thumbs (74%) and
using the QWERTY layout (87%). Most did not use third-
party keyboard apps (79%). Android and iOS devices were
used almost equal to Mobile Safari (43%) and Chrome Mobile
(38%) browsers used most often.
U.S. subsample (N = 1475). In the rest of the paper, we re-
port comparative statistics from a stratied subsample that
better matches the general population of the U.S., the best-
represented country in our sample. We randomly selected
U.S. participants to match the distributions of gender [
age groups [
], and mobile operating systems [
], resulting
in a subsample of 1475 participants. See Table 1 for details.
Figure 1: The web-based transcription task. One sentence
was presented at a time with the progress shown at the top.
Task and Procedure
We followed the same procedure as Dhakal et al. [
]. The
task was to transcribe 15 English sentences, shown one after
another. Participants were shown instructions requesting
they rst read and memorize the sentence, then type it as
quickly and accurately as possible. Breaks could be taken be-
tween the sentences. After acknowledging that they had read
the instructions and giving their consent for data collection,
the rst sentence was displayed. Upon pressing Next or the
Enter key, the user’s progress, their speed and error rate were
updated and the next sentence was shown. The sentence was
visible at all times. The user interface is shown in Figure 1.
When all sentences had been transcribed, participants were
asked to ll in a questionnaire before they were shown their
nal result. In addition to the questions related to demograph-
ics and typing experience asked by Dhakal et al. [
], we also
asked for their typing posture (1- or 2-hand, index nger(s),
thumb(s) or other) the keyboard app and layout they used,
and whether they used autocorrection, word prediction, or
gesture typing (see below). Then, performance results were
shown as a histogram over all participants with details on the
fastest/slowest and most error-prone sentences (see [
] for
details). Finally, participants were oered to transcribe more
sentences to improve the performance assessment, which
we did not include in the following analysis.
We used the same sentences as [
], drawn randomly from a
set of 1,525 sentences, composed of the Enron mobile email
corpus (memorable set from [
], 400 sentences) and English
Gigaword Newswire corpus. The former one is represen-
tative of the language people use when typing on mobile
devices but too small to be used alone. The latter one com-
plements the set with more diverse sentences with a higher
Out-Of-Vocabulary (OOV) rate (0.8% versus 2.2% [37]). Mo-
bile text entry can exhibit much higher OOV rates (e.g. >20%
on Twitter [
]) with respect to a general text corpus. How-
ever, modern ITE methods adapt to users’ vocabularies. We
thus assume that the low OOV rates of our sentences are
representative of mobile text input in practice.
We implemented the front-end of our typing test using HTML,
CSS, and JavaScript. The back-end was implemented in Scala
via the Play framework, using a MySQL database for storing
the timestamp, key characteristics, and state of the input eld
at every key press, as well as meta-data of the participant
and session. The data were stored on the same university
owned server where the application was hosted on.
Limitations of web-based logging. In contrast to typing on
Desktop keyboards [
] which redirect raw device-level
events to the input eld, the access privileges of web ap-
plications are limited on most mobile devices. Similar to
other online transcription tests [
], our browser-side logging
has the following limitations: (1) the keycode is reported as
for some Android devices
; (2) for many devices
touch-down and -up events are generated together at the
moment of touch-up, resulting in a keystroke duration of
<10 ms; (3) in the case of multi-touch / rollover [
], the
events are not transmitted correctly: the key-up event of the
rst keystroke is dispatched as soon as the second nger
touches the screen, although the rst key is still pressed
down. As a result, the keycode of pressed key was often not
available and the accuracy of timestamps was poor. To en-
sure a consistent analysis, we did not analyze metrics related
to the timing of individual keystrokes, such as inter-key in-
tervals, keystroke durations, and rollover ratio [
]. Similar
to prior work (e.g. the WebTEM application [
]), we inferred
the pressed key from changes in the text on input eld.
The collected dataset was preprocessed to remove incom-
plete, inaccurate, or corrupted items. We included only par-
ticipants who nished 15 sentences and completed the ques-
tionnaire. This only included 19% of the over 260,000 people
that started the typing test. Such high dropout rates are com-
mon in online studies [
]. Of these, we conservatively
excluded about 25% of participants who did not use a mobile
device, who reported to be younger than than ve or older
than 61 years (more than 2 SD away from mean age), whose
2test e.g. at
average typing speed was over 200 WPM, who left more than
25% uncorrected errors, or who took long breaks within a
sentence (inter-key interval >5s). This yielded a dataset of
37,370 participants typing 15 sentences each.
Analyzed Metrics
We followed the de facto standard denition of performance
metrics in text entry research [
], where some of the follow-
ing metrics were already computed during runtime. Further
analysis was conducted using these measures, computed per
sentence and then averaged for each user:
Words per minute (WPM). Computed as the length of the
input (one word dened as ve characters) divided by the
time between the rst and the last keystroke.
Keystrokes per character (KSPC). The number of input events
(including non-scribed key presses) divided by the number
of characters in the produced string.
Uncorrected error rate (ER). Calculated as the Levenshtein
edit distance [
] between the presented and transcribed
string, divided by the larger size of the two strings. The un-
corrected errors were further classied into insertion,omis-
sion, and substitution errors as suggested by MacKenzie and
Soukore [25] using the TextTest tool [42].
Backspaces (BSP). The average number of backspace presses
per sentence.
Recognition of ITE from Logs
As described above, web-based logging is limited in the infor-
mation it receives about each keystroke. As a result, we could
not reliably identify the ITE methods from the dispatched
events. Instead, we had to rely on the changes in the input
eld for inferring the use of ITE methods. Therefore, we de-
veloped a simple but eective heuristics-based recognition
scheme. It compares the state of the input eld before and
after an input event. Therefore, it uses the last character of
the input eld, the length of the text and the Levenshtein
edit distance, which captures the amount of change in the
input eld. Each input event is characterized as one of four
None: is a “normal” keystroke where no ITE method was
used. We recognize this event in the case where only a single
character was inserted.
Autocorrection (A):. is the event where the keyboard auto-
matically changes the word after the user nishes it (e.g. by
pressing space). We recognize autocorrection if the previous
input was a normal keystroke and then multiple characters
were changed while the length of the text remained about
the same.
Table 2: Confusion matrix of ITE recognition. A = Autocor-
rection; P = Prediction; G = Gesture.
Recognized as
A P G none
A32 1 0 1
P371 013
G17425 27
none 7 23 21 7022
Prediction (P):. is the event where a user nishes the cur-
rently typed word by selecting it from a word prediction list.
