Phonetic variation occurs naturally in conversational speech, both
within and across talkers. Indeed, to the extent that no two words are
ever acoustically identical, variation occurs in all speech. Extensive
research on variation has documented the influence of multiple factors on
variation, including phonological context, properties of words, as well as
a range of extra-linguistic variables, such as age, gender, and
socioeconomic status. However, little is known about how factors relating
to language usage influence variation, or how variation arises. In fact,
for most phenomena, even information on the relative frequencies of
specific words can be difficult to come by.
In this study, we analyze one common variation phenomenon in
spontaneous American English speech: the deletion of the alveolar stops
/t/ and /d/ when they occur word-internally, that is, in
any position other than the beginning or end of an orthographic word, as
in stop, better, advice, or it's.
Our study examines numerous factors affecting internal alveolar stop
deletion, and we use the analytic results as a means of understanding the
nature of /t,d/ deletion. In modeling deletion, we assume the
notion of a variable rule (Cedergren & Sankoff,
1974), in which the probability of a phonetic output is a function
of a number of contextual factors. The contribution of a factor to
predicting deletion is assessed statistically with logistic regression in
the complete dataset and partial datasets, or by chi-square tests in some
smaller subsets. Using statistical tests of significance in a variable
rule framework allows us to quantify the nondeterministic nature of the
variation contingent on multiple factors and their interactions.
The data analyzed in this study were taken from the Buckeye
corpus of spontaneous interview speech, which will be described in more
detail later. The decision to use spontaneous speech was motivated by a
number of factors, including the observation that spontaneous speech data
reveal a richer range of phonetic variation than is typically found in
laboratory speech (Keating, 1998), and that
variation generally occurs in a broader set of environments in spontaneous
speech than in read speech. The focus on /t,d/ deletion was due in
part to practical considerations. Both /t/ and /d/ are
frequent sounds in the English lexicon and occur in many very frequent
words, making it possible to test usage-based factors. These stops are
also produced with considerable variability in spontaneous speech. In the
Buckeye corpus, /t/ and /d/ occurred in the
canonical, or dictionary, pronunciation of approximately 40% of all word
tokens in the corpus. Furthermore, of the words canonically containing
alveolar stops, 45% of all tokens resulted in a phonetic realization other
than [t] for /t/ or [d] for /d/, and
deletion was a common realization of the stops, occurring in 16.5% of the
tokens. High deletion rates for /t/ and /d/ are consistent
with findings from studies of word- (and stem-) final alveolar stops
(e.g., Guy, 1980; Neu,
1980).
Another key reason for choosing to study word-internal /t,d/
is that there already exists an extensive literature on word-final
alveolar stop deletion that has used similar methodological approaches
(Fasold, 1972; Guy, 1980, 1997; Jurafsky et al., 2001; Labov,
1967; Neu, 1980; Wolfram, 1969, among others). Factors shown to be significant
in the study of word-final deletion were also considered in the present
study. Additional factors were identified from the extensive literature on
variation in spontaneous speech (see later discussion). Although our
results confirm that many of the factors shown to influence word-final
stop deletion also influence internal deletion, important differences also
emerged, which was expected given the differing properties of the two
environments. For example, although for a word-internal segment the
distribution of sounds on both sides of a stop is phonologically
constrained, for a word-final stop only the preceding context is.
Moreover, word-final stops are typically also syllable-final, unless they
are resyllabified in connected speech. By examining word-internal
sequences we were thus also able to study the effect of syllable position
on deletion more directly. As we will show, a number of crucial
differences in the deletion patterns emerged depending on whether the stop
was in coda or onset position. Besides the expected finding that a stop
deletes more often in coda than in onset position, there are also
differences based on position of phonological environment, unit
frequencies, prominence, and speaker age.
A goal of this study is to use knowledge of variation patterns to
explain the origins of deletion. Deletion is generally considered a type
of reduction, but it has been alternatively explained as resulting from
segment lenition or cluster simplification, depending on phonological
environment. Simplification is proposed for segments in consonant
clusters. In a lenition account, Rhodes (1992)
noted that deletions occur in flapping environments, and thus proposed
that flaps and deletions are both outcomes of a single lenition process,
with deletion being a more extreme form of lenition. Zue and
Laferriere's (1979) study of read speech
found some deletion in the post-nasal flapping environment, providing some
support for this view. Although the amount of deletion that Zue and
Laferriere found in read speech was limited, we would expect deletion that
is part of a spectrum of gradient lenition to be more likely in fluent
spontaneous speech, such as faster speech, because greater fluency is
associated with segment shortening (Byrd & Tan,
1996). On the other hand, when contextual factors strengthen one
outcome, such as flapping, deletion would be less likely. As a means of
coming to a better understanding of deletion as a potential type of
lenition, throughout the course of this study deletion is compared to the
alternate alveolar stop process of flapping.
