Content uploaded by James R. Lewis
Author content
All content in this area was uploaded by James R. Lewis on Aug 07, 2014
Content may be subject to copyright.
INTERNATIONAL JOURNAL OF SPEECH TECHNOLOGY 7, 69–79, 2004
c
2004 Kluwer Academic Publishers. Manufactured in The Netherlands.
Models of Throughput Rates for Dictation and Voice Spelling
for Handheld Devices
PATRICK M. COMMARFORD
IBM Corporation, 8051 Congress Ave, Suite 2228, Boca Raton, FL 33487, USA
commarfo@us.ibm.com
JAMES R. LEWIS
IBM Corporation, 8051 Congress Ave, Suite 2227, Boca Raton, FL 33487, USA
Abstract. Since the emergence of the personal digital assistant (PDA), developers have attempted to create input
methods that allow users to enter accurate data at speeds that approach those achieved with the personal computer.
Common text entry methods (handwriting and soft keyboard) allow for rates that are unacceptably slow for many
purposes. The objective of this paper is to consider the possible benefits of speech-to-text input mechanisms
(dictation and voice spelling) for handheld devices. By modeling throughput based on varying rates of speech,
correction speeds, and system recognition accuracies, we can compare expected speech throughput rates to current
throughput rates for PDAs.
Keywords: personal digital assistant (PDA), dictation, voice spelling, speech interface
Introduction
Recent years have seen the emergence and rising pop-
ularity of handheld personal digital assistants (PDAs).
These devices have many benefits, including being
small, lightweight, and extremely mobile. To date, there
have been two primary methods of inputting data to a
PDA—tapping a small onscreen (soft) keyboard and
using highly constrained handwriting recognizers such
as Graffiti (Graffiti is a registered trademark of Palm
Computing, Inc.) or Unistrokes (Unistrokes is a reg-
istered trademark of Xerox Corp.). The current input
speeds for these methods, however, are substantially
slower than input rates achieved with a personal com-
puter and keyboard.
Virtually all users of PDAs have experience with
personal computers and have some familiarity with
the standard computer keyboard (the QWERTY key-
board), with which expert typists can enter data at a
rates of approximately 55 words per minute (WPM)
with near perfect accuracy (Marklin et al., 1998;
Norman and Fisher, 1982). Prior research (discussed
below) has shown substantially slower rates of input for
various handwriting recognizers and soft keyboards.
Hand printing speeds are typically in the 12–
23 WPM range, and cursive handwriting speeds
range from 16 to just over 30 WPM (Soukoreff and
MacKenzie, 1995). These, necessarily, provide esti-
mates of the upper limits for this sort of text entry.
MacKenzie and Chang (1999), using two discrete print-
ing recognizers and a 9.5-inch tablet, found a mean text
entry speed of 17.1 WPM. The mean recognition accu-
racy was 92% when the recognizer was constrained to
lowercase letters and 90% when constrained to upper
and lower case letters. MacKenzie and Zhang (1997),
found high (95.8%) recognition accuracy for experi-
enced users of the Graffiti handwriting recognition sys-
tem, but did not report the entry speeds. Sears and
Arora (2001) reported a much slower text entry rate
of 4.95 WPM for the Graffiti recognizer with a recog-
nition accuracy of 95% for participants using a PDA
(rather than a tablet). They found that participants us-
ing the Jot recognizer were able to produce 7.74 WPM
with an average recognition accuracy of 88% (Jot is
70 Commarford and Lewis
aregistered trademark of Communication Intelligence
Corporation). Each of the previously mentioned studies
reported rates of entry for uncorrected text. Kleid and
Bonto (1995) asked users to enter a rather complex set
of letters, numbers, and special characters (a person’s
contact information) using Graffiti on a 6
(diagonal)
screen. Participants were to attempt 100% accuracy and
to use any editing tools that they felt would be helpful.
The participants were only able to enter 1.98 corrected
words per minute (CWPM).