We recognize prediction if the previous input was a normal
keystroke, multiple characters were changed, and the length
of text increased by more than two characters.
Gesture (G):. is the event where the user continuously draws
from one letter to another to input a full word. We recognize
a gesture if a whole word is inserted after a space character
input or after a gesture.
The exact algorithm used to recognize each input event
is available at https://userinterfaces.aalto./typing37k. Note,
that the denition of these events corresponds to the interac-
tion of the user: in the case of autocorrection, the user does
not perform any additional action, while prediction requires
the user to actively shift their attention to the word predic-
tion list and select the right word, and the use of gestures
requires them to change their input actions from tapping
to swiping. From an algorithmic point of view, these events
might be entangled. For example, the keyboard might apply
autocorrection to the detection of a gesture, in which case
only “Gesture” is recognized.
Empirical validation: To validate our ITE recognition, we
collected ground truth data from fteen volunteers who did
the typing test in our lab using their mobile device. Each ITE
method was used by at least ve participants. We externally
recorded the device’s screen and manually labeled the ITE
input they used for each keypress.
Like this, we collected 7,654 manually labeled input events:
34 autocorrections, 87 predictions, 460 gestures and 7,073
none-ITE inputs. We labeled the events with our recognition
algorithm; Table 2 shows the confusion matrix. Overall, the
algorithm recognized 90.9% of ITE events correctly (=9.1%
false-negative rate), with a low false-positive rate of only
0.7% (none-ITE events recognized as any of the ITE methods).
We report on indicators of typing performance and analyze
how demographic factors, typing behavior, and use of ITE
are associated with performance. Since most of our data are
not normally distributed, we used the Mann-Whitney U test
Figure 2: Histogram and density estimate of WPM (left), uncorrected error rate (middle), and KSPC (right) over all data and
the U.S. subsample.
to test dierences between distributions. In the case of more
than two groups, we rst used a Kruskal-Wallis test; if a
signicant dierence was found we performed follow up
pairwise comparisons using the Mann-Whitney U test with
Holm-Bonferroni correction. Eect sizes are reported using
Cohen’s d.
Performance Measures
An overview of the performance measures overall data in
comparison to the U.S. subsample is given in Table 3.
Words per minute. On average, participants typed at 36.17 WPM
SD =
22) with 75% of participants having a performance
below 43.98 WPM. The fastest typists reached over 80 WPM.
Figure 2 shows the distribution over all participants. It has
skewness of 0.72 and kurtosis of 1.13. The average WPM
of participants in the U.S. subsample is similar, 35.99 WPM
SD =
15). The distribution shown in Figure 2 is slightly
dierent, with a skewness of 1.08 and a kurtosis of 4.14.
Uncorrected error rate. On average, participants left 2.34%
(SD=2.08) of errors uncorrected; 75% of participants left less
than 3.07%. The skewness of the distribution shown in Fig-
ure 2 is 2.54, kurtosis is 12.12. The distribution of the U.S.
subsample is similar to an average uncorrected error rate
of 2.25% (SD=2.04). The skewness of the distribution is 3.02,
kurtosis is 18.65. The uncorrected error consisted of 11.1% in-
sertion errors, 55.6% substitution errors, and 33.3% omission
errors. Substitution was the most salient error type, which is
in line with a study of text entry on physical keyboards [
Keystrokes per character. The average KSPC value for partici-
pants is 1.18 (
SD =
18), similar to that of the U.S. subsample
SD =
2). In both, 75% of participants made less
than 1.28 keystrokes per character. Figure 2 shows a similar
distribution for both groups.
Backspaces. On average, participants performed 1.89 back-
spaces per entered sentence, with a large standard deviation
of 1.96. Participants in the U.S. subsample performed fewer
corrections, a statistically signicant dierence, but with a
small eect size.
Discussion. Typing performance in our sample is relatively
high in comparison to prior smaller-sample studies that re-
quired participants to use a dedicated app and keyboard.
They reported input rates of 32 WPM [
] and 31 WPM [
The large sample allows us to make a statistically reliable
comparison of typing on mobile soft keyboards versus phys-
ical desktop keyboards, which were studied with the same
method by Dhakal et al. [
]. Average performance is about
15 WPM slower in mobile typing. Participants left more er-
rors uncorrected on mobile devices (2.34% versus 1.17% [
Accordingly, the amount of backspacing is also lower (1.89
versus 2.29 on average). A possible explanation is the higher
interaction cost of correcting mistakes on mobile devices
and the limited text editing methods (see [
] for a discus-
sion). Nevertheless, KSPC is remarkably similar (
SD =
09 in [
]) with only the standard deviation being
smaller compared with desktop keyboard entry potentially
due to the varying use of intelligent text entry methods on
mobile devices (see below).
Note that we did not report corrected errors here. There
is no standard metric that allows to include ITE methods
because ITE input breaks the assumption of keystroke level
analysis. We call for future work to develop metrics that take
this into account, as this is beyond the scope of this paper.
However, BSP is a related metric to corrected error.
Table 3: Typing performance of the participants, for the full
sample and in the U.S. subsample.
Overall U.S. subsample Statistics
XσXσp d Sign.
WPM 36.17 (13.22) 35.99 (14.15) .045 .04
ER 2.34 (2.08) 2.25 (2.04) .716
KSPC 1.18 (0.18) 1.17 (0.20) .061
BSP 1.89 (1.96) 1.70 (1.84) .002 .09 ∗∗
∗∗∗ p<0.001,∗∗ p<0.01,p<0.05,d: Cohen’s d value
Table 4: Overview of WPM for dierent demographic factors
overall and for the U.S. subsample. Signicance is indicated
for dierences within a demographic factor. For factors with
more than two groups, detailed statistics of pairwise com-
parisons are given in the supplementary material.
WPM all WPM U.S.
XσStat. XσStat.
