Squibs and Discussions 
Aligning Phonetic Segments 
Articulation Assessment 
for Children's 
Harold L. Somers* 
UMIST 
In a recent paper published in this journal (Covington 1996), an algorithm is described which 
aligns segments within a pair of words for the purpose of identifying historical cognates. This 
algorithm could have a further application in the field of speech therapy, and in particular in 
the practice of articulation assessment of young children. The present author developed a similar 
algorithm some years ago for this purpose. In this paper, we explore some points of comparison 
between the two applications. 
1. Articulation Testing 
It is well known that a child's acquisition of phonology is gradual, and can be charted 
according to the appearance of phonetic distinctions (e.g., stops vs. fricatives), the 
disappearance of childish mispronunciations, especially due to assimilation (\[9o9\] for 
dog), and the ability to articulate particular phonetic configurations (e.g., consonant 
clusters). Childhood speech impediments, often a symptom of other learning disor- 
ders, can often be diagnosed early on by the identification of delayed acquisition of 
these articulatory skills. Whether screening whole populations of children, or assessing 
individual referrals, the articulation test is an important tool for the speech clinician. 
A child's articulatory development is usually described with reference to an adult 
model, and in terms of deviations from it: a number of phonological "processes" 
(Table 1) can be identified (see Ingrain \[1976\]), and their significance with respect 
to the chronological age of the child assessed (though often processes interact, so 
for example when spoon is pronounced \[fun\] we have consonant-cluster reduction 
and assimilation). In Somers (1978a, 1979) I reported a computer program called CAT 
(Computerised Articulation Test), which I had developed to automate the assessment 
of children's articulation tests. At the heart of this program was an algorithm very 
similar to the one reported by Covington. 
Whereas Covington seeks to align the segments in possible historical cognates, 
CAT aligns the segments of a child's articulation with those of the adult model, and on 
the basis of this looks for evidence of the phonological processes listed in Table 1. For 
example, if elephant \[ehfAnt\] is pronounced \[evot\], we need to decide which of several 
possible alignments is the most plausible (cf. Covington 1996, 481): 
~--vo-t ev~t--- cv~---t e--vot- etc. 
~iIfont clIfont ¢iIfont ¢iIfont 
If applied to a body of articulation data, e.g., a corpus of, say, 45 words elicited 
from the child as in a standardized articulation test, the evidence of each phonological 
process can be quantified, and the overall picture compared with the model of the 
average child, to give an individual's "articulation age" and profile. 
* Centre for Computational Linguistics, UMIST, Manchester, England. E-mail: harold@cd.umist.ac.uk 
(~) 1999 Association for Computational Linguistics 
Computational Linguistics Volume 25, Number 2 
Table 1 
Phonological processes. 
Process Example Adult Model Child's Version 
Final consonant deletion queen kwin kwi 
Unstressed syllable deletion elephant elif0nt cf0nt 
Consonant cluster reduction stamps stamps tam 
Stopping kiss kIs kit 
Fronting key ki ti 
Denasalisation mummy mAmi bAbi 
Affrication tent tent tents 
Vocalisation (1) bottle botl botu 
Vocalisation (2) chimney tj~Imni tJ~immi 
Depalatalisation fish flJ~ hs 
Devoicing dogs do9 z doks 
Voicing tent tent dent 
Assimilation dog do9 9 o9 
Lisping kiss kls ki0 
S-lateralisation fish fif fi4 
Ejectivisation tent tent tent' 
Metathesis remember rImemb o mIrcmb o 
Gliding look 1uk wuk 
2. Use of the Computer by Speech Clinicians 
Early studies reporting the use of computers by speech pathologists include Faircloth 
(1971), van Demark and Tharp (1973), and Telage (1980), none of which involves au- 
tomatic analysis of the input, though the last named uses binary articulatory features 
in a way almost identical to CAT. Comparatively little has appeared in the speech- 
language disorders literature on the specific topic of computerized articulation testing 
in the nearly 20 years since the CAT program was developed. Software for computer- 
ized language analysis does exist, but is mainly for grammatical and lexical analysis. 
