Frequency Distributions for Names and Unconstrained Words Associated with the Letters of the English Alphabet

Abstract

This paper provides frequency distributions (constrained and unconstrained) for words and names associated with the letters of the English alphabet. These distributions can be useful when developing voice spelling user interfaces (Lewis & Commarford, 2003). The data suggest that the unconstrained word distribution is preferable to the names distribution because, contrary to expectation, the overall consistency of unconstrained responses was not poorer than the responses in the names distribution and the occurrence of cases in which participants did not provide a response for the names distribution was significantly greater than for the unconstrained distribution.

Full text

Frequency Distributions for Names and Unconstrained Words
Associated with the Letters of the English Alphabet
James R. Lewis
IBM Pervasive Computing Division
8051 Congress Ave, Suite 2227
Boca Raton, FL 33487
jimlewis@us.ibm.com
Abstract
This paper provides frequency distributions (constrained and unconstrained) for words and names
associated with the letters of the English alphabet. These distributions can be useful when
developing voice spelling user interfaces (Lewis & Commarford, 2003). The data suggest that the
unconstrained word distribution is preferable to the names distribution because, contrary to
expectation, the overall consistency of unconstrained responses was not poorer than the responses
in the names distribution and the occurrence of cases in which participants did not provide a
response for the names distribution was significantly greater than for the unconstrained
distribution.
1 Introduction
Voice spelling is an important component of many speech systems (embedded in other appliances,
accessed by telephone, or for hands-free use of a desktop dictation system). If a user needs to use
a word that is not in the system’s built-in lexicon and either cannot or chooses not to use a
keyboard, then the user must spell the word by voice.
The standard approach to voice spelling is to let users say the letter names or, because some letter
names have high acoustic confusability, to also permit the use of the military alphabet (alpha, beta,
Charlie, etc.). The military alphabet has an advantage over letter names in that the words chosen
for the military alphabet have very low acoustic confusability. The military alphabet has a
disadvantage over letter names in that it is somewhat difficult to memorize the military alphabet.
The purpose of the current research was to find out what words come first to people’s minds when
asked to produce a word for each letter of the alphabet. In one experiment participants provided
words without constraint. In a separate experiment, participants provided first names for each
letter. The reason for constraining participants in this way was to see if participants under this
condition produced more consistent responses than those in the unconstrained condition.

2 Method
2.1 Participants
The source for participants in each experiment was a set of 400 IBM employees (800 employees in
all, 400 for each experiment), selected at random from an internal e-mail directory of all of the
IBM employees in the United States. Each participant received an e-mail invitation to participate.
Of the employees invited to participate, 103 (26%) responded to the invitation for the first
experiment (no constraints) and 120 (30%) responded to the invitation for the second experiment
(names for letters).
2.2 Materials
I used WebSurveyor
1
to construct a web-based form for the evaluation. The form simply provided
a space by each letter of the alphabet (arranged vertically on the page in alphabetical order), with
instructions appropriate for the specific experiment (to either provide any word for each letter, or
to provide only first names).
2.3 Procedure
After receiving the e-mailed invitation, participants clicked a link in the message that brought up
the WebSurveyor page containing the survey. Participants read an introduction explaining the
purpose of the survey, then completed and submitted the form. The WebSurveyor program kept
track of the raw data. After collection, I used a spreadsheet to organize, summarize, and analyze
the results.
3 Results
From the survey data I created a set of 52 tables, one for each letter of the alphabet by production
condition (unconstrained or constrained). Each table contained the submitted words, listed in
decreasing frequency of submission. Each table also showed the cumulative frequency and
indicated (1) the point at which the cumulative frequency for a given frequency of submission
came as close as possible to 80%, (2) the number of words included in the list up to that point, (3)
the standard deviation of the frequencies, and (4) the total number of different words submitted.
The choice of 80% as a cutoff point was arbitrary, but seemed reasonable for the purpose of
ensuring fairly comprehensive coverage and for comparing the cutoff points of the different
distributions. See the Appendix for a summary table of the three most-frequently produced words
for each condition (for the complete set of tables, see Lewis, 2001).
I computed the correlations among the 80% cutoff points, standard deviations, and total word
counts for both word distributions. In both distributions the cutoff points and total word counts
had a positive correlation (r = .84, p = 0.0000001 for unconstrained; r = .50, p = 0.01 for name-
constrained), and both of these variables had negative correlations with the standard deviations (r
ranging from -.64 to -.87, all p < .0004). These correlations show the interesting (though perhaps
not surprising) fact that the more words produced for a letter (which increases the 80% cutoff
point and the total word count), the smaller the standard deviation.
1
WebSurveyor is a trademark of WebSurveyor Corp.