Because lenition and simplification have been proposed for specific
segmental environments, it is reasonable to assume that the primary
correlates of deletion can be found in phonological contexts. If so, then
processes sensitive to context, such as speech planning and articulation,
will underlie these reductions. This observation suggests an approach to
explaining deletion that is based on the language production process. To
anticipate our findings, evidence did indeed emerge in support of the view
that two different deletion processes are involved: one finding its
explanation in consonant cluster simplification and the other in segment
lenition. Moreover, the two processes are argued largely to arise at
different stages of speech production, with the consequence that the two
processes are differentially affected by factors that influence deletion,
as we will show.
We begin the article with a description of the Buckeye corpus
and the dataset used in our study. We then introduce the methods and
details of our variable coding, drawing on the results of previous studies
on alveolar stop deletion to motivate the selection of specific factors
and to make predictions of their effects on word-internal deletion.
Results of the analyses are then presented for the extra-linguistic,
lexical, and phonological variables. Finally, we present a summary of our
results and interpret them in terms of a model of speech production.
THE DATASET: THE BUCKEYE
CORPUS
The data analyzed in this study were created from a lexically and
phonetically transcribed subset of the Buckeye corpus of
spontaneous speech (Pitt et al., 2005). The
Buckeye corpus contains over 300,000 words of speech from 40
individual speaker interviews. Speakers were asked to express opinions on
a variety of topics for about an hour. The speaker sample for the corpus
was stratified for age (under 35 and over 40) and gender, and all speakers
were natives of central Ohio.
The transcribed subset of the corpus consisted of about 100,000 words
from 14 of the corpus speakers. The transcribed speech from the 14
speakers comprises all or the vast majority of each speaker's
complete interview (in a continuous interval from the beginning of the
interview). Interviews from 7 male (2 older and 5 younger) and 7 female (2
older and 5 younger) speakers were included in the subset.
The source of the dataset taken from the transcribed subset consisted
of the set of all orthographic words in the transcribed subset whose
dictionary pronunciation contained an internal /t/ or /d/.
Internal /t,d/ tokens were defined as /t/ or /d/
phones in complete words that were neither the first nor last phone in a
word's dictionary pronunciation. Internal /t,d/ tokens may
thus be in initial onsets of words if preceded by /s/ (e.g.,
still, stable) or in final codas of words if followed by
/s/ (e.g., that's, kids), as well as in
onset and coda syllable positions elsewhere in words (e.g.,
better, advice). Dictionary pronunciations were taken
from the Buckeye corpus dictionary, which consists of over 66,000
lexeme entries based on the CELEX English dictionary lexicon (Baayen, Piepenbrock, & van Rijn, 1993) augmented
to include all vocabulary in the corpus and with Standard American
pronunciations.
Phonetic labeling and label alignment with the speech signal were
performed using a combination of automated and manual transcription
procedures by trained transcribers. Transcribers labeled and aligned
speech with the aid of spectrograms of the speech signal. In transcribing
the internal /t,d/ phones, transcribers used 18 different segment
labels. Of these 18 labels, only five labels were selected by transcribers
for more than 2% of the /t,d/ tokens in the transcribed subset.
The five frequently used variant labels were [t], [d],
flap, nasal flap, and glottal stop. (The [t] and [d]
variants were not further subcategorized with respect to other features,
such as the presence or absence of aspiration or release.) Internal
alveolar stop realizations other than the five most frequent variants were
excluded from the final dataset. The excluded outcomes were largely
assimilations (e.g., “abministration” for
administration). Deletions (i.e., no transcribed label) accounted
for 16.5% of the realizations of canonical internal alveolar stops, making
deletion the fourth most likely realization of word-internal /t/
or /d/ phones.
In the phonetic transcription protocol for the corpus, deletion was
defined as the absence in the speech signal of any acoustic evidence for
some segment realization, with clear segment boundaries, corresponding to
/t/ or /d/ in the dictionary pronunciation of a word.
Intervocalically, stop outcomes were recognized by a perceptible closure,
perhaps with an accompanying burst, in a spectrogram of the token. Oral
and nasal flaps are quite short and generally lack a release burst, but
result in a brief closure or a significant disturbance in the formants of
adjacent vowels that clearly distinguishes them from deletion. When a
/t/ or /d/ variant was adjacent to another stop consonant
or a silence, it could be identified as a stop (alveolar or glottal) based
on details of formant transitions, the presence of a release burst,
and/or closure length.