Fairly thorough overviews are given by Rushakoff (1984), Rushakoff and Schwartz 
(1986), Long (1991), Long and Masterson (1993), and Miller and Klee (1995), though 
of course these may be more or less out of date. Other very general works such as 
Schwartz (1984), Curtis (1987), Silverman (1987), Cochran and Masterson (1995), and 
Masterson (1995) cover the use of computers by clinicians for all aspects of their work, 
including screening and diagnosis of various language skills (lexis, grammar, under- 
standing, and auditory skills, as well as articulation) but also research (use of statistics), 
treatment (computer-based games), and clerical uses. 1 
Those programs reported in the literature which specifically address the problem 
of articulation testing are listed below. For many of these programs, it seems that 
the only published information is in the user manual that accompanies the software. 
As far as one can tell, in none of the packages is the data analysis fully automatic. 
The following packages have been reviewed or discussed in the articles as cited: CAPP 
(Computer Analysis of Phonological Processes) (Long 1991; Kennedy 1986); Computer 
Managed Articulation Diagnosis (Bardzik 1986; Long 1991); Computerized Profiling 
1.0 (Klee and Sahlie 1994); Computerized Profiling 2.0 (Gregg and Andrews 1995); ISPA 
1 Several papers refer to articles in the Journal for Computer Users in Speech and Hearing, but at the time of writing I was unfortunately unable to locate any copies of this obviously relevant journal. 
268 
Somers Aligning Phonetic Segments 
(Interactive System for Phonological Analysis) (Ball 1994); Lingquest 2 (Long 1991); PAL 
(Pye Analysis of Language) (Pye and Ingram 1988; Leonard 1989); PDAC (Phonological 
Deviation Analysis by Computer) (Perry 1995); PEPPER (Programs to Examine Phonetic 
and Phonological Evaluation Records) (Dyson 1987; Pollock 1988); Process Analysis 2.0 
(Long 1991). 
The best-known computer application in speech-language pathology research is 
the CHILDES database of language samples and associated software (MacWhinney and 
Snow 1985; MacWhinney 1992; Sokolov and Snow 1994). This is primarily aimed at fa- 
cilitating the storage and search of large databases of transcribed clinical data, where 
the transcription is basically orthographic, with mark-up for gestures, pauses, and 
other conversational features. Provision is made for a phonetic transcription too, us- 
ing a "translation" of the IPA (International Phonetic Association) alphabet into ASCII 
symbols called "UNIBET" (MacWhinney 1992, 61ff). Although the organizers of the 
CHILDES database have had input from computational linguists on the question of 
mark-up, there is little or no automatic analysis. Crucially, no attempt is made to com- 
pare on a phone-by-phone basis the child language data with adult models, so data 
on the types of phonological process listed in Table 1 cannot be extracted. 
This situation is typical of child language software, exemplified by Pye and Ingram 
(1988), whose PAL system uses a simple transcription, based on the IPA consonant chart, 
without the possibility of diacritics or special symbols to indicate specifically childish 
articulations. The system is unable to compare adult models with the child's output, 
and can only produce a "phonological lexicon", i.e., a list of the different sounds 
attested: it is then up to the clinician to analyze this inventory, e.g., to see if sounds 
are used contrastively, or in complementary distribution. The authors suggest that 
matching the child's utterances to an adult model would involve a procedure which 
"would have to be very sophisticated indeed to handle complex cases of metathesis 
and deletion" (p. 124). As we show in Somers (1979) and in the next section, CAT was 
able to handle metathesis and deletion without being "very sophisticated indeed." 
3. The CAT Algorithm 
Since the Somers (1979) article was aimed at speech therapists, it did not describe 
the alignment algorithm as such, which is described only in a local journal (Somers 
1978a) and--in great detail--in an M.A. thesis (Somers 1978b). It bears comparison 
with Covington's algorithm, though it should be said that the implementation in Pascal 
would be judged crude in the light of modem programming practice. 