Research has shown stylus tapping on a soft
QWERTY keyboard to be slightly faster than using
a handwriting recognizer, but these rates are still sub-
optimal. Zha and Sears (2001) reported that partici-
pants could input text on a PDA at a rate of 12.62 WPM
with an average input accuracy of 96% using the soft
keyboard. Using a tablet, MacKenzie et al. (1994)
found that users could type text at a rate of 22.9 WPM
with 99% accuracy. MacKenzie et al. (1999) reported
text entry rates of 20.2 WPM for participants tapping on
a full-sized paper QWERTY layout. Kleid and Bonto
(1995) had their participants use a soft keyboard to en-
ter the previously described set of complex text with
100% accuracy. Under these conditions, participants
were only able to obtain mean throughput rates of
5.17 CWPM.
In the past decade, users have been introduced to a
new method of inputting data to a personal computer—
speech dictation. While voice throughput is not yet typ-
ically as fast as typing with a full-sized QWERTY key-
board, it may offer a faster rate of throughput for hand-
held devices than the currently available options. Lewis
(1999) defined true throughput as the number of correct
words produced per minute, and found that participants
could achieve rates of 31.0 CWPM with multimodal
(manual and vocal) correction and 19.0 CWPM with
voice-only correction, using two commercially avail-
able desktop speech dictation products. Voice-only cor-
rection was significantly slower (29.1 seconds per cor-
rection) than multimodal correction (13.2 seconds per
correction).
If system designers could embed voice recognition
technology into the PDA environment in a way that al-
lows users to achieve throughput rates similar to those
observed by Lewis (1999), users would benefit greatly.
By 1999, estimates of recognition accuracy for com-
mercial desktop dictation systems ranged from 90–
95% (Koester, 2001; Lewis, 2001). Resource limita-
tions of these handheld devices, however, will probably
prevent the high levels of recognition accuracy reached
by the desktop software. In addition, multimodal cor-
rection on a PDA will include the use of a method of
input (soft keyboard or handwriting recognizer) that is
less efficient than those used with the desktop systems
in the Lewis study.
It is reasonable to consider the efficiency of two
methods of inputting data by voice in a PDA environ-
ment: speech dictation and voice spelling. These meth-
ods can exist in isolation or in combination. To dictate,
users would simply say the words they want to appear
on the screen. To voice spell, the user would say codes
for each letter of the alphabet. Due to the well-known
acoustic similarities of certain letters (for example, “B”
and “D” or “S” and “F”), recognition accuracy for letter
names is much lower than for a carefully selected set of
code words (Lewis and Commarford, 2002). For exam-
ple, rather than saying “d” the user would say, “dog.” It
would be possible to present a reminder display when
in voice spelling mode for users who do not have the
codes committed to memory.
The user would certainly be able to produce more
uncorrected words per minute via dictation than by
voice spelling. However, because of the limited gram-
mar set, voice spelling would probably achieve higher
levels of system recognition accuracy than dictation,
potentially reducing the need to correct. Furthermore,
corrections speeds for voice spelling (single-character
corrections) would probably be faster than correction
speeds for dictation (full-word corrections). Prior mod-
eling (Lewis, 1999) suggests that system recognition
accuracy and correction speeds are more important de-
terminants of true throughput than speaking rate. Mod-
eling true throughput (CWPM) based on input method,
system accuracy, time per correction, and speaking rate
can help us understand how substantial the accuracy
difference between spelling and dictation would have
to be for it to be advantageous to use the spell mode.
These models would also allow comparisons of each
of these speech input methods to reported throughput
rates for other PDA input methods.
The aforementioned research suggests faster rates
of throughput for soft keyboard input than for hand-
writing recognizers with a PDA. The Zha and Sears
(2001) study seems to provide the best estimate of PDA
throughput with a soft keyboard. These researchers
used a PDA and asked participants to input a 41-word
passage characteristic of a short business email mes-
sage. They found a mean input rate for this task of
12.62 WPM, with a 4% error rate. Assuming a very
high speed of correction, it seems reasonable to set the
Modeling Dictation 71
benchmark for the true throughput rate for soft key-
board input at 12 CWPM.
The rest of this report describes performance mod-
eling conducted to make the following comparisons of
throughput rates:
(a) dictation for the 150-WPM speaker vs. the 100-
WPM speaker
(b) expert spelling for the 150-WPM speaker vs. the
100-WPM speaker
(c) novice spelling for the 150-WPM speaker vs. the
100-WPM speaker
(d) dictation vs. expert spelling for the 150-WPM
speaker
(e) dictation vs. expert spelling for the 100-WPM
speaker
In addition, there will be a comparison of each
speech input rate to the target rate of 12 CWPM, which
appears to be the best estimate of the most efficient
method of text input currently available for stand-
alone PDAs (not docked and connected to a personal
computer).
Method
The source text for this evaluation was a 107-word
passage with 483 letters. This passage contained 101
spaces and seven punctuation marks. The passage con-
tained no capitalization, except for characters follow-
ing periods. We created six models of true through-
put (using CWPM) for speech dictation and voice
spelling (see Table 1). Each model contains a range
of system recognition accuracies from 50 to 100%
and a range of correction times from 5 to 35 sec-
onds.