A study of the transcription consistency in the corpus indicated that
agreement on labeling for internal /t,d/ tokens overall in the
Buckeye corpus was seen in 81% of transcriber pairs (Raymond et al., 2002). This agreement rate is
comparable to other studies of consistency in labeled spontaneous speech.
Raymond et al. (2002) found differences in
labeling consistency among the labels used for lexical alveolar stops.
Most notably, there was little consistency on glottal transcription, with
transcribers agreeing that the realization of a canonical /t/ or
/d/ was a glottal stop only 33% of the time. Although most
disagreement in the study involved glottal stop and [t] (and
never deletion), tokens transcribed with glottal stops were excluded from
the final dataset. Exclusion of glottals was due to the desire to consider
variables based on knowledge of contexts that promote alternate
realizations, which would have been unreliable for glottalizing
environments. Unlike the transcription of glottals, transcribers agreed on
identification of deletions (vs. all other variants) in 80% of transcriber
pairs, making deletion more consistently transcribed than [t]
(78% agreement) or [d] (51% agreement in all environments).
Although agreement on word-internal [d] transcription was low,
there was disagreement between [d] and deletion in only 2% of
transcriber pairs. There were similar low levels of disagreement between
deletion and other noncanonical labels. Labelers disagreed between
deletion and nasal flapping in only 5% of transcriber pairs, and between
deletion and oral flapping in only 4% of transcriber pairs. Transcription
of deletions was thus generally consistent and reliable.
After excluding tokens with infrequently used labels and glottal
labels, the transcribed subset yielded a dataset of 7,241 internal
/t,d/ tokens (2,014 /d/ tokens and 5,227 /t/
tokens).The distribution of the five variant outcomes (including deletion)
in the final dataset is shown in Table 1.
Distribution of transcribed variants of canonical internal
/t,d/ phones
VARIABLE IDENTIFICATION AND CODING
Factors examined in the study include extra-linguistic and lexical
variables, and a range of phonological influences. For each class of
variables, identification of factors will be motivated by previous
studies. Discussion of previous work is followed by the details of our
selection and coding of variables in the dataset. Predictions for their
influences on medial deletion are found at the end of the section.
Extra-linguistic variables
Speaker group: Age and gender
In Guy's (1992) study of word-final
/t,d/ deletion, older speakers deleted word-final stops less often
than younger speakers. Furthermore, speaker group differences were found
in Wolfram's (1969) study of final
deletion, in which there were more deletions among men than women.
Neu's (1980) study of word-final alveolar
stop deletion found that the relative tendencies of a preceding sibilant,
stop, or nasal consonant to promote deletion differed according to the
speaker's gender, with men showing significantly more deletion after
sibilants than after nasals or stops, but women having no significant
difference in deletion rates among the three classes. However, in the
study by Zue and Laferriere (1979) of
/t,d/ variation in flapping environments, there was more
/t/ deletion in a nasal flapping environment (VntV) by females
than by males. The full range of preceding consonant types examined by Neu
was not tested by Zue and Laferriere, however, so it is difficult to
compare the results of the two studies.
Given the results from these studies, speaker gender and age were
analyzed in the present study. Age was coded both as a continuous variable
(age in years) and also as a binary variable (older, younger), based on a
speaker's age category in the Buckeye corpus. Gender was
coded as a binary variable (male, female).
It should be noted that other social factors, such as dialect or
social class, were not analyzed in this study despite the importance of
such factors to understanding variation in the realization of alveolar
stops (e.g., Labov, 1967; Labov & Cohen, 1967; Labov et
al., 1968; Wolfram, 1969). The reason is
that many dimensions of the social group were controlled in the design of
the Buckeye corpus by limiting participants to a small geographic
area and excluding nonstandard dialects.
Speech rate and speech fluency
Speech rate is known to affect many reduction phenomena, including
final /t,d/ deletion (Fosler-Lussier &
Morgan, 1999; Guy, 1980; Jurafsky et al., 2001; Labov &
Cohen, 1967; Labov et al., 1968; Wolfram, 1969). Guy (1980)
noted that, impressionistically, word-final /t,d/ deletion
increased with speech rate. Fosler-Lussier and Morgan (1999) found that the rate of segment deletion
increased from 9.3% in very slow conversational speech to 13.6% in very
fast speech. Higher deletion rates at faster speech rates would be
consistent with articulatory studies showing that faster speech is
associated with segment shortening and gestural overlap (Byrd & Tan, 1996), assuming that deletion is the
end result of shortening and so is also sensitive to rate.