3.1 Coding the Input 
The articulation data is coded as a narrow transcription, identifying phonetic detail 
such as secondary articulations, which can be important in speech therapy, in a fairly 
transparent notation, despite the limitations of the (capitals only) character set: pri- 
mary phones are identified by single characters, with diacritics indicated in brackets, 
for example N (D) would indicate a dental (rather than alveolar) \[n\]. The notation is 
interpreted internally as bundles of articulatory features. The adult models are stored 
in a similar form. In each word, one vowel is marked as the primary stress, and 
this is taken as an anchor point for the alignment. The program as a whole ignores 
vowel quality, and in the CAT transcription any one of five vowel symbols (the vowel 
characters AEIOU in fact) can be used, the choice of one or the other being merely cos- 
metic. This treatment of vowels is a reasonable expedient. Primarily, CAT is aimed at 
consonant articulation, which is also the main concern of speech clinicians: see Stoel- 
Gammon and Herrington (1990) who state that "vowels are mastered earlier \[than 
269 
Computational Linguistics Volume 25, Number 2 
TARGET RESPONSE F/L LAB DEN ALV RET PAL VEL UVU GLT STP FRC LAT NAS VWL 
S OMITTED 
T D /-/ - + ..... + .... 
A(') A(') ............ + 
M 9 - /-I .... I+/ ..... + - 
P OMITTED 
S X + - - /-/ - - /+/ - - + - - 
Figure 1 
Alignment of stamps pronounced as \[daI3x \]. The features where the target and response differ 
are highlighted with slant brackets. 
consonants\] and tend to evidence fewer errors" (p. 145). Regarding the identification 
of a single stressed vowel as an anchor point for the alignment, stress patterning (at 
least in stress-timed languages like English) is one of the first features of phonology 
to be acquired by children: again, if this is still a problem, then the distinctions tested 
by CAT will certainly be too fine-grained for such a subject. 2 
3.2 Alignment 
The alignment is based on taking the highest-scoring matches in terms of features, 
much as suggested by Covington (p. 490), so that in the elephant example above, the 
alignment \[v\]:\[f\] is preferred over the alignment \[v\]:\[1\]. Since the number of features 
for each segment remains constant, it is a simple matter of adding up the number of 
common features (+ or -), and taking the highest total. The algorithm works on the 
basis of "syllables" centred around a vowel. With the stressed vowel as an anchor 
point, the search-space is reduced to a comparison of the syllables either side of it: 
note that "vowel" is also marked as a feature. This is generally straightforward if the 
words are mono- or disyllabic, or trisyllabic with the stress on the second syllable. In 
other cases, if there is gross distortion of the consonants as well as inserted or omitted 
syllables, alignment can become somewhat arbitrary. 
The algorithm takes some other factors into account, and is "tuned" to look out 
for certain processes that undermine the simplistic sequential skip-and-match search 
(which is also the basis of Covington's algorithm): two such processes are metathesis 
(e.g., remember:\[mIr~mb~\]) and merging. In merging, a consonant cluster is simplified 
so that the resulting phone shares features of the two merged phones, e.g., box \[boks\] 
pronounced \[bot\], where the It\] has the place of articulation of the \[s\], but the manner 
of articulation of the \[k\]. Identifying metathesis can be rendered more complex by the 
coincidence of some other process, e.g., stopping, so that elephant becomes \[cpIlont\] 
with the \[1\] and \[f\] swopped round, and the \[f\] replaced by a \[p\]. The CAT alignment 
algorithm looks for these explicitly. Figure 1 shows the result of the alignment of 
stamps:\[daoxs\], as it was actually presented. 