The expert voice spelling models assume complete
automaticity in the assignment of the letters to their
Table 1. Description of six throughput models.
Input method User Speaking rate
Dictation All 100 WPM
150 WPM
Voice spelling Novice 100 WPM
150 WPM
Expert 100 WPM
150 WPM
respective codes (in other words, users need no pro-
cessing time to match letters to their codes or to retrieve
them from short term memory). The novice models as-
sume a 230 ms eye movement time to locate each letter
in the passage on the spell vocabulary reminder display.
The 230 ms eye movement time is consistent with that
given as the typical or “middle man” time by Card et al.
(1983).
The Models
Dictation Model
This model is a replication of that created by Lewis
(1999), but extends the upper limit for correction times
to 35 s because it is reasonable to anticipate longer cor-
rection times with a handheld device. Table 2 shows
the expected throughput data (in CWPM) for a 150-
WPM speaker and a 100-WPM speaker at varying lev-
els of system recognition accuracy and varying correc-
tion speeds. The bold numbers in the table indicate the
point at which speech becomes competitive with soft
keyboard input for each combination of speaking rate,
correction speed, and recognition accuracy (in other
words, exceed 12 CWPM). Figure 1 illustrates this
relationship.
The following worked example illustrates the
method by which we calculated CWPM for dictation.
The example is based on the cell in Table 2 that pro-
vides the throughput rate for a 150-WPM speaker who
takes 25 seconds on average to complete a correction
and is using a system that accurately recognizes 80%
of the dictated speech.
We calculated CWPM by dividing the number of
words to appear in the target application on the PDA
screen (107 in this case) by the total amount of time
it would take, in minutes, for the corrected text to ap-
pear. Total time consists of time to speak the passage
and time to correct the misrecognitions. We calculated
time to speak by dividing the number of words to be
spoken (107) by the speaking rate (150 WPM). The
result was 0.71 minutes to speak the 107-word pas-
sage. Multiplying the number of required corrections
by the correction speed (25 seconds) provided the cor-
rection time. We determined the number of required
corrections by multiplying the number of words spoken
(107) by the error rate (1 − recognition accuracy). For
the example data, the error rate is .20, resulting in 21.4
errors. Multiplying 21.4 errors by the 25-second correc-
tion speed produced 535 seconds, or 8.92 minutes. The
72 Commarford and Lewis
Table 2. Model of expected throughput (CWPM) rates for dictation.
Recognition accuracy
Speaking rate Correction speed 100% 95% 90% 85% 80% 75% 70% 65% 60% 55% 50%
150 WPM 5 sec 150 92.31 66.67 52.17 42.86 36.36 31.58 27.91 25.00 22.64 20.69
15 sec 150 52.17 31.58 22.64 17.65 14.46 12.24 10.62 9.38 8.39 7.59
25 sec 150 36.36 20.69 14.46 11.11 9.02 7.59 6.56 5.77 5.15 4.65
35 sec 150 27.91 15.38 10.62 8.11 6.56 5.50 4.74 4.17 3.72 3.35
100 WPM 5 sec 100 70.59 54.55 44.44 37.50 32.43 28.57 25.53 23.08 21.05 19.35
15 sec 100 44.44 28.57 21.05 16.67 13.79 11.76 10.26 9.09 8.16 7.41
25 sec 100 32.43 19.35 13.79 10.71 8.76 7.41 6.42 5.66 5.06 4.58
35 sec 100 25.53 14.63 10.26 7.89 6.42 5.41 4.67 4.11 3.67 3.31
0
20
40
60
80
100
120
140
160
100 95 908580757065605550
Recognition Accuracy
Throughput (CWPM)
S150- E 05
S150- E 15
S150- E 25
S150- E 35
S100- E 05
S100- E 15
S100- E 25
S100- E 35
S150 = 150 WPM speaking rate
S100 = 100 WPM speaking rate
E05 = 5 seconds per correction
E15 = 15 seconds per correction
E25 = 25 seconds per correction
E35 = 35 seconds per correction
Figure 1. Model of expected throughput (CWPM) rates for dictation.
sum of 0.71 minutes (speaking time) and 8.92 minutes
(correction time) was 9.63 minutes (total time). Finally,
dividing the 107 corrected words by 9.63 minutes of to-
tal time produced the presented value of 11.11 CWPM
(see Table 2).