Gender ∗∗ ∗∗∗
female 36.0 (12.5) p=.002 34.6 (12.3) p<.001
male 36.1 (14.4) d=.003 37.4 (15.7) d=.22
Age ∗∗∗ ∗∗∗
10-19 39.6 (14.3) p<.001 38.0 (14.8) p<.001
20-29 36.5 (12.6) 39.0 (13.3)
30-39 32.2 (10.8) 34.3 (13.8)
40-49 28.9 (9.2) 27.3 (9.0)
50-59 26.3 (9.9) 24.8 (9.1)
Language use ∗∗∗
native 37.8 (13.6) p<.001 36.5 (14.5) p=.03
always 35.9 (13.1) 35.7 (12.7)
usually 34.5 (11.8) 33.5 (14.1)
sometimes 30.4 (10.5) 28.8 (12.2)
rarely 29.6 (11.2) 20.5 (9.4)
never 25.6 (12.4) 26.4 (1.2)
Training ∗∗∗
no 36.4 (13.1) p<.001 36.0 (14.1) p=.026
yes 35.7 (13.4) d=.05 35.9 (14.2) d=.002
Fingers ∗∗∗ ∗∗∗
2 37.7 (13.2) p<.001 37.9 (13.8) p<.001
1 29.2 (10.7) d=.66 28.6 (12.9) d=.65
Posture ∗∗∗ ∗∗∗
both, thumbs 38.0 (13.1) p<.001 38.2 (13.7) p<.001
both, index 32.6 (12.7) 32.7 (13.3)
right, thumb 30.2 (10.5) 30.1 (10.7)
right, index 26.7 (9.7) 25.4 (9.8)
le, thumb 30.8 (12.4) 28.9 (11.1)
le, index 25.0 (11.4) 21.8 (4.2)
∗∗∗ p<0.001,∗∗ p<0.01,p<0.05,d: Cohen’s d value
Demographic Factors
We analyze the dierences in WPM between dierent pop-
ulation groups categorized by gender, age, use of language,
typing training, and nger usage. Table 4 summarizes the
results. Details of all the statistical tests for all pair-wise com-
parisons (Holm-Bonferroni corrected) are available on our
project page userinterfaces.aalto./typing37k.
Gender: Average performance of men and women was simi-
lar. They both typed at about 36 WPM with only the SD of
WPM being smaller for female typists. Note that this analysis
excludes 4.6% of participants who did not report their gender.
Figure 3: Performance (gray) and time spent typing on mo-
bile devices (white) for dierent age groups with 95% con-
dence intervals and percentage of participants in each
Age: Participants’ performance vary with age groups, as
shown in Figure 3. Dierences were signicant for all groups
001). Participants of age between 10 and 19 typed
the fastest (
SD =
3), participants of age below
10 the slowest (
SD =
2). Interestingly, par-
ticipants aged 10–19 were not the ones who reported the
most time spent typing, as shown in Figure 3). Note that this
analysis excludes 0.2% of participants older than 60.
English experience: Native users of English were the fastest
typists (
SD =
6); those who never type in Eng-
lish the slowest (
SD =
4). Figure 4 shows how
the the typing speed decreased with reported level of experi-
ence (adj.
001 for all comparisons except sometimes
versus rarely, adj. p=0.0173).
Typing training: Surprisingly, users who reported to have
taken a touch typing course for typing on desktop keyboards
were slightly slower (
SD =
4) than those who
reported to not have taken such a course (
SD =
1). Note that this dierence was signicant, albeit with a
small eect size (adj. p<0.001,d=0.002).
Finger usage: Participants who reported to use two ngers
were signicantly faster than those who used only one nger
SD =
SD =
66). A closer look at the reported typing posture shows
that the use of dierent hands and ngers had a signicant
impact on performance. Over 82% of participants typed using
two thumbs. Conrming the ndings of prior work [
this was the fastest way to enter text (
SD =
001 in comparison to all other groups). Those who
typed with the index nger of the left hand were the slowest
SD =
4, adj.
001 in comparison to
all other groups but right, index, adj.
014). Figure 4
shows the performance and frequency for all typing postures.
Note, that this analysis excludes a small percentage of people
(<1%) who reported to use the middle nger(s).
Figure 4: Typing speed versus use of the English language
and posture in daily typing with 95% condence intervals
and percentage of participants in each group.
Discussion. The observed performance dierences are large
with over 12 WPM dierence between native English speak-
ers and those that never type in English. This holds important
implications for text entry studies which are often performed
with non-natives typing in English, but disregarding the lan-
guage experience in the analysis. From the results above,
we also conclude that a touch typing course on physical
keyboards has no recognizable association with typing per-
formance on mobile phones operated with only 1 or 2 ngers.
Intelligent text entry
Based on our ITE recognition method, we classied each
input event into prediction, autocorrection, gesture, or none.
We here analyze the actual use of ITE in practice and its cor-
relation with typing behavior. For each ITE, we computed the
percentage of words entered using the ITE per participant.
Usage and performance. We found 13.9% of participants did
not use any ITE method. More than half of the participants
used a mix of ITEs. Exact numbers are given in Figure 5. On
average, across participants who used any of the ITEs, 8%
of words were automatically corrected, 10% of words were
selected from the prediction list, and 22% of words were
entered using a gesture.
Impact of ITE on performance. The use of dierent ITE meth-
ods is associated with WPM and other performance mea-
sures. Figure 5 compares the WPM between each group.
Participants that used autocorrection were faster (
SD =
4) than all other participants (
001 for all
comparisons with other groups, using Holm-Bonferroni cor-
rection). Participants using prediction only or in combina-
tion with gestures were the slowest, with 10 WPM less
than those using autocorrection. Pairwise-comparisons us-
ing Holm-Bonferroni correction showed signicant dier-
ences between participants using no ITE and those using
prediction (adj.
p< .
15), a mix of prediction
Figure 5: ITE method versus typing speed with 95% con-
dence intervals and percentage of participants in each group.
P = Prediction; A = Autocorrection; G = Gesture.
and gestures (adj.
p< .
07), or all ITEs (adj.
49). These dierences are less pronounced in
the U.S. subsample where the dierence between autocor-
rection and normal typing was not found to be signicant,
nor were the dierence between prediction and normal typ-
ing. The detailed statistics can be found on the project page
userinterfaces.aalto./typing37k. Exact numbers of perfor-
mance are given in Table 5.
A correlation analysis between the dierent ITE methods
and performance metrics further conrms this observation.
As shown in Table 6, autocorrection has a moderate positive
correlation with WPM (
237). This is plotted in Figure 6c.
Conversely, word prediction has a small negative correlation
with performance (
183), as shown in Figure 6b. Our
correlation analysis shows that the use of ITE aects KSPC.
Using gestures and word prediction reduces the amount
of keystrokes (
251 and
232, respectively).
In contrast, autocorrection has a positive correlation with
KSPC, indicating an increase in keystrokes. Similar eects
are observed for the U.S. subsample, as shown in Table 6.