The algorithm first aligns the marked vowel. It then takes the sequence of seg- 
ments either side of the vowel. For \[st\]:\[d\], \[d\] is aligned with \[t\] rather than \[s\] as 
\[t\]:\[d\] represents a difference of only one feature, while \[s\]:\[d\] differs in three fea- 
tures. The evidence for a merge is the same as for the simpler devoicing analysis, 
2 I am grateful to the anonymous reviewer who queried this aspect of the algorithm. 
270 
Somers Aligning Phonetic Segments 
so the latter is preferred. In the case of \[mps\]:\[~x\], the algorithm compares the four 
possibilities \[m\]:\[~\], \[p\]:\[~\], \[p\]:\[x\], and \[s\]:\[x\], as well as the possible merges \[mp\]:\[~\] 
and \[ps\]:\[x\], in that order. It does not consider the matches \[m\]:\[x\] or \[s\]:\[~\] as these 
would involve a simultaneous insertion and deletion (cf. Covington's "no-alternating- 
skips rule," p. 482). As usual, the solution with the least "cost" in terms of feature 
differences is chosen. If the sequence includes a "no cost" match, this would imme- 
diately be preferred. The test for metathesis would also be made when there is an 
"unstressed" vowel in the sequence, though not in consonant clusters (so vest:\[vcts\] 
would be analyzed incorrectly). 
3.3 Comparing CAT and Covington's Algorithm 
Comparing the CAT algorithm with Covington's, it seems that a key difference is the 
manual identification of a favored segment the "stressed" vowel-- as an anchor 
point. This can drastically reduce the search-space, especially if it happens to occur 
near the middle of the string, as in the above example. Apart from this, both algorithms 
work on a sequential match-or-skip, comparing the relative cost of each match, and 
narrowing the search-space by halting the search if a perfect match is found. The CAT 
algorithm has the additional task of searching explicitly under certain circumstances 
for metathesis and merges. 
Apart from Covington's more sophisticated programming style, the only other 
difference between our techniques is in the scoring method. Covington's (p. 487) 
seems simpler than my own, in that the penalties reflect different types of (mis-)match, 
whereas in CAT the score derives more directly from the phonetic nature of the match. 
Covington states, on the same page, that "excessively narrow phonetic transcriptions 
do not help; they introduce too many subtle mismatches that should have been ig- 
nored." The CAT alignment algorithm, however, makes quite the opposite assumption, 
since the nature of the task demands a particularly narrow transcription. Covington 
also states (p. 490) that his algorithm could be improved by using phonetic features. 
It is enlightening to take Covington's cognate alignment examples and to see what 
CAT would make of them. Looking first at the Spanish-French pairs (pp. 488f), we find 
that CAT agrees with Covington in 16 of the 20 cases. CAT has problems in three cases 
where the French has lost syllables that are stressed in Spanish, as in cabdza:cap (1); 3 
in the case of drboharbre (2), CAT gets the correct alignment as identified by Covington 
(p. 488) if we omit the schwa in the French transcription (as would be normal for 
Parisian French (Armstrong 1967, 117). 
Example1 
(1) k a b e 0 a k a b 6 0 a 
k ap- - - k- - &p- 
(2) a r b - o 1 6 r b - o 1 
arbro- 6Rb--R 
For the English-German data (pp. 490f) CAT gets exactly the same alignments as 
Covington for all 20 pairs (including the incorrect analysis of this:dieses), though in CAT 
we would not transcribe the second element of the diphthongs in four of the examples. 
Like Covington's algorithm, CAT would correctly assign the \[0\] of mouth with the \[t\] 
rather than the \[n\] of Mund. 
3 In this and subsequent examples, the CAT alignments (on the fight) are shown in IPA; an acute accent 
marks the "stressed" vowel. Covington's alignments are shown on the left. 
271 
Computational Linguistics Volume 25, Number 2 
The examples considered so far have been quite straightforward (and much easier 
to align than typical child language data). The English-Latin cognates (pp. 492f) present 
more of a challenge. Applying the accepted rules of Latin stress, the CAT and Cov- 
ington alignments differ in five of the 20 cases: In four of these, blow'.fl~re, fish:piscis, 
full:pl~nus, and tooth:dentis (3), CAT does better than Covington, and in three other 
cases (grass:gr~men, heart:cordis, and mountain:mSns), CAT gets as first choice the align- 
ment Covington ranks third, second, and second respectively. With just one exception 
(knee:genF~), CAT does as well as or better than Covington. 