A number of measures were developed to investigate effects of local
speech rate on deletion. An adequate measure should average over a domain
large enough to estimate rate independently of segment deletion while
ensuring reasonably local scope. Consequently, rate measures were computed
over four domains: (1) the /t,d/ word; (2) a three-word window
centered on the /t,d/ word; (3) a five-word window centered on the
/t,d,/ word; and (4) the pause-bounded utterance containing the
/t,d/ word. The duration of speech units were calculated from the
phonetically annotated and aligned speech signal. Although there are a
number of different units over which rate could be calculated, the
syllable was chosen because it is less likely to be deleted in speech than
the phone (Fosler-Lussier & Morgan, 1999),
and so the canonical syllable count can be used.
Note that measures of local rate do not take into account
speakers' overall rate differences, which ranged from 4.77
syllables/sec to 7.20 syllables/sec. To accommodate for these
potential differences, we created a binary variable that encoded relative
rate. Each token's three-word rate was categorized relative to the
speaker's mean speech rate (fast, slow).
Many studies have reported effects of dysfluent production on deletion
and other reduction phenomena. Dysfluency types that have been reported to
affect deletion include adjacent pauses, fillers (e.g., uh or
um), word repetitions, and word cutoffs. A dysfluency following a
word has generally been reported to result in durational lengthening,
fuller phonetic forms, and the inhibition of final segment deletion (Fox Tree & Clark, 1997; Jurafsky et al., 1998; Shriberg,
1999; Van Santen, 1992). In a study of
function word reduction in spontaneous speech, Jurafsky et al. (1998) found that the final /d/ of and
was more likely to be present when the word was followed by a repetition
of and, a filled pause (uh or um), or a silence
than when it was not. They found no effect, however, on the likelihood of
deletion in dysfluent contexts of the final /t/ in it or
that, suggesting that fluency may differ with word function, word
identity, or perhaps alveolar stop identity. In some sociolinguistic
studies, a following silence promoted deletion (Fasold,
1972; Labov, 1967), but in other studies
it inhibited deletion (Wolfram, 1969). Differing
results of the effect of a following silence reflected dialectal
differences in Guy (1980). For New Yorkers in
his study, a following pause promoted deletion, but for Philadelphians a
pause inhibited deletion.
To investigate the influence of dysfluent speech, tokens in the
dataset were coded using binary variables (fluent, dysfluent) for four
types of dysfluency with preceding lexical context and following lexical
context. The variables indicate whether or not a word containing an
internal /t,d/ token was preceded or followed by: (1) a filled
pause (i.e., uh or um, e.g., “a a former uh
basketball player”); (2) a word repetition (e.g.,
“that's that's really hard to say um”); (3)
a lexical cutoff (e.g., preceding, “his po- [position]
title was”, and following, “… was pretty
moderate actually and it is pretty mod- [moderate]
people …”); or (4) a pause (e.g., “〈silence〉
absolutely no meaning”).
Predictions for word-internal alveolar stop
deletion
Based on previous studies, we anticipate that overall deletion rates
will be higher in more fluent speech, where fluent speech is defined as
being faster and with fewer pauses, fillers, repetitions, and word cutoffs
than less fluent speech. The effects of fluency should be seen in those
phonetic environments in which deletion results from gestural shortening,
but may not be as strong in environments where deletion is not a
consequence of gestural shortening. There may also be an effect of gender,
with males deleting more than females, and perhaps an effect of age, with
younger speakers having higher deletion rates than older speakers. Age and
gender differences may themselves ultimately be attributable to
differences in fluency or variant choice between speaker groups. The
origins of group differences can be explored by controlling for other
fluency factors when performing group comparisons.
Lexical variables
Word form
There is evidence that suggests word-level effects on final
/t,d/ deletion. Neu (1980) found greater
rates of deletion of /d/ in the word and than would be
expected from the phonetic context of the final /d/. Labov (1975) found similar high deletion rates of the
/t/ in the word just. Note that the effects may be
word-specific, or may be generalizable to wider subsets of the lexicon
based on word structure or word use.
At the broadest level of word classification, a distinction based on
use is often made between (closed-class) function words and (open-class)
content words, although this classification is confounded with word
structure and word probabilities. The two classes are generally
distinguished by semantics and productivity. Function words, unlike
content words, have low semantic content, and new members are not readily
admitted to the set. Function words are also common, generally short, and
usually unaccented in production. For these reasons, they are generally
less salient than content words (Pollack & Pickett,
1964), and are often cliticized to content words in connected
speech, signaling their susceptibility to reduction, including deletion.
On the other hand, content words are frequently multisyllabic and more
likely to be accented in production, resulting in fuller forms.
To tease apart various aspects of word form, tokens of internal
alveolar stops were categorized in the present study by a binary
classification of the word containing the token (function, content).