Table 5: Performance measures for each group of intelli-
gent text entry methods and their combinations, overall
and in the U.S. subsample. P=Prediction; A=Autocorrection;
Overall participants U.S. subsample
none 34.8 (12.6) 2.3 1.2 42.6 (14.7) 2.2 1.2
P 32.8 (12.1) 2.3 1.2 35.4 (15.1) 2.2 1.1
A 43.4 (14.4) 2.4 1.2 46.1 (14.4) 2.3 1.2
G 32.2 (13.4) 2.4 1.0 38.8 (12.5) 2.1 0.8
P+A 35.7 (12.6) 2.4 1.2 37.3 (12.8) 2.4 1.2
P+G 31.5 (13.6) 2.2 0.9 37.5 (19.3) 2.1 0.7
A+G 33.8 (12.1) 2.4 1.1 33.3 (10.9) 2.6 1.2
P+A+G 28.8 (11.3) 2.4 1.1 30.9 (13.4) 2.1 1.1
Figure 6: (a) Regression analysis for the relation between error rate and performance for dierent groups of ITE methods.
The bands denote 95% condence interval. Note that the Gesture group is not shown because of small sample size (<100). (b, c)
Typing performance in relation to the use of intelligent text entry methods. There is (b) a negative correlation with percentage
of words typed using prediction, and (c) a positive correlation with percentage of autocorrected words.
Impact of ITE on error. To analyze the eect of ITE methods
on error and how it changed performance, we performed
a regression analysis for each ITE condition as shown in
Figure 6a. For participants using no ITE, we found a weak
negative correlation between error rate and performance
16). This means, without ITE, the faster typists tend
to generate less error. This is in line with earlier ndings on
desktop typing [
]. In contrast, other groups do not show
clear trends between the error rates and performance. Almost
zero correlation were found for autocorrection (
prediction (
05) and mix of ITE (
04). Note that
gesture input was not analyzed due to its small sample size.
Discussion. In comparison to what has been reported in a
smaller-scale study that used a dedicated typing application
and smaller convenience sample [
], users of our typing test
used autocorrection and prediction for more words. This
might be due to the keyboard being more familiar to them.
Prior work has noted that autocorrection can be detri-
mental to performance because of high cost of erroneous
corrections [
]. It is interesting to see that nevertheless, par-
ticipants using autocorrection have the highest performance
in our dataset.
Table 6: Pearson correlation between performance and ITE
measures, overall and in the U.S. subsample. Gray: not sta-
tistically signicant (p>
), bold: weak or moderate cor-
relations. P=Prediction; A=Autocorrection; G=Gesture.
Overall participants U.S. subsample
Measure G A P G A P
WPM -0.003 0.237 -0.183 -0.017 0.272 -0.152
ER -0.012 0.086 -0.037 0.006 0.051 -0.005
KSPC -0.251 0.181 -0.232 -0.219 0.171 -0.283
As discussed at the beginning of the paper, prior work on
the usefulness of word prediction has presented conicting
results. The performance benet depends on many factors.
For mobile typing, a recent study showed decreased per-
formance rates for heavy use of word prediction [
]. Our
correlation analyses reveal similar trends. However, the wide
spread of data points in Figure 6 shows the need for more
detailed analyses to better understand the usefulness of ITE
in dierent contexts and for dierent users. While faster typ-
ists generally make fewer mistakes [
], we found no such
relation in the case where ITE methods are used. This indi-
cates that the use of autocorrection and prediction mitigates
the higher error rate of novice users.
We release a dataset containing typing events from over
37,000 participants. It includes all data reported on here, in-
cluding demographics, key log data, stimuli and transcribed
sentences, key press events and corresponding state of the in-
put eld. In addition, we captured each device’s screen width
and height, the device type and brand, the keyboard app as
reported by the participants, as well as the device’s orienta-
tion at every key press. The dataset and preprocessing code
are available at https://userinterfaces.aalto./typing37k.
In this work, we collected typing data from 37,370 volun-
teers using a browser-based transcription test. Previous work
on gathering typing data outside the traditional lab exper-
iment has relied on crowdsourcing [
] or custom mobile
apps [
]. In contrast to previous studies, the dataset
in this paper is on an unprecedented scale. However, this
comes with limitations. Generalizability of the sample is an
issue: our participants are likely exhibiting a self-selection
bias due to the nature of the website, which is a typing test
website. Many participants are young females from the U.S.
interested in typing. This is not representative of the general
population and might bias the data towards representing
a western, young, more technology-ane group of people.
We compared our results to a subsample that better repre-
sents U.S. demographics and could not nd any signicant
dierences for the basic performance measures. Neverthe-
less, results might not be generalizable to other user groups.
One example of likely sampling bias inuence is in the low
proportion of gesture keyboard users. Researchers interested
in using this dataset for their research should rst consider
this sampling method and its limitation.
On the other hand, previous solutions for collecting mobile
typing data typically relied on either opportunity-sampling
from a university campus population or recruiting partici-
pants from microtask markets or app markets, which also
introduces bias, though not necessarily in terms of the same
factors. A fruitful avenue of future work would be to per-
form a factor analysis and identify the dominant user factors
inuencing typing performance and typing behavior. Such
work could be used to correct sampling errors in text entry
studies regardless of the participant recruitment source.
We observed a higher text entry rate in our sample for
auto-correction and a lower entry rate for word prediction.
The ecacy of word prediction is an open research problem
as it depends on several factors. The primary one is the
accuracy of word prediction and the unaided entry rate of
the user, in other words, to which degree the user is rate-
limited. We conjecture that the relatively high entry rates
we observed overall in our sample make it challenging for
word prediction to provide a substantial performance benet
for users. These results are in line with prior lab studies on
mobile typing [29].
Note that our analysis is limited by the accuracy of the
ITE recognition. Due to the security and privacy restrictions
of mobile devices, we were often unable to log keycode in-
formation of each keypress. To detect and analyze the use of
intelligent text entry methods, we had to rely on a heuristic
recognition scheme based on changes in the input eld. To
evaluate our recognition we collected a ground-truth dataset
from video recordings of 15 participants. We found that our
technique was simple but eective, classifying over 90% of
ITE events correctly with a low false-postive rate (
However, there were a few cases where the changes in the
input eld were ambiguous and correct recognition was not
always possible. Our evaluation study showed that for such
edge cases, ITE use was not recognized resulting in false-
negatives the majority of misclassications. We argue that
our ndings on the eect of ITE on performance should
not be aected by this; if so eects should be even more
pronounced. Future work could investigate the use of more
advanced learning-based approaches to recognize ITE usage
from changes in the input eld.