Referring to Table 2, beginning with the right hand
side of the table, we can see that with a recogni-
tion accuracy of 50%, the user would need to be able
to make a correction every five seconds for dicta-
tion to compete with soft keyboard input. This speed
Modeling Dictation 73
of correction seems unrealistically fast given Lewis’
(1999) research, in which he found average multimodal
correction speeds of 13.2 seconds per correction with
desktop speech dictation systems. Given the size and
processing limitations of the PDA, correction times for
dictation would likely be a little slower, perhaps be-
tween 15 and 20 seconds. If we assume 15 seconds
per correction, recognition accuracies as low as 70–
75% would produce mean true throughput rates that are
highly competitive with the target of 12 CWPM, even
for a 100-WPM speaker. If correction speeds were as
slow as 25 seconds per correction, a recognition accu-
racy of about 85% would be necessary for dictation to
compete with soft keyboard input. At 35 seconds per
correction, the target accuracy would be 90%.
Expert Speller Model
The expert speller model allows estimation of the voice
spelling throughput rates for a speaker for whom the
spell letter codes have become automatic. Table 3
shows the expected spell data for an expert speller,
speaking 150 and 100 WPM at varying levels of recog-
nition accuracy and varying correction speeds. The
bold numbers in the table indicate the point at which
speech becomes competitive with soft keyboard input.
Figure 2 illustrates this relationship.
The calculations for the expert spelling model are
similar to those for the dictation model, except that the
number of words spoken is now 591 (the user must say a
code word for each of 483 letters, must say, “space” 101
times, and must say, “period” 7 times) and corrections
are at the character rather than the word level. The
following example goes through the steps of calculating
throughput for an expert speller speaking 150-WPM,
correcting one character every 25 seconds, and using a
Table 3. Model of expected throughput (CWPM) rates for the expert speller.
Recognition accuracy
Speaking rate Correction speed 100% 95% 90% 85% 80% 75% 70% 65% 60% 55% 50%
150 WPM 5 sec 27.16 16.71 12.07 9.45 7.76 6.58 5.72 5.05 4.53 4.10 3.75
15 sec 27.16 9.45 5.72 4.10 3.19 2.62 2.22 1.92 1.70 1.52 1.38
25 sec 27.16 6.58 3.75 2.62 2.01 1.63 1.38 1.19 1.04 0.93 0.84
35 sec 27.16 5.05 2.79 1.92 1.47 1.19 1.00 0.86 0.75 0.67 0.61
100 WPM 5 sec 18.10 12.78 9.88 8.05 6.79 5.87 5.17 4.62 4.18 3.81 3.50
15 sec 18.10 8.05 5.17 3.81 3.02 2.50 2.13 1.86 1.65 1.48 1.34
25 sec 18.10 5.87 3.50 2.50 1.94 1.59 1.34 1.16 1.02 0.92 0.83
35 sec 18.10 4.62 2.65 1.86 1.43 1.16 0.98 0.85 0.74 0.66 0.60
system that accurately recognizes 80% of the speech.
Dividing the number of utterances to be spoken (591)
by the 150-WPM speaking rate produced 3.94 min-
utes to speak the passage. Multiplying the number of
utterances spoken (591) by the error rate (.20) pro-
duced 118.2 errors, which would require 2955 seconds
(49.25 minutes) to correct at 25 seconds per error. Sum-
ming the time to speak (3.94 minutes) and the time
to correct (49.25 minutes) produced a total time of
53.19 minutes. Dividing 107 words in the passage by
53.19 produced 2.01 CWPM (see Table 3).
Table 3 shows that, with low levels of system recog-
nition accuracy, voice spelling will produce unaccept-
ably slow input rates, regardless of correction speed.
Assuming a correction speed of 15 seconds, recogni-
tion accuracy would need to be greater than 95% for
voice spelling to be a competitive method of inputting
text. However, the previously cited correction speeds
(Lewis, 1999) were for full-word corrections that re-
sulted from system misrecognitions of user dictation.
Correcting the misrecognition of a single character
would presumably be much quicker (see Koester, 2001,
p. 123), requiring only two actions (tapping or saying
backspace, followed by tapping the appropriate key or
saying the appropriate code). Assuming five seconds
per correction, 90% recognition accuracy would allow
150-WPM speakers to voice spell at a rate competi-
tive with soft keyboard input (12.07 CWPM). Slower
speaking expert spellers would require a recognition
accuracy of 95% for voice spelling to be a competitive
alternative.