Function words included personal and indefinite pronouns, prepositions,
articles, modal and auxiliary verbs, and wh-words. All other
words were classified as content words. Tokens were also coded for length
(in canonical syllables) of the word in which they occurred. We also took
into account the relative position of the internal /t,d/ token
within the word, which was coded as one continuous variable and two binary
variables. The continuous position variable was calculated as the syllable
in which the token occurred (counting from the left edge of the word)
divided by word length. For example, the length of better is 2,
and the relative position of the canonical /t/ in the words is 1
(= 2/2, being in the second syllable of a two syllable word). The
binary position variables coded: (1) whether the token occurred in the
first syllable of a word or a subsequent syllable, and (2) whether it
occurred in the last syllable of a word or an earlier syllable.
Word probabilities
Although word frequencies, which may be considered nonconditional word
probabilities, and other conditional word probabilities are all highly
correlated with word class and word structure, previous studies have
established effects of word probabilities on a variety of processes in
spontaneous speech, including deletion, that appear to be independent of
word class and many other factors (Bybee, 2000;
Jurafsky et al., 2001). Jurafsky et al. (2001) found that after controlling for many correlated
factors, more frequent content words had higher rates of final alveolar
stop deletion than less frequent words. Final /t,d/ deletion was
also found to be encouraged in content word pairs with high word bigram
frequencies, that is, when a content word ending in a /t/ or
/d/ and either of its neighbors in the speech context were likely
to occur together. Jurafsky et al. (2001) also
reported that the word probabilities examined in their study affected word
duration, with more likely words being durationally shorter. Importantly,
they showed that durational shortening was not simply the result of final
segment deletion, so that word probabilities may influence phonetic
reduction within the word and not just at word ends.
Word frequency measures were coded for each token in the dataset from
two sources, the CELEX English corpus counts, and the Buckeye
corpus. Raw counts of words from both corpora were log transformed for
analysis. The two log frequencies are correlated (r2 =
.67), but because of its larger text base, the CELEX measure may be more
discriminating. However, CELEX counts are based largely on text counts
rather than spontaneous speech, and counts are not available for some
words in the Buckeye corpus, in particular, proper names and
nouns ending with 's (n = 336; e.g.,
Mauldin and kid's). Although Buckeye token
counts are available for the words of all /t,d/ tokens in the
dataset, the counts of many words are low, with 428 words (5.9%) occurring
only once (and thus estimated to occur only about three times per million
words of speech). As a result, a combined measure of frequency per million
was created from the Buckeye counts and the CELEX counts, and
this measure, log transformed, was used in the analyses.
Frequency measures of word bigrams, or pairs of words that occurred
contiguously, were calculated from the Buckeye corpus. Word
bigram counts were calculated from the full (300,000 word)
Buckeye corpus by counting the fluent word pair types (i.e.,
unique pairs). A fluent pair was defined as any two consecutive
orthographic words not separated in speech by a perceptible silence. Using
the bigram counts, a preceding and a following word bigram frequency
measure were coded for each /t,d/ token that occurred in a word
not adjacent to a silence. Bigram frequencies were undefined for words
adjacent to a silence.
Bigram counts were also calculated for dysfluent environments. A
dysfluent pair was defined as a word preceded or followed by a filled
pause or a word cutoff, where the dysfluent element was not separated from
the word by a silence. Using these bigram counts, preceding and following
word bigram frequencies were coded for all /t,d/ tokens preceded
or followed by a dysfluency.
The predictability of a /t,d/ token word based on the
immediately preceding or following word was also tested. Predictability
measures were calculated using the word and the word bigram frequencies
described earlier. Specifically, the predictability of a word w
from an adjacent word w′ was estimated using (1). For
example, the predictability of let's from the preceding word
in the context so let's would be the word bigram frequency
of so let's divided by the frequency of so.
The predictability of a word preceding or following a silence is
undefined. As with word frequencies, word predictabilities were log
transformed to create variables used in the analyses.
Morphological structure
In studies of word-final /t,d/ deletion, independent
morphological status of /t/ or /d/ was shown to inhibit
deletion (Guy, 1980, 1992; Labov et al., 1968;
Neu, 1980). When a final stop is the realization
of a past tense morpheme (e.g., missed) there is less deletion
than when it is not (e.g., mist). Irregular past tense forms
(e.g., left) have an intermediate rate of deletion. Although past
tense morphology will not apply to internal alveolar stop segments, other
word morphology may, as where an alveolar stop is stem-final but not
word-final (e.g., writing, let's). Neu (1980) included stem-final stops in her study, but
found no difference in deletion rates between stem- and word-final stops.