We used a transcription task to assess typing performance
which requires the participant to dedicate part of their atten-
tion to the transcribed sentence in addition to the entered text
and the keyboard. It is also possible to use alternative meth-
ods, such as a composition task [
] or even object-based
methods, such as instructing users to annotate an image [
Given the high trac volume for the data tap underpinning
this work, we see promising follow-up work in both chang-
ing the nature of the task and the parameters of individual
tasks. For example, using this webpage, we could investigate
the eect of dierent composition tasks and individual task
parameters, such as the eect of diculty of a sentence set
] on transcription task performance. Such investigations
are dicult to perform using traditional text entry experi-
mental methods and we hope the data tap approach will be
inspirational for other text entry researchers.
In this paper, we have reported observations from a tran-
scription task mobile text entry study with 37,370 volunteers.
The set-up allowed us to carry out detailed statistical anal-
yses of distributions and correlates of typing performance,
including demographics, device, and technique. Due to the
size of the dataset, this paper has been able to reveal the
distributions of key text entry metrics, such as entry rate, un-
corrected error rate and keystrokes per character for both the
entire sample and a stratied subsampled dataset designed
to represent U.S. mobile text entry users. Also, we have clas-
sied the participants’ typing into four dierent technique
categories: autocorrect, word prediction, gesture keyboard,
and plain typing. This allowed us to explore the eects of
these techniques. Among other ndings, the data indicates
that autocorrect users tend to be faster while those that rely
on prediction tend to be slower. The collected dataset is
very rich. The presented analysis conrms prior ndings on
smaller more controlled studies and gives us new insights
into the complex typing behavior of people and the large
variations between them. However, more research is needed
to disentangle confounds, and to investigate other factors
and their interactions. To this end, we are releasing the code
and the dataset to assist further eorts in modeling, machine
learning and improvements of text entry methods.
This project has received funding from the European Re-
search Council (ERC) under the European Union’s Horizon
2020 research and innovation programme (grant agreement
No 637991), the ERC Grant OPTINT (StG-2016-717054) and
EPSRC (grant EP/R004471/1). The data collection was sup-
ported by Typing Master, Inc.
Denis Anson, Penni Moist, Mary Przywara, Heather Wells, Heather
Saylor, and Hantz Maxime. 2006. The Eects of Word Completion
and Word Prediction on Typing Rates Using On-Screen Keyboards.
Assistive Technology 18, 2 (sep 2006), 146–154.
Ahmed Sabbir Arif and Ali Mazalek. 2016. WebTEM: A Web Applica-
tion to Record Text Entry Metrics. In Proceedings of the 2016 ACM In-
ternational Conference on Interactive Surfaces and Spaces (ISS ’16). ACM,
New York, NY, USA, 415–420.
Shiri Azenkot and Shumin Zhai. 2012. Touch behavior with dierent
postures on soft smartphone keyboards. In Proceedings of the 14th
international conference on Human-computer interaction with mobile
devices and services - MobileHCI ’12.
Timothy Baldwin, Paul Cook, Marco Lui, Andrew MacKinlay, and Li
Wang. 2013. How noisy social media text, how dirnt social media
sources?. In Proceedings of the Sixth International Joint Conference on
Natural Language Processing. 356–364.
Nikola Banovic, Varun Rao, Abinaya Saravanan, Anind K. Dey, and
Jennifer Manko. 2017. Quantifying Aversion to Costly Typing Errors
in Expert Mobile Text Entry. In Proceedings of the 2017 CHI Conference
on Human Factors in Computing Systems (CHI ’17). ACM, New York,
NY, USA, 4229–4241.
Xiaojun Bi, Yang Li, and Shumin Zhai. 2013. FFitts Law: Modeling
Finger Touch with Fitts’ Law. In Proceedings of the SIGCHI Conference
on Human Factors in Computing Systems (CHI ’13). ACM, New York,
NY, USA, 1363–1372.
Daniel Buschek, Benjamin Bisinger, and Florian Alt. 2018. Re-
searchIME: A Mobile Keyboard Application for Studying Free Typing
Behaviour in the Wild. In Proceedings of the 2018 CHI Conference on
Human Factors in Computing Systems - CHI ’18. ACM Press, New York,
New York, USA, 1–14.
Kelly Caine. 2016. Local Standards for Sample Size at CHI. In Pro-
ceedings of the 2016 CHI Conference on Human Factors in Comput-
ing Systems (CHI ’16). ACM, New York, NY, USA, 981–992. https:
Pew Research Center. 2015. U.S. Smartphone Use in 2015. http:
// 2015/
[10] 2019. Live United States of America (USA) popula-
accessed January 30, 2019.
Vivek Dhakal, Anna Feit, Per Ola Kristensson, and Antti Oulasvirta.
2018. Observations on Typing from 136 Million Keystrokes. In Pro-
ceedings of the 2018 CHI Conference on Human Factors in Comput-
ing Systems (CHI ’18). ACM, New York, New York, USA. https:
Mark D. Dunlop, Emma Nicol, Andreas Komninos, Prima Dona, and
Naveen Durga. 2016. Measuring inviscid text entry using image de-
scription tasks. In Proceedings of the 2016 CHI Workshop on Inviscid
Text Entry and Beyond.
M Goel, A Jansen, T Mandel, S Patel, and J Wobbrock. 2013. Con-
textType: using hand posture information to improve mobile touch
screen text entry. In Proceedings of the 31st ACM international con-
ference on Human factors in computing systems CHI ’10. https:
Joshua Goodman, Gina Venolia, Keith Steury, and Chauncey Parker.
2002. Language Modeling for Soft Keyboards. In Proceedings of the 7th
International Conference on Intelligent User Interfaces (IUI ’02). ACM,
New York, NY, USA, 194–195.
Asela Gunawardana, Tim Paek, and Christopher Meek. 2010. Usability
guided key-target resizing for soft keyboards. In Proceedings of the 15th
international conference on Intelligent user interfaces - IUI ’10. ACM
Press, New York, New York, USA, 111.
Niels Henze, Enrico Rukzio, and Susanne Boll. 2012. Observational
and experimental investigation of typing behaviour using virtual key-
boards for mobile devices. In Proceedings of the 2012 ACM annual confer-
ence on Human Factors in Computing Systems - CHI ’12. ACM Press, New
York, New York, USA, 2659.