Example2 
(3) - -- t uwO t ~- - - 0 
denti-s d~ntis 
On the Fox-Menomini data (p. 494), CAT gets the same results as Covington on 
all ten examples if we assume either the first or the second vowel is stressed. Finally, 
Covington presents a variety of language-pair examples (p. 495). Again, the correct 
placement of the stressed vowel is important, leading to a wrong alignment for cen- 
tum:hekaton (4), and preventing the \[0\]:\[x\] alignment in thugat~r:Tochter (5). CAT does 
worse than Covington in one other case did~rni:d8 (6), but better in three cases daugh- 
ter:thugat~r (7), ager:ajras (8), and bhar~mi:pher~ (9). For centum:satem they both get the 
same alignment. 
Example3 
(4) --kentum --k--4ntum 
heka-t on hekat 6n--- 
(5) thug a t ~ r thug ~ - t e r 
tox-tor t--6xter 
(6) d i d 8 m i d i d 6 m i 
--dS-- d--6-- 
(7) - - d o t o r d - - 6 t a r 
thug a t ~ r thug 6 t e r 
(8) a - g e r 6 g e r - - 
aJ r as 6d3- r as 
In summary then, CAT does worse on the Spanish-French, better on the English- 
Latin, and about the same on the rest. Considering that Covington's algorithm is aimed 
at dealing with this sort of data, this is a good result for CAT. 
In a reciprocal comparison, Michael Covington was kind enough to run his algo- 
rithm on some child language data that I sent him. Of 25 examples, all of which CAT 
handles correctly, Covington's algorithm also got the correct alignment, but often it 
was unable to distinguish between alternative alignments, all of which received the 
same score. For example, with the stamps:\[da~xs\] alignment mentioned above, all six 
different combinations of consonant alignment either side of the vowel are proposed 
with an equal score. This is because, as Covington (personal communication) readily 
points out, "it doesn't know anything about place of articulation." 
272 
Somers Aligning Phonetic Segments 
4. Conclusions 
4.1 Connolly's New Algorithm 
Since the appearance of Covington's article (and even since the first draft of this reply), 
a highly relevant article has appeared, which--coincidentally--addresses the issues 
raised here (Connolly 1997). In this two-part article, Connolly first suggests ways of 
quantifying the difference between two individual phones, on the basis of perceptual 
and arficulatory differences, and using either a Euclidean distance metric or, like CAT, 
a feature-based metric. Connolly's proposals are more elaborate, however, in that they 
permit specific differences to be weighted, so as to reflect the relative importance 
of each opposition. In the second part of the article, Connolly introduces a distance 
measure for comparing sequences of phones, based on the Levenshtein distance well- 
known in the speech processing and corpus alignment literature (inter alia). Again, this 
metric can be weighted, to allow substitutions to be valued differentially (presumably 
on the basis of the individual phone distance measure as described in the first part), 
and to deal with merging and metathesis. Connolly also considers briefly the effects 
of nonlinear prosodic structure on the distance measure. Although his methods are 
clearly computational in nature, Connolly reported (personal communication, 1997) 
that he had not yet implemented them. Taken together, these measures are certainly 
more sophisticated than either CAT's or Covington's, so this contribution could well be 
an extremely significant one towards the development of articulation testing software. 
In Somers (1998), I report an implementation and comparison of Connolly's measures 
with my own earlier work. 