Novice Speller Model
The novice speller model allows the estimation of the
throughput rates for a speaker who is just learning to
74 Commarford and Lewis
0
5
10
15
20
25
30
100 95 908580757065605550
Recognition Accuracy
Throughput (CWPM)
S150- E 05
S150- E 15
S150- E 25
S150- E 35
S100- E 05
S100- E 15
S100- E 25
S100- E 35
S150 = 150 WPM speaking rate
S100 = 100 WPM speaking rate
E05 = 5 seconds per correction
E15 = 15 seconds per correction
E25 = 25 seconds per correction
E35 = 35 seconds per correction
Figure 2. Model of expected throughput (CWPM) rates for the expert speller.
use spell mode and has not yet memorized the letter
codes. Table 4 shows the expected throughputs for a
novice speller at 150 WPM and at 100 WPM with
varying levels of system recognition accuracy and vary-
ing correction speeds. The bold numbers in the table
indicate the point at which speech becomes competi-
tive with soft keyboard input. Figure 3 illustrates this
relationship.
We calculated novice speller throughput times in
much the same way as the expert speller models. The
one exception is the addition of a time component,
code word search time, to the total time. The fol-
lowing describes the method of calculating through-
put for a novice speller speaking 150-WPM, requiring
25 seconds per correction for each misrecognized char-
acter, and using a system that accurately recognizes
80% of the spoken words.
The 3.94 minutes to speak the passage and the
49.25 minutes to correct the misrecognitions remain
the same as in the expert speller example above. How-
ever, we also assumed a 230 msec (.0038 minute)
search time for code words for each of the 483 let-
ters in the passage (but not for “period” or “space”).
Multiplying 483 code words by 0.0038 minutes per
search yielded 1.85 minutes of total search time. Sum-
ming time to speak (3.94 minutes), time to correct
(49.25 minutes), and time to search (1.85 minutes) re-
sulted in 55.04 minutes. Dividing 107 by 55.04 minutes
produced 1.94 CWPM (see Table 4).
The expected novice speller data follows a pattern
similar to the expert speller, but is slightly slower. At
five seconds per correction, novice spellers would re-
quire 95% or greater recognition accuracy to be com-
petitive with the soft keyboard.
Modeling Dictation 75
Table 4. Model of expected throughput (CWPM) rates for the novice speller.
Recognition accuracy
Speaking rate Correction speed 100% 95% 90% 85% 80% 75% 70% 65% 60% 55% 50%
150 WPM 5 sec 18.47 12.96 9.98 8.12 6.84 5.91 5.20 4.65 4.20 3.83 3.52
15 sec 18.47 8.92 5.20 3.83 3.03 2.50 2.14 1.86 1.65 1.48 1.34
25 sec 18.47 5.91 3.52 2.50 1.94 1.59 1.34 1.16 1.03 0.92 0.83
35 sec 18.47 4.65 2.66 1.86 1.43 1.16 0.98 0.85 0.74 0.66 0.60
100 WPM 5 sec 13.79 10.47 8.43 7.06 6.08 5.33 4.75 4.28 3.90 3.58 3.30
15 sec 13.79 7.06 4.75 3.58 2.87 2.39 2.05 1.80 1.60 1.44 1.31
25 sec 13.79 5.33 3.30 2.39 1.88 1.54 1.31 1.14 1.01 0.90 0.82
35 sec 13.79 4.28 2.53 1.80 1.39 1.14 0.96 0.83 0.73 0.66 0.59
0
2
4
6
8
10
12
14
16
18
20
100 95 90 85 80 75 70 65 60 55 50
Recognition Accuracy
Throughput (CWPM)
S150-E05
S150-E15
S150-E25
S150-E35
S100-E05
S100-E15
S100-E25
S100-E35
S150 = 150 WPM speaking rate
S100 = 100 WPM speaking rate
E05 = 5 seconds per correction
E15 = 15 seconds per correction
E25 = 25 seconds per correction
E35 = 35 seconds per correction
Figure 3. Model of expected throughput (CWPM) rates for the novice speller.
150-WPM Speaker Model
This model provides expected throughputs for some-
one who speaks at a rate of 150 WPM for dictation
and for expert spelling. The data enable the determi-
nation, for a given correction speed, of how much dif-
ference in recognition accuracy must exist for expert
voice spelling to be more efficient than dictation. Ta-
ble 5 shows the expected dictation and expert spelling
throughput rates for a user who speaks at a rate of
76 Commarford and Lewis
Table 5. Model of expected throughput (CWPM) rates for the 150-WPM speaker.