[17] 2019. United States Age structure - Demograph-
accessed January 30, 2019.
Heidi Horstmann Koester and Simon Levine. 1996. Eect of a word
prediction feature on user performance. Augmentative and Alternative
Communication 12, 3 (jan 1996), 155–168.
Andreas Komninos, Mark Dunlop, Kyriakos Katsaris, and John Garo-
falakis. 2018. A glimpse of mobile text entry errors and corrective
behaviour in the wild. In Proceedings of the 20th International Confer-
ence on Human-Computer Interaction with Mobile Devices and Services
Adjunct - MobileHCI ’18. ACM Press, New York, New York, USA, 221–
Per Ola Kristensson. 2007. Discrete and continuous shape writing for text
entry and control. Ph.D. Dissertation. Institutionen för datavetenskap.
Per Ola Kristensson. 2009. Five challenges for intelligent text entry
methods. AI Magazine 30, 4 (2009), 85.
Per Ola Kristensson and Keith Vertanen. 2012. Performance compar-
isons of phrase sets and presentation styles for text entry evaluations.
In Proceedings of the 2012 ACM international conference on Intelligent
User Interfaces. ACM, 29–32.
Per-Ola Kristensson and Shumin Zhai. 2004. SHARK2: a large vo-
cabulary shorthand writing system for pen-based computers. In Pro-
ceedings of the 17th annual ACM symposium on User interface software
and technology - UIST ’04. ACM Press, New York, New York, USA, 43.
Vladimir I Levenshtein. 1966. Binary codes capable of correcting
deletions, insertions, and reversals. In Soviet physics doklady, Vol. 10.
I Scott MacKenzie and R William Soukore. 2002. A character-level
error analysis technique for evaluating text entry methods. In Proceed-
ings of the second Nordic conference on Human-computer interaction.
ACM, 243–246.
I. Scott MacKenzie and Shawn X. Zhang. 1999. The design and eval-
uation of a high-performance soft keyboard. In Proceedings of the
SIGCHI conference on Human factors in computing systems the CHI
is the limit - CHI ’99. ACM Press, New York, New York, USA, 25–31.
Josip Musić and Roderick Murray-Smith. 2016. Nomadic Input on
Mobile Devices: The Inuence of Touch Input Technique and Walking
Speed on Performance and Oset Modeling. HumanâĂŞComputer
Interaction (2016).
Antti Oulasvirta, Anna Reichel, Wenbin Li, Yan Zhang, Myroslav
Bachynskyi, Keith Vertanen, and Per Ola Kristensson. 2013. Im-
proving two-thumb text entry on touchscreen devices. In Proceed-
ings of the SIGCHI Conference on Human Factors in Computing Sys-
tems - CHI ’13. ACM Press, New York, New York, USA, 2765. https:
Philip Quinn and Shumin Zhai. 2016. A Cost-Benet Study of Text
Entry Suggestion Interaction. In Proceedings of the 2016 CHI Conference
on Human Factors in Computing Systems - CHI ’16. ACM Press, New
York, New York, USA, 83–88.
Katharina Reinecke and Krzysztof Z. Gajos. 2015. LabintheWild: Con-
ducting Large-Scale Online Experiments With Uncompensated Sam-
ples. In Proceedings of the 18th ACM Conference on Computer Supported
Cooperative Work &#38; Social Computing (CSCW ’15). ACM, New York,
NY, USA, 1364–1378.
Shyam Reyal, Shumin Zhai, and Per Ola Kristensson. 2015. Perfor-
mance and User Experience of Touchscreen and Gesture Keyboards in
a Lab Setting and in the Wild. In Proceedings of the 33rd Annual ACM
Conference on Human Factors in Computing Systems (CHI ’15). ACM,
New York, NY, USA, 679–688.
Sayan Sarcar, Jussi PP Jokinen, Antti Oulasvirta, Zhenxin Wang, Chak-
lam Silpasuwanchai, and Xiangshi Ren. 2018. Ability-Based Optimiza-
tion of Touchscreen Interactions. IEEE Pervasive Computing 17, 1
(2018), 15–26.
I Scott MacKenzie and R William Soukore. 2002. A Model of Two-
Thumb Text Entry. Graphics Interface (2002), 117—-124. http://www.
Keith Trnka, John McCaw, Debra Yarrington, Kathleen F. McCoy, and
Christopher Pennington. 2009. User Interaction with Word Prediction.
ACM Transactions on Accessible Computing 1, 3 (feb 2009), 1–34. https:
Paul D. Varcholik, Joseph J. LaViola, and Charles E. Hughes. 2012.
Establishing a baseline for text entry for a multi-touch virtual keyboard.
International Journal of Human-Computer Studies 70, 10 (oct 2012), 657–
K. Vertanen, D. Gaines, C. Fletcher, A.M. Stanage, R. Watling, and P.O.
Kristensson. 2019. VelociWatch: designing and evaluating a virtual
keyboard for the input of challenging text. In Proceedings of the 2019
CHI Conference on Human Factors in Computing Systems - CHI ’19.
ACM Press, New York, New York, USA, to appear.
Keith Vertanen and Per Ola Kristensson. 2011. A versatile dataset for
text entry evaluations based on genuine mobile emails. In Proceedings
of the 13th International Conference on Human Computer Interaction
with Mobile Devices and Services. ACM, 295–298.
Keith Vertanen and Per Ola Kristensson. 2014. Complementing text
entry evaluations with a composition task. ACM Transactions on
Computer-Human Interaction (TOCHI) 21, 2 (2014), 8.
Keith Vertanen, Haythem Memmi, Justin Emge, Shyam Reyal, and
Per Ola Kristensson. 2015. VelociTap: Investigating fast mobile text
entry using sentence-based decoding of touchscreen keyboard input.
In Proceedings of the 33rd Annual ACM Conference on Human Factors
in Computing Systems. ACM, 659–668.
Daryl Weir, Simon Rogers, Roderick Murray-Smith, and Markus
Löchtefeld. 2012. A User-specic Machine Learning Approach for
Improving Touch Accuracy on Mobile Devices. In Proceedings of
the 25th Annual ACM Symposium on User Interface Software and
Technology (UIST ’12). ACM, New York, NY, USA, 465–476. https:
Jacob O. Wobbrock. 2007. Measures of Text Entry Performance. In
Text Entry Systems, Scott I. MacKenzie and Kumiko Tanaka-Ishii (Eds.).