4.2 What Would a New Version of CAT Be Like? 
In the light of the above remarks, it is interesting to think about how we might specify 
a reimplementation of CAT. One area where there could be considerable improvement 
is in the data input. CAT uses a very crude phonetic transcription based only on a 
minimal character set, not even including lower-case letters. Clearly this restriction 
would not be necessary nowadays. The software system PDAC (Phonological Devia- 
tion Analysis by Computer) uses a software package called LIPP (Logical International 
phonetic Programs) for input of transcriptions (Perry 1995). Alternatively, it seems 
quite feasible to allow the transcriptions to be input using a standard word processor 
and a phonetic font, and to interpret the symbols accordingly. For a commercial im- 
plementation it would be better to follow the standard proposed by the IPA (Esling 
and Gaylord 1993), which has been approved by the ISO, and included in the Unicode 
definitions. 
Despite the reservations of all the speech-language pathology experts, it seems 
to me that the work on alignment discussed here (Somers 1978b; Covington 1996; 
Connolly 1997) suggests that this aspect of computerized articulation test analysis is a 
research aim well worth pursuing, especially if collaborators from the speech-language 
pathology field can be found. It would be rewarding if this article were to awaken 
interest in the problem. 
Acknowledgments 
I would like to thank the following people 
for their help in gathering the information 
presented in this paper: Catherine Adams 
(University of Manchester), Lawrence 
Shriberg (University of Wisconsin-Madison), 
Julie Masterson (Southwest Missouri State 
University), Carol Stoel-Gammon 
(University of Washington) and John 
Connolly (Loughborough University); 
Michael Covington, for collaborating on the 
"bake-off"; Joe Somers, for providing some 
of the example data; and the three 
anonymous reviewers for their suggestions, 
which have been extremely valuable. 
273 
Computational Linguistics Volume 25, Number 2 
References 
Armstrong, Lilias E. 1967. The Phonetics of 
French: A Practical Handbook. G Bell, 
London. 
Ball, Martin J. 1994. Review of ISPA: 
Interactive System for Phonological 
Analysis. Child Language Teaching and 
Therapy, 10:97-102. 
Bardzik, Paul P. 1986. Review of Computer 
Managed Articulation Analysis. ASHA, 
28(2):74--75. 
Cochran, Paula S. & Julie J. Masterson. 1995. 
NOT using a computer in language 
assessment/intervention: In defense of 
the reluctant clinician. Language, Speech, 
and Hearing Services in Schools, 26:213-222. 
Connoll~ John H. 1997. Quantifying 
target-realization differences. Clinical 
Linguistics & Phonetics, 11:267-298. 
Covington, Michael A. 1996. An algorithm 
to align words for historical comparison. 
Computational Linguistics, 22:481--496. 
Curtis, Jack F. 1987. An Introduction to 
Microcomputers in Speech, Language, and 
Hearing. College-Hill, Boston, MA. 
Dyson, Alice T. 1987. Review of PEPPER: 
Programs to Examine Phoentic and 
Phonological Evaluation Records. Child 
Language Teaching and Therapy, 3:329-335. 
Esling, John H. & Harry Gaylord. 1993. 
Computer codes for phonetic symbols. 
Journal of the International Phonetic 
Association, 23:83-97. 
Faircloth, Marjorie A. 1971. 
Computer-assisted articulation analysis. 
Paper presented at the Annual 
Convention of the American Speech and 
Hearing Association, Chicago, IL, Sept. 
1971. ASHA, 13:534. 
Gregg, Ellen Meyer. & Valorie Andrews. 
1995. Review of Computerized Profiling 
(1993). Child Language Teaching and Therapy, 
11:209-216. 
Ingram, David. 1976. Phonological Disability 
in Children. Edward Arnold, London. 
Kennedy, Kathleen A. 1986. Review of 
Computer Analysis of Phonological 
Processes. ASHA, 28(8):71. 
Klee, T. & E. Sahlie. 1994. Review of 
Computerized Profiling Version 1.0. Child 
Language Teaching and Therapy, 3:87-93. 
Leonard, Laurence B. 1989. Review of the 
Pye Analysis of Language. Child Language 
Teaching and Therapy, 5:79-86. 
Long, Steven H. 1991. Integrating 
microcomputer applications into speech 
and language assessment. Topics in 
Language Disorders, 11(2):1-17. 