A set of t-tests (treating letters as subjects) indicated that there was no significant difference
between the distributions for the variables of 80% cutoff point, standard deviation, and total word
counts. There was a significant difference between the distributions for the variable of Nones –
the percentage of responses for which respondents indicated that they couldn’t think of a word for
a given letter (t(25) = 3.61, p = .001). This never occurred in the unconstrained distribution, but
happened for a number of letters in the names distribution (most notably, Q, U, X, Y and Z).
4 Discussion
The hypothesis that asking for first names would produce more consistent responses than
completely unconstrained responding did not hold. For some letters this was true, but for others it
was not. Overall, the distributions had similar statistical properties with regard to 80% cutoff
points, standard deviations, and total word counts (all measures of response consistency).
However, the number of letters for which participants were not able to think of any appropriate
response was greater for first names than for unconstrained words. For this reason, designers
should avoid requiring users to provide first names when using words to perform voice spelling.
One reasonable approach for limited voice spelling grammars would be to provide three
candidates for each letter – the word from the military alphabet, the word with the greatest
frequency for each letter from the unconstrained tables, and the word with the greatest frequency
for each letter in the first names tables. Depending on the capacity of the system, it would be
reasonable to include more of the words from each source to increase the odds of having a match
to a user’s choice of word when voice spelling.
Another use for the results of these experiments is in the spoken feedback of alphanumeric
information. Many current systems simply feed back letter names, but the confusability of letter
names is a liability in spoken language output as well as in spoken language input. Some systems
use the military alphabet for feedback (for example, “A as in alpha”). The output of such a system
would sound more familiar (less formal) if it used more familiar words, such as “A as in apple”.
5 References
Lewis, J. R. (2001). Frequency distributions for names and unconstrained words associated with
the letters of the English alphabet (Tech. Report 29.3437). West Palm Beach, FL: International
Business Machines Corp. (http://drjim.0catch.com/vspellwords.pdf )
Lewis, J. R., and Commarford, P. M. (2003). Developing a voicespelling alphabet for PDAs. IBM
Systems Journal, 42, 624-638.