Morgan Kaufmann, 47–74.
Jacob O. Wobbrock and Brad A. Myers. 2006. Analyzing the Input
Stream for Character- Level Errors in Unconstrained Text Entry Evalu-
ations. ACM Trans. Comput.-Hum. Interact. 13, 4 (Dec. 2006), 458–489.
Ying Yin, Tom Y. Ouyang, Kurt Partridge, and Shumin Zhai. 2013.
Making Touchscreen Keyboards Adaptive to Keys, Hand Postures,
and Individuals - A Hierarchical Spatial Backo Model Approach. In
Proceedings of the SIGCHI Conference on Human Factors in Computing
Systems. ACM.
Shumin Zhai and Per-Ola Kristensson. 2003. Shorthand writing on
stylus keyboard. In Proceedings of the SIGCHI conference on Human
factors in computing systems. ACM, 97–104.
... The first Transformer-based encoder [9] used for keystroke biometrics utilizes the Gated Transformer Network to generate sequence embeddings. Performance was significantly improved compared to the LSTM-based TypeNet model with a global EER (Equal Error Rate) of 6.26% and an average EER of 3.93% on the Aalto Mobile Dataset [10]. The current state-of-the-art model, the TypeFormer [11], is also based on a Gated Transformer Network, the difference being the insertion of a recurrent layer between the Transformer layers of the stepwise encoder. ...
... Our models were trained and evaluated using the Aalto University datasets. The university made public 2 typing datasets for scientific use: Aalto Desktop Dataset [13] and Aalto Mobile Dataset [10]. ...
... The use of web-based logging for data collection presents certain limitations when compared to inputting data through desktop keyboards, which directly capture raw devicelevel events and funnel them into input fields. The browser-side logging method used by Palin et al. [10] for mobile data acquisition is subject to several constraints akin to those encountered in various online transcription tests. The limitations are the following: ...
Full-text available
Keystroke authentication is a well-established biometric technique that has gained significant attention due to its non-intrusive and continuous characteristics. The method analyzes the unique typing patterns of individuals to verify their identity while interacting with the keyboard, both virtual and hardware. Current deep-learning approaches like TypeNet and TypeFormer focus on generating biometric signatures as embeddings for the entire typing sequence. The authentication process is defined using the Euclidean distances between the new typing embedding and the saved biometric signatures. This paper introduces a novel approach called DoubleStrokeNet for authenticating users through keystroke analysis using bigram embeddings. Unlike conventional methods, our model targets the temporal features of bigrams to generate user embeddings. This is achieved using a Transformer-based neural network that distinguishes between different bigrams. Furthermore, we employ self-supervised learning techniques to compute embeddings for both bigrams and users. By harnessing the power of the Transformer’s attention mechanism, the DoubleStrokeNet approach represents a significant departure from existing methods. It allows for a more precise and accurate assessment of user authenticity, specifically emphasizing the temporal characteristics and latent representations of bigrams in deriving user embeddings. Our experiments were conducted using the Aalto University keystrokes datasets, which include 136 million keystrokes from 168,000 subjects using physical keyboards and 63 million keystrokes acquired on mobile devices from 60,000 subjects. The DoubleStrokeNet outperforms the TypeNet-based authentication system using 10 enrollment typing sequences, achieving Equal Error Rate (EER) values of 0.75% and 2.35% for physical and touchscreen keyboards, respectively.
... The only information available concerns a study on typing speed, commonly considered a measure of typing ability on a keyboard. In detail, a large-scale dataset on mobile text entry collected via a web-based transcription task [32] performed by 37,370 volunteers (age range from 10 to 60 years) revealed the impact of demographic factors and typing styles on performance. Specifically, the number of written Words Per Minute (WPM) decreased with age. ...
... We reasoned, based on the parallelism with the physical world, that the recognition process should be rooted in the capacity to detect very subtle differences in the agent's typing. We expected the recognition process to increase with decreasing age, given the greater typing ability in younger subjects [32], which, according to embodied cognition, should increase the accuracy in detecting differences in the observed digital action. ...
... Unfortunately, there are currently no data to evaluate the ability to type on a touchscreen. The only information currently available concerns a study on typing speed [32], commonly considered an indication of typing ability. The results of that study showed that the number of written Words Per Minute (WPM) decreases with age, and that those who report using two thumbs are significantly faster than those who use only one finger. ...
Full-text available
Every day, we make thousands of finger movements on the touchscreen of our smartphones. The same movements might be directed at various distal goals. We can type “What is the weather in Rome?” in Google to acquire information from a weather site, or we may type it on WhatsApp to decide whether to visit Rome with a friend. In this study, we show that by watching an agent’s typing hands, an observer can infer whether the agent is typing on the smartphone to obtain information or to share it with others. The probability of answering correctly varies with age and typing style. According to embodied cognition, we propose that the recognition process relies on detecting subtle differences in the agent’s movement, a skill that grows with sensorimotor competence. We expect that this preliminary work will serve as a starting point for further research on sensorimotor representations of digital actions.
... As a built-in means of text input, a QWERTY soft keyboard has been responsible for most users' communication activities on smartphones. Palin et al. [1] reported that 87.0% of approximately 37,000 survey participants from 163 countries primarily used a QWERTY layout for text entry in 2018, and Statista [2] confirmed that more than 70.0% of smartphone users worldwide used chat, messaging, and email on their smartphones between July 2022 and June 2023. As such, a QWERTY soft keyboard has been popularly employed for text input on smartphones [1,[3][4][5], and communication using text input is considered the most common activity among smartphone users [2,6]. ...
... Palin et al. [1] reported that 87.0% of approximately 37,000 survey participants from 163 countries primarily used a QWERTY layout for text entry in 2018, and Statista [2] confirmed that more than 70.0% of smartphone users worldwide used chat, messaging, and email on their smartphones between July 2022 and June 2023. As such, a QWERTY soft keyboard has been popularly employed for text input on smartphones [1,[3][4][5], and communication using text input is considered the most common activity among smartphone users [2,6]. In general, studies [7][8][9] have predicted that the primary reason for this popularity is that smartphone users are familiar with a QWERTY layout based on their previous experiences with personal computers, and thus, they prefer a QWERTY soft keyboard as a means of text input. ...