However, Hay (2000) found less deletion of
stem-final /t/ in words whose stems were more transparently
related to the morphologically complex form (e.g., swiftly) than
in words with a less transparent relationship between stem and word (e.g.,
listless), even after controlling for frequency of word and stem.
Morphological structure is not controlled for in our study.
Predictions for internal stop deletion
Function words and more likely words have been shown to be associated
with higher final alveolar stop deletion rates, and both have been
associated with shorter word durations. This suggests that processing of
these words may promote internal stop deletion, as well, assuming that
these factors are not limited in their effects on word ends but influence
production parameters across the word. However, the correlation between
word class and word probabilities may make distinguishing the underlying
source of deletion difficult. Moreover, because function and content words
differ in their segmental and metrical phonology, interactions with
phonological variables are likely, and an effect of word class may
ultimately be the result of the structural properties of words and not
fundamental processing differences between the two classes. In addition,
previous studies have shown a correlation between phonological structure
and word frequency (Frauenfelder et al., 1993;
Landauer & Streeter, 1973), so that effects
of word probabilities may be ultimately explained more locally in terms of
the probabilities of their phones or phone sequences, which will be
discussed in the next section. These more local phone probabilities have
more transparent consequences for articulation, but word probabilities are
more easily interpreted as influencing lexical access and speech planning.
Finally, there may be effects of morphological structure on deletion, with
more deletion in words that are less readily semantically decomposed, in
cases where morphological structure is a factor.
Phonological variables
Syllable position
Studies of word-final stop deletion indicate that deletion is high
when there are flanking consonants, and unless the stop is resyllabified
as onset of the following word-initial syllable, word-final stops occur in
syllable codas. Consistent with this, Greenberg (1999) found that productions of onset consonants
resulted in less variation in spontaneous speech than productions of coda
consonants. Because segmental context and syllable structure are
correlated, explanations of deletion involving specific phonological
contexts and morphological structure also implicate the role of syllable
position in deletion.
To examine the effects of syllable structure on deletion, tokens of
/t,d/ were coded for syllable position (onset, coda) based on the
CELEX dictionary syllabification of canonical pronunciations that were
modified to Standard American pronunciation. Adjustments were made to make
syllable structure consistently a function of phonological context and not
morphological structure. In particular, alveolar stops whose dictionary
syllabification placed them in coda position, but whose following context
was a vowel, a liquid, or a glide that allowed for resyllabification of
the stop, were coded as onsets (e.g., wait.ing >
wai.ting, had.n't > ha.dn't,
broad.way > broa.dway). Taking into account these
adjustments, the possible onset and coda values coded for the analyses in
all CV contexts are as in Table 2. Note that the
set defined by C can differ by context in keeping with English
phonotactics. For example, onsets can only be followed by the consonants
/w, r/.
Possible consonant–vowel combinations in onsets and
codas
Prominence
The importance of prominence factors, such as stress and accent, is
indicated by findings showing that stress-accented syllables exhibit less
variation and reduction than unaccented syllables (Greenberg, Carvey, & Hitchcock, 2002), and that
the segments in these syllables are lengthened and strengthened (Turk & White, 1999). Studies of word-final stops
have also found that deletion is more likely in unstressed syllables than
in stressed syllables (Fasold, 1972; Labov, 1972; Labov et al.,
1968; Wolfram, 1969). Zue and Laferriere
(1979) found that a prevocalic /t/ is
shorter when it is the onset of an unstressed syllable than a stressed
syllable. Stop deletions only occurred, however, when the stop was also
preceded by a stressed syllable, suggesting that the stress level on the
syllable preceding a word-internal /t,d/ is important, as well. In
these studies, we can interpret lexical stress as prosodic prominence,
because of the association of pitch accents to lexically stressed
syllables in words produced in isolation.
In the present study, dictionary representations were used to code
three levels of prominence (stressless, secondary stress, primary stress)
on the syllable containing the canonical alveolar stop and for the
preceding and following syllables. Because of the low proportion of
secondarily stressed syllables (15.6% of the syllables containing an
internal /t,d/ segment), a binary stress variable (stressed,
unstressed) was also coded by conflating primary and secondary stress
levels. Finally, note that monosyllabic function words are lexically
stressed, but are less likely to be accented in production than content
words. In fact, an examination of one hour of speech in the
Buckeye corpus coded for pitch accent revealed that content words
were 50% more likely than function words to have an accent of any kind. A
second binary prominence variable was thus coded that conflated primary
and secondary stress levels for all syllables except for the monosyllabic
function words, which were coded as “unstressed” to reflect
the observation that they are less likely than content words to receive
prominence in production. The assumption that a monosyllabic function word
is unaccented may be thought of as a predictive variable to which speakers
are sensitive in a manner similar to their sensitivity to word
probabilities. Note that defining stress levels in this way could also be
handled as an interaction of stress and word class.