Recognition accuracy
Input method Correction speed 100% 95% 90% 85% 80% 75% 70% 65% 60% 55% 50%
Dictation 5 sec 150.00 92.31 66.67 52.17 42.86 36.36 31.58 27.91 25.00 22.64 20.69
15 sec 150.00 52.17 31.58 22.64 17.65 14.46 12.24 10.62 9.38 8.39 7.59
25 sec 150.00 36.36 20.69 14.46 11.11 9.02 7.59 6.56 5.77 5.15 4.65
35 sec 150.00 27.91 15.38 10.62 8.11 6.56 5.50 4.74 4.17 3.72 3.35
Spelling 5 sec 27.16 16.71 12.07 9.45 7.76 6.58 5.72 5.05 4.53 4.10 3.75
15 sec 27.16 9.45 5.72 4.10 3.19 2.62 2.22 1.92 1.70 1.52 1.38
25 sec 27.16 6.58 3.75 2.62 2.01 1.63 1.38 1.19 1.04 0.93 0.84
35 sec 27.16 5.05 2.79 1.92 1.47 1.19 1.00 0.86 0.75 0.67 0.61
150 WPM for varying levels of system recognition ac-
curacy and varying correction speeds. The bold num-
bers in the table indicate the point at which speech be-
comes competitive with soft keyboard input. Figure 4
illustrates this relationship.
0
20
40
60
80
100
120
140
160
100 95 90 8580757065605550
Recognition Accuracy
Throughput (CWPM)
D-E05
D-E15
D-E25
D-E35
ES-E05
ES-E15
ES-E25
ES-E35
D = Dictation
ES = Expert Spelling
E05 = 5 seconds per correction
E15 = 15 seconds per correction
E25 = 25 seconds per correction
E35 = 35 seconds per correction
Figure 4. Model of expected throughput (CWPM) rates for the 150 WPM speaker.
As shown in Table 5, given equivalent system recog-
nition accuracy, dictation will be superior across all
correction speeds, with the greatest differences com-
ing at the higher levels of accuracy. Given the large
difference in the size of the grammar sets for the two
Modeling Dictation 77
modes, however, recognition accuracy is expected to be
higher and correction speeds faster when the user is in
spell mode. Lewis and Commarford (2002) developed
avoice spelling alphabet and tested its accuracy with
a desktop system and headset microphone. The gram-
mar set (which included letter codes for voice spelling,
punctuation, and cursor control commands) produced
results that were 97.5% accurate under these condi-
tions. Given that accuracy should be no higher (and
might be lower) with a PDA microphone, 97.5% is an
estimate of the upper limit to voice spelling accuracy.
Assuming 95% recognition accuracy and 5 seconds per
correction for voice spelling and assuming 15 seconds
per correction for dictation, dictation with recogni-
tion accuracy as low as 80% would be more efficient
than voice spelling (with a voice spelling through-
put of 16.71 CWPM and a dictation throughput of
17.65 CWPM).
100-WPM Speaker Model
This model provides expected dictation and voice
spelling throughput rates for an expert speller who
speaks at a rate of 100 WPM. Table 6 shows the ex-
pected throughput rates. The bold numbers in the table
indicate the point at which speech becomes competi-
tive with soft keyboard input. Figure 5 illustrates the
relationship.
Again, the data show that, given equivalent system
recognition accuracy, dictation should be superior to
voice spelling across all correction speeds, with the
greatest differences at the higher levels of accuracy.
Assuming 95% recognition accuracy and 5 seconds per
correction for voice spelling and assuming 15 seconds
per correction for dictation, dictation with recogni-
tion accuracy as low as 75% would be more efficient
Table 6. Model of expected throughput (CWPM) rates for the 100-WPM speaker.
Recognition accuracy
Input method Correction speed 100% 95% 90% 85% 80% 75% 70% 65% 60% 55% 50%
Dictation 5 sec 100.00 70.59 54.55 44.44 37.50 32.43 28.57 25.53 23.08 21.05 19.35
15 sec 100.00 44.44 28.57 21.05 16.67 13.79 11.76 10.26 9.09 8.16 7.41
25 sec 100.00 32.43 19.35 13.79 10.71 8.76 7.41 6.42 5.66 5.06 4.58
35 sec 100.00 25.53 14.63 10.26 7.89 6.42 5.41 4.67 4.11 3.67 3.31
Spelling 5 sec 18.10 12.78 9.88 8.05 6.79 5.87 5.17 4.62 4.18 3.81 3.50
15 sec 18.10 8.05 5.17 3.81 3.02 2.50 2.13 1.86 1.65 1.48 1.34
25 sec 18.10 5.87 3.50 2.50 1.94 1.59 1.34 1.16 1.02 0.92 0.83
35 sec 18.10 4.62 2.65 1.86 1.43 1.16 0.98 0.85 0.74 0.66 0.60
than voice spelling (with a voice spelling through-
put of 12.78 CWPM and a dictation throughput of
13.79 CWPM).
General Discussion
Relatively Low Throughput for Handwriting
Recognition
Although it might seem counterintuitive, the evidence
from the published literature indicates that textual
throughput with a soft keyboard is more efficient than
handwriting recognition for PDAs, despite the im-
provements in recognition accuracy associated with
constrained alphabets such as Graffiti and Unistrokes.