Full-text available
This study aims to observe which thumb, the left or right, is used for keystrokes and examine the patterns during two-thumb key entry on a smartphone QWERTY soft keyboard. A total of 36 college students, including 18 left-handed and 18 right-handed, were recruited for testing, and they had 9.7 years of smartphone use experience on average. A smartphone application was implemented, and whether the left or right thumb was used for touch interactions was recorded for each of the 26 letter keys. As a result, it was found that there were slightly more letter keys that were statistically more often tapped by the left thumb during the two-thumb key entry on the QWERTY soft keyboard, regardless of the participant’s handedness. In addition, all the letter keys were touched statistically more often with the relatively closer one of both thumbs, except for the letter keys G and V in the center. It seemed that the distance between keys and thumbs was regarded as the most important factor influencing the thumb choice for keystrokes, followed by the habituated experience of using physical QWERTY keyboards.
... However, with their Bluetooth connectivity and small size, these devices also advertise their use on mobile devices such as tablets and phones. According to a similar study, users typing on a mobile screen achieved an average of 36.17 words per minute, with a high of over 80 [8]. When comparing these average words per minute to the advertised speeds from Tap and Tek Gear, both devices claim that users will be able to type within this average as seen in our graph in Fig. 1. ...
Conference Paper
Full-text available
This paper presents a study on the use of the Tap Strap 2 input device for creative applications. The Tap Strap 2 is a wearable device that allows users to input commands through a combination of finger taps. While the device has been marketed as a versatile input solution for a wide range of devices, we argue that it is not well-suited for everyday use, and that its main target audience is hobbyists and creatives. Our paper has two main goals: (1) to devise unique functionality for the Tap Strap 2 beyond that which is advertised by the manufacturer, and (2) to perform an analysis of the learning curve associated with learning the tap combinations of the Tap Strap 2. To achieve the first goal, we train a trajectory clustering machine learning (ML) model that can recognize gestures that resemble dial rotation, using trajectory data from the optical mouse sensor of the Tap Strap 2. Overall, the paper aims to demonstrate how the Tap Strap 2 can be used to expand the creative/artistic repertoire of hobbyists and creatives without the need for expensive specialized devices.
Conference Paper
Full-text available
We report on typing behaviour and performance of 168,000 volunteers in an online study. The large dataset allows detailed statistical analyses of keystroking patterns, linking them to typing performance. Besides reporting distributions and confirming some earlier findings, we report two new findings. First, letter pairs that are typed by different hands or fingers are more predictive of typing speed than, for example, letter repetitions. Second, rollover-typing, wherein the next key is pressed before the previous one is released, is sur- prisingly prevalent. Notwithstanding considerable variation in typing patterns, unsupervised clustering using normalised inter-key intervals reveals that most users can be divided into eight groups of typists that differ in performance, accuracy, hand and finger usage, and rollover. The code and dataset are released for scientific use.
Conference Paper
Full-text available
We present a data logging concept, tool, and analyses to facilitate studies of everyday mobile touch keyboard use and free typing behaviour: 1) We propose a filtering concept to log typing without recording readable text and assess reactions to filters with a survey (N=349). 2) We release an Android keyboard app and backend that implement this concept. 3) Based on a three-week field study (N=30), we present the first analyses of keyboard use and typing biometrics on such free text typing data in the wild, including speed, postures, apps, auto correction, and word suggestions. We conclude that research on mobile keyboards benefits from observing free typing beyond the lab and discuss ideas for further studies.
Conference Paper
" Virtual keyboard typing is typically aided by an auto-correct method that decodes a user's noisy taps into their intended text. This decoding process can reduce error rates and possibly increase entry rates by allowing users to type faster but less precisely. However, virtual keyboard decoders sometimes make mistakes that change a user's desired word into another. This is particularly problematic for challenging text such as proper names. We investigate whether users can guess words that are likely to cause auto-correct problems and whether users can adjust their behavior to assist the decoder. We conduct computational experiments to decide what predictions to offer in a virtual keyboard and design a smartwatch keyboard named VelociWatch. Novice users were able to use the features of VelociWatch to enter challenging text at 17 words-per-minute with a corrected error rate of 3%. Interestingly, they wrote slightly faster and just as accurately on a simpler keyboard with limited correction options. Our finding suggest users may be able to type difficult words on a smartwatch simply by tapping precisely without the use of auto-correct.
Conference Paper
Research in mobile text entry has long focused on speed and input errors during lab studies. However, little is known about how input errors emerge in real-world situations or how users deal with these. We present findings from an in-the-wild study of everyday text entry and discuss their implications for future studies.
Ability-based optimization is a computational approach for improving interface designs for users with sensorimotor and cognitive impairments. Designs are created by an optimizer, evaluated against task-specific cognitive models, and adapted to individual abilities. The approach does not necessitate extensive data collection and could be applied both automatically and manually by users, designers, or caretakers. As a first step, the authors present optimized touchscreen layouts for users with tremor and dyslexia that potentially improve text-entry speed and reduce error.
Conference Paper
Text entry is an increasingly important activity for mobile device users. As a result, increasing text entry speed of expert typists is an important design goal for physical and soft keyboards. Mathematical models that predict text entry speed can help with keyboard design and optimization. Making typing errors when entering text is inevitable. However, current models do not consider how typists themselves reduce the risk of making typing errors (and lower error frequency) by typing more slowly. We demonstrate that users respond to costly typing errors by reducing their typing speed to minimize typing errors. We present a model that estimates the effects of risk aversion to errors on typing speed. We estimate the magnitude of this speed change, and show that disregarding the adjustments to typing speed that expert typists use to reduce typing errors leads to overly optimistic estimates of maximum errorless expert typing speeds.
Conference Paper
WebTEM is a Web application to record text entry metrics. It is developed with common Web technologies, thus works on any device with a modern Web browser, and with any keyboard. We evaluated its effectiveness in an empirical study that compared the default Apple iOS, Google Android, and Microsoft Windows Phone keyboards. Results of the study highlighted mobile users' increased dependency on autocorrections.
Conference Paper
We describe the primary ways researchers can determine the size of a sample of research participants, present the benefits and drawbacks of each of those methods, and focus on improving one method that could be useful to the CHI community: local standards. To determine local standards for sample size within the CHI community, we conducted an analysis of all manuscripts published at CHI2014. We find that sample size for manuscripts published at CHI ranges from 1 -- 916,000 and the most common sample size is 12. We also find that sample size differs based on factors such as study setting and type of methodology employed. The outcome of this paper is an overview of the various ways sample size may be determined and an analysis of local standards for sample size within the CHI community. These contributions may be useful to researchers planning studies and reviewers evaluating the validity of results.