Alveolar segment identity
There is some evidence that /t/ and /d/ may pattern
differently with respect to deletion. Zue and Laferriere (1979) found articulatory differences between
/t/ and /d/ released into a nasal (e.g., sweeten,
Sweden). A tendency for /t/ in this environment to be
produced without a burst suggests that /t/ is more likely to be
deleted than /d/. Conversely, Rhodes (1992) noted that /d/ is more likely to flap
than /t/, suggesting that /d/ is also more likely to
delete than /t/ in flapping environments. To examine the effects
of alveolar stop identity on deletion, tokens in the dataset were coded
based on canonical stop identity using a binary variable (/t/,
/d/).
Following segmental context
Studies of word-final alveolar stop deletion have shown that a
following consonant promotes deletion of final alveolar stops, but a
following vowel inhibits it (Guy, 1980; Labov, 1967; Labov & Cohen,
1967; Labov et al., 1968; Neu 1980; Wolfram, 1969). The
tendency for following consonants to promote deletion of final alveolar
stops over vowels is consistent with the more specific claim that the less
sonorous the following segment, the greater the likelihood of deletion,
assuming the sonority ranking: obstruent > liquid > glide > vowel
(Fasold, 1972; Guy,
1980; Labov, 1972, 1975; Labov et al., 1968;
Neu, 1980; Wolfram,
1969). Guy (1991) and Labov (1995) noted, however, that for following liquids and
glides the sonority of the following segment cannot account for deletion
patterns. Both studies observed that the probability of final stop
deletion was much higher before /l/ than before /r/,
although the sounds are of comparable sonority. In fact, deletion before
/l/ approaches the high level for obstruents in Labov's
study, and it exceeds the rate of deletion before obstruents in Guy (1991). Labov also found the glide /w/ to
behave like an obstruent in promoting deletion, although deletion before
/y/ is very low, and before /h/ it is intermediate.
There is also some evidence that the place of articulation of a
following consonant affects final alveolar stop deletion. Fasold (1972) found that if the consonant following an
alveolar stop was homorganic, the deletion rate was higher than if it was
nonhomorganic. In Neu's (1980) study, a
following consonant's place was not significant, however, leaving the
role of place uncertain.
Preceding segmental context
As with following context, studies have shown a clear difference
between preceding consonants and vowels in their effect on final alveolar
stop deletion. Although preceding consonants have been found to promote
deletion, for speakers of standard American English, preceding vowels have
been assumed to almost categorically inhibit deletion, given the absence
of post-vocalic deletion observed in the data used in previous word-final
stop deletion studies (Guy, 1992; Guy & Boberg, 1997). Postvocalic deletion was also
not observed in the Zue and Laferriere (1979)
study. However, they did find that the quality of the preceding vowel
influenced flap production. Vowels ending with high front tongue positions
(i.e., high off-glides, such as the diphthong
) were associated with longer flap durations than other vowels.
This class may thus inhibit deletion in flapping environments, because of
the strengthening of the flap segment.
In general, preceding obstruents (limited to stops and fricatives)
have been seen to favor deletion, although sonorant consonants, such as
nasals and liquids, inhibit it (Guy, 1980, 1997; Labov et al., 1968;
Neu, 1980; Wolfram,
1969). For most speakers in the study by Guy (1980), the rate of deletion decreased in the order:
sibilant > stop > nasal > liquid (> vowel = 0). Note that the
strength of manner effects on deletion is ranked roughly the same in
preceding and following contexts, with greater sonority associated with
less deletion.
Fasold's (1972) finding that a
homorganic following consonant promotes deletion also holds for a
preceding homorganic consonant (cf. Neu 1980).
Zue and Laferriere (1979) found that slightly
over 10% of internal alveolar stops are deleted after the homorganic nasal
/n/ (e.g., winter), the only environment in which they
observed deletion.
Coding segmental context
As we have seen, numerous differences have emerged in the patterning
of segments in the context of alveolar stops. Consequently, in the present
study consonants adjacent to /t,d/ tokens were coded by phone
identity rather than assuming a priori groupings based on broader
phonological categories (e.g., based on place and manner of articulation).
We chose this more detailed classification to allow for potential
similarities and differences in the effects of individual consonants to
emerge. To the extent that consonants patterned as natural classes, the
groups were subsequently coded and tested. This approach led to the
eventual creation of codes for categories of consonant place (homorganic,
nonhomorganic; labial, alveolar, palato-alveolar, velar, glottal) and
manner (sonorant, obstruent; stop, fricative, affricate, liquid, glide).