6 Appendix. The Three Most-Frequently Produced Words as a
Function of Constraint and Letter
Word
(Unconstrained)
Percent Cumulative
Word
(Name)
Percent Cumulative
apple 67.0 67.0 Adam 10.8 10.8
alpha 11.7 78.6 Alan 9.2 20.0
able 4.9 83.5 Albert 5.8 25.8
boy 40.4 40.4 Bob 33.3 33.3
baker 9.6 50.0 Brian 8.3 41.7
bravo 9.6 59.6 Barbara 6.6 48.3
cat 38.5 38.5 Charlie 32.3 32.3
Charlie 37.5 76.0 Cathy 12.8 45.1
Charles 2.9 78.8 Charles 7.5 52.6
dog 66.0 66.0 David 45.8 45.8
David 13.6 79.6 Dave 9.2 55.0
delta 8.7 88.3 Dog 7.5 62.5
elephant 21.7 21.7 Edward 29.4 29.4
Edward 14.2 35.8 Ed 7.6 37.0
echo 12.3 48.1 Eric 5.0 42.0
Frank 34.0 34.0 Frank 63.8 63.8
fox 21.4 55.3 Fred 11.8 75.6
foxtrot 4.9 60.2 Fox 4.7 80.3
George 25.0 25.0 George 48.8 48.8
girl 16.3 41.3 Gary 9.9 58.7
good 9.6 51.0 Greg 9.9 68.6
help 13.6 13.6 Harry 32.5 32.5
Henry 13.6 27.2 Henry 25.8 58.3
Harry 11.7 38.8 Harold 5.0 63.3
igloo 13.5 13.5 Irene 11.1 11.1
India 12.5 26.0 Ian 10.3 21.4
Indian 7.7 33.7 Isabel 9.4 30.8
Jack 16.5 16.5 Jack 25.6 25.6
jump 11.7 28.2 John 23.1 48.7
John 9.7 37.9 James 6.8 55.6
king 19.6 19.6 Kevin 16.2 16.2
kite 16.7 36.3 Karen 14.3 30.5
kilo 10.8 47.1 Ken 13.3 43.8
Larry 22.3 22.3 Larry 40.0 40.0
love 10.7 33.0 Linda 11.7 51.7
Lima 6.8 39.8 Laura 4.2 55.8
Mary 42.7 42.7 Mary 46.1 46.1
Mike 8.7 51.5 Mike 10.2 56.3
man 6.8 58.3 Michael 9.4 65.6

Word
(Unconstrained)
Percent Cumulative
Word
(Name)
Percent Cumulative
Nancy 49.5 49.5 Nancy 72.9 72.9
nice 5.8 55.3 Nathan 3.4 76.3
no 5.8 61.2 Nick 3.4 79.7
open 21.4 21.4 Oscar 38.1 38.1
Oscar 18.4 39.8 Oliver 6.8 44.9
orange 12.6 52.4 Oprah 6.8 51.7
Paul 23.5 23.5 Paul 40.0 40.0
peter 13.7 37.3 Peter 35.8 75.8
Papa 5.9 43.1 Papa 2.5 78.3
queen 31.1 31.1 Quency 13.6 13.6
quick 11.7 42.7 Quenten 13.6 27.1
Quebec 8.7 51.5 Queen 8.5 35.6
Robert 17.5 17.5 Robert 38.3 38.3
rabbit 7.8 25.2 Roger 9.2 47.5
Romeo 6.8 32.0 Richard 8.3 55.8
Sam 37.5 37.5 Sam 49.2 49.2
snake 3.8 41.3 Sally 8.3 57.5
star 3.8 45.2 Steve 6.7 64.2
Tom 35.3 35.3 Tom 55.0 55.0
tango 6.9 42.2 Thomas 6.7 61.7
Thomas 4.9 47.1 Tony 5.0 66.7
under 20.0 20.0 Ursula 20.5 20.5
uncle 14.7 34.7 Ulysses 10.3 30.8
ugly 8.0 42.7 Uncle 6.0 36.8
vertical 49.0 49.0 Victor 58.3 58.3
victory 16.7 65.7 Vicky 6.7 65.0
violin 4.9 70.6 Victoria 4.2 69.2
water 9.0 9.0 William 42.0 42.0
William 9.0 18.0 Walter 13.4 55.5
whiskey 8.0 26.0 Wayne 5.9 61.3
x-ray 64.6 64.6 Xavier 31.1 31.1
xylophone 16.7 81.3 X-ray 13.4 44.5
Xerox 11.5 92.7 Xena 10.1 54.6
yellow 45.5 45.5 Yolanda 30.5 30.5
yes 13.1 58.6 Yello 6.8 37.3
Yankee 10.1 68.7 Yvonne 5.9 43.2
zebra 75.2 75.2 Zebra 18.4 18.4
zoo 5.9 81.2 Zach 9.6 28.1
zero 5.0 86.1 Zachary 7.9 36.0