The only cause of error in key tapping is tapping
the wrong key, but errors in handwriting recogni-
tion can be due to forming an incorrect character
(user error) or misrecognition of a correctly formed
character (system error). Furthermore, corrections
performed with handwriting recognition are also prob-
abilistic, making them more error prone than deter-
ministic key tapping. Thus, the low throughput ob-
served for handwriting recognition systems is most
likely the result of a longer total correction time.
Regardless of the cause, the throughputs reported in
the literature indicate that handwriting recognition is
a less effective input method than typing with soft
keyboards.
For these reasons, we have not included handwriting
throughput in our models. Instead, we have focused on
comparisons of speech input methods with the more
competitive method of soft keyboard input. For read-
ers interested in comparing speech throughput rates
with handwriting on a PDA, a generous estimate for
the true throughput of current handwriting recognition
78 Commarford and Lewis
0
20
40
60
80
100
120
100 95 90 8580757065605550
Recognition Accuracy
Throughput (CWPM)
D-E05
D-E15
D-E25
D-E35
ES-E05
ES-E15
ES-E25
ES-E35
D = Dictation
ES = Expert Spelling
E05 = 5 seconds per correction
E15 = 15 seconds per correction
E25 = 25 seconds per correction
E35 = 35 seconds per correction
Figure 5. Model of expected throughput (CWPM) rates for the 100 WPM speaker.
is 7 CWPM (based on the results for the Jot recognizer
in Sears and Arora, 2001).
Potential Usefulness of Voice Spelling and Dictation
as PDA Input Methods
The range of values used to model true throughput for
the speech input methods described in this paper have a
firm basis in previously published human performance
data, and were selected to encompass them. The typical
speaking rate for users who are transcribing text into a
dictation system is about 110 WPM (Lewis, 1999), so
we set the values for this rate in the models at 100 and
150 WPM.
Correction speeds ranged from about 13 to
30 seconds per correction in Lewis (1999), depending
on the correction strategy employed by users (multi-
modal for the faster times, voice only for the slower
times). Correction speeds for entry with standard key-
boards are faster, averaging about 3 seconds per cor-
rection (Koester, 2001). Given a soft rather than stan-
dard keyboard, it therefore seems reasonable to assume
about 5 seconds per correction for individual char-
acters, such as those produced during voice spelling.
Thus, the range of correction speeds included in our
models (5 to 35 seconds per correction) encompasses
correction speeds for keyed corrections, multimodal
corrections, and voice-only corrections.
Estimates of recognition accuracy for commercial
desktop speech dictation systems typically range from
90 to 95%, but with some additional variation above
and below that range (Koester, 2001; Lewis, 2001).
Resource and hardware limitations might prevent users
from achieving such high recognition accuracies with
aPDA. Therefore, the range of recognition accuracies
modeled (50 to 100%) seems appropriate to consider.
The models clearly show that both voice spelling and
dictation can be competitive with soft keyboard entry,
Modeling Dictation 79
and are therefore potentially useful methods for the in-
put of text into PDAs. All other things being equal, dic-
tation will be much more effective than voice spelling.
Due to their differing system requirements, however,
voice spelling might be the only viable method for
speech input of text into handheld devices with rela-
tively low system resources.
The models also illustrate the likely limits to dicta-
tion throughput as a function of accuracy and correc-
tion speed. Users can easily speak text at rates between
100 and 150 WPM, but are very unlikely to experience
true throughput much greater than 45 to 50 CWPM
(given 95% accuracy and multimodal correction speeds
of 15 seconds per correction) in the near future.
Summary of Key Findings
1. Assuming 15 seconds per correction with a PDA,
dictation would be as efficient as soft keyboard in-
put as long as speech recognition accuracy were
70% or greater. Assuming dictation recognition ac-
curacies as high as 85–90%, dictation would be ap-
proximately twice as productive as soft keyboard
entry.
2. Under ideal conditions (expert speller, 150-WPM
speaker, and 5-second correction speed), voice-
spelling accuracy must exceed 90% to be competi-
tive with soft keyboard entry.