Adjacent canonical phones were also categorized as peak (syllabic) or
nonpeak (nonsyllabic). Syllabic segments included vowels, syllabic
/n/, and syllabic /l/. All other segments were coded as
nonsyllabic. The syllabic segments will be referred to hereafter as V or
syllabics; nonsyllabic segments will be referred to as C or
consonants.
Vowels preceding and following alveolar stops were categorized in two
ways. Following Zue and Laferriere (1979), one
classification of vowels was based on whether the vowel has a high
off-glide
or not (all others). Vowels were also categorized as nonlax
or lax (all others).
Phone bigrams
Labov (1995) proposed that the frequency
with which consonant clusters are learned may explain final stop deletion
rankings for some following segments. This suggests that a gradient
measure of phonotactic likelihood, such as phone bigram frequencies or
phone transitional probabilities, may be correlated with deletion
likelihood. These measures reflect acquisition order because more common
phone sequences are also more likely to be encountered earlier by language
learners. To our knowledge, nonstructural, probabilistic properties of
segments have not previously figured into studies of alveolar stop
deletion. We see these measures as potentially important, given
psycholinguistic research on speakers' use of probabilistic knowledge
in production and perception (Coleman &
Pierrehumbert, 1997; Frisch, Large, &
Pisoni, 2000; Luce, 1986; Makashay, 2001; Pierrehumbert,
1994; Pitt & McQueen, 1998; Pitt & Samuel, 1990; Saffran,
Newport, & Aslin, 1996; Savin, 1963;
Vitevitch & Luce, 1999). For example, Pitt
and McQueen (1998) found that the transitional
probabilities of voiceless coronal fricatives at the end of nonwords
influenced listeners' identification of an ambiguous fricative, as
well as that of the following stop consonant. We thus expect deletion to
occur at a higher rate in contexts where the alveolar stop is highly
predictable from its immediate phonological context.
As a means of studying the potential effect of the predictability of
sound sequences on stop deletion, two frequencies were coded for each
/t,d/ token, one for the frequency of the stop in combination with
the preceding segment, and one for the frequency of the stop and the
following segment. These bigram frequencies were estimated using the
corpus dictionary pronunciations of words and CELEX word frequencies.
Bigram frequencies were calculated by summing the word token frequencies
of all words in the corpus dictionary containing the bigram in its
pronunciation. To simulate the casual interview speech of the corpus, only
words with frequencies over 10 per million words were used. The result of
the estimations was a bigram count for each canonical phone before and
after /t/, and a count for each phone before and after
/d/. The most frequent preceding phone bigram was /st/
(e.g., listed, start),
accounting for 16.4% of all bigrams of a phone followed by an alveolar
stop of either kind. Other frequent preceding phone bigrams were
/nt/ (12.7%; e.g., hints,
enter),
(10.4%; e.g., it's,
limits), and /nd/ (7.6%; e.g.,
hundred,
nintendo). The most frequent following phone
bigram, with /t/ or /d/ as the first element, was
(9.7%; e.g., noticed,
stereotypical), followed by /ts/
(9.5%; e.g., hates, sorts) and
(8.1%; e.g., hated,
protestant). Preceding and following bigram
frequencies were log transformed for use in the analysis.
Phonological factors: Predictions for word-internal
/t,d/ deletion
Consistent with earlier studies, we expect to find phonological
context to play an important role in the word-internal deletion process.
In consonant contexts, word-internal deletion rates may be higher when
neighboring segments are homorganic, less sonorous, or more generally
share more features with the alveolar stop (Guy &
Boberg, 1997). Effects of phonological context must also include
investigation of syllable structure, although syllable structure and
segmental context are confounded to a certain degree because of
phonotactic constraints. Syllable position of a word-internal stop should
play a role in deletion, given reports of low rates of deletion in
intervocalic contexts (stops in onsets) but high rates of word-final
/t,d/ deletion in consonant contexts (stops most likely in codas).
In intervocalic contexts, features of the neighboring vowels may affect
deletion of the stop. Furthermore, deletion rates for /t/ and
/d/ may differ in flapping environments and in any other
environments where /t/ and /d/ are produced differently in
pertinent ways.
Earlier studies have shown that more probable words and word
combinations exhibit more variation (Bush, 2001;
Bybee, 1995; Bybee &
Scheibman, 1999; Cooper, Egido, & Paccia,
1978; Hooper, 1976; Jurafsky et al., 2001; Phillips,
1984). Thus, we expect /t,d/ deletion to