Long, Steven H. & Julie J. Masterson. 1993. 
Computer technology: Use in language 
analysis. ASHA 35(9):40--41,51. 
MacWhinne~ Brain. 1992. The CHILDES 
Project: Tools for Analyzing Talk. Lawrence 
Erlbaum Associates, Hillsdale, NJ. 
MacWhinney, Brian & Catherine Snow. 1985. 
The child language data exchange system. 
Journal of Child Language, 12:271-296. 
Masterson, Julie J. 1995. Computer 
applications in the schools: What we can 
do--what we should do. Language, Speech, 
and Hearing Services in Schools, 26:211-212. 
Miller, Jon F. & Thomas Klee. 1995. 
Computational approaches to the analysis 
of language impairment. In Paul Fletcher 
& Brian MacWhinney, editors, The 
Handbook of Child Language. Blackwell, 
Oxford, pages 545--572. 
Perr~ Cecyle K. 1995. Review of 
Phonological Deviation Analysis by 
Computer (PDAC). Child Language Teaching 
and Therapy, 11:331-340. 
Pollock, Karen E. 1988. Review of PEPPER: 
Programs to Examine Phonetic and 
Phonological Evaluation Records. ASHA, 
30(8):57-58. 
Pye, Clifton & David Ingram. 1988. 
Automating the analysis of child 
phonology. Clinical Linguistics & Phonetics, 
2:115-137. 
Rushakoff, Gary E. 1984. Clinical 
applications in communication disorders. 
In Arthur H. Schwartz, editor, The 
Handbook of Microcomputer Applications in 
Communication Disorders. College-Hill 
Press, San Diego, CA, pages 147-171. 
Rushakoff, Gary E. & Arthur H. Schwartz. 
1986. Clinical assessment software. In 
Michael L. Grossfeld & Cathleen A. 
Grossfeld, editors, Microcomputer 
Applications in Rehabilitation of 
Communication Disorders. Aspen, 
Rockville, MD, pages 1-24. 
Schwartz, Arthur H. 1984. Microcomputer 
applications: Facts, functions, fads, and 
fallacies. Journal of Childhood 
Communication Disorders, 8:89-111. 
Silverman, Franklin H. 1987. Microcomputers 
in Speech-Language Pathology and Audiology: 
A Primer. Prentice-Hall, Englewood Cliffs, 
NJ. 
Sokolov, Jeffrey L. & Catherine E. Snow, 
editors, 1994. Handbook of Research in 
Language Development Using CHILDES. 
Lawrence Erlbaum Associates, Hillsdale, 
NJ. 
Somers, H. 1978a. Computer analysis of 
speech therapists' articulation tests. 
UMRCC Journal, 5(1):9-17 (University of 
Manchester Regional Computing Centre). 
Somers, H. L. 1978b. Computerised 
Articulation Testing. M.A. thesis, 
Department of General Linguistics, 
274 
Somers Aligning Phonetic Segments 
University of Manchester. 
Somers, H. 1979. Using the computer to 
analyse articulation test data. British 
Journal of Disorders of Communication, 
14:231-240. 
Somers. H. 1998. Similarity metrics for 
aligning children's articulation data. In 
Proceedings of COLING-ACL '98: 36th 
Annual Meeting of the Association for 
Computational Linguistics and 17th 
International Conference on Computational 
Linguistics, Montreal Quebec, Canada, 
pages 1227-1232. 
Stoel-Gammon, Carol & Paula Beckett 
Herrington. 1990. Vowel systems of 
normally developing and phonologically 
disordered children. Clinical Linguistics & 
Phonetics, 4:145-160. 
Telage, Kal M. 1980. A computerized 
place-manner distinctive feature program 
for articulation analyses. Journal of Speech 
and Hearing Disorders, 45:481-494. 
Van Demark, D. R. & Rosemary Tharp. 1973. 
A computer program for articulation 
tests. Cleft Palate Journal, 10:378-386. 
275 