3. Voice spelling will not be as efficient as dictation
unless voice spelling recognition accuracy is much
higher than dictation recognition accuracy.
Conclusions
Current methods for textual input to a PDA do not result
in a very high throughput. Of the two most common
contemporary methods, typing on a virtual keyboard
has a higher throughput than any approach to hand-
writing recognition. The best current estimate of true
throughput for typing on a PDA’s virtual keyboard is
about 12 CWPM. The models presented in this paper
show that dictation on a PDA would have a substan-
tial throughput advantage over current methods. Voice
spelling is a viable alternative to current methods for
PDAs that do not have sufficient computing power to
support dictation. The true throughput for any speech
input method with less than perfect recognition accu-
racy is critically dependent on correction speed, making
the design of efficient correction procedures of utmost
importance.
References
Card, S.K., Moran, T.P., and Newell, A. (1983). The Psychol-
ogyofHuman Computer Interaction. Hillsdale, NJ: Lawrence
Erlbaum.
Kleid, N.A. and Bonto, M.A. (1995). Handwriting Recognition and
Soft Keyboard Study.(Tech. Report 29.3008). Raleigh, NC: Inter-
national Business Machines Corp.
Koester, H.H. (2001). User performance with speech recognition: A
literature review. Assistive Technology, 13:116–130.
Lewis, J.R. (1999). Effect of error correction strategy on speech
dictation throughput. Proceedings of the Human Factors and
Ergonomics Society 43rd Annual Meeting, Santa Monica, CA:
Human Factors Society, pp. 457–461.
Lewis, J.R. (2001). The Accuracy Wars: Journalists’ Estimates of
Continuous Speech Product Dictation Accuracy from 1997–1999
(Tech. Report 29.3465). Raleigh, NC: International Business Ma-
chines Corp.
Lewis, J.R. and Commarford, P.M. (2002). Developing and Tuning
aVoice Spelling Alphabet for Devices with Small Displays. (Tech.
Report 29.3517). Raleigh, NC: International Business Machines
Corp.
MacKenzie, S.I. and Chang, L. (1999). A performance compari-
son of two handwriting recognizers. Interacting with Computers,
11:283–297.
MacKenzie, S.I., Nonnecke, R.B., McQueen, J.C., Riddersma, S.,
and Meltz, M. (1994). A comparison of three methods of character
entry on pen-based computers. Proceedings of the Human Factors
and Ergonomics Society 38th Annual Meeting, Santa Monica, CA:
Human Factors Society, pp. 330–334.
MacKenzie, S.I. and Zhang, S.X. (1997). The immediate usability of
Graffiti. Proceedings of Graphics Interface’97,Toronto, Canada:
Canadian Information Processing Society, pp. 129–137.
MacKenzie, S.I., Zhang, S.X., and Soukoreff, R.W. (1999). Text en-
try using soft keyboards. Behaviour and Information Technology,
18:235–244.
Marklin, R.W., Simoneau, G.G., and Hoffman, D. (1998). Effects of
computer keyboard setup parameters and user’s anthropometric
characteristics on wrist deviation and typing efficiency. Proceed-
ings of the Human Factors and Ergonomics Society 42nd Annual
Meeting, Santa Monica, CA: Human Factors Society, pp. 876–
880.
Norman, D.A. and Fisher, D. (1982). Why alphabetic keyboards are
not easy to use: Keyboard layout doesn’t much matter. Human
Factors, 24:509–519.
Sears, A. and Arora, R. (2001). An evaluation of gesture recognition
for PDAs. In M.J. Smith, G. Salvendy, D. Harris, and R.J. Koubek
(Eds.), Usability Evaluation and Interface Design: Cognitive En-
gineering, Intelligent Agents and Virtual Reality, Mahwah, NJ:
Lawrence Erlbaum Associates, pp. 1–5.
Soukoreff, R.W. and MacKenzie, S.I. (1995). Theoretical up-
per and lower bounds on typing speed using a stylus and
soft keyboard. Behaviour & Information Technology, 14:370–
379.
Zha, Y. and Sears, A. (2001). Data entry for mobile devices using
soft keyboards: Understanding the effect of keyboard size. In M.J.
Smith, G. Salvendy, D. Harris, and R.J. Koubek (Eds.), Usability
Evaluation and Interface Design: Cognitive Engineering, Intelli-
gent Agents and Virtual Reality,Mahwah, NJ: Lawrence Erlbaum
Associates, pp. 16–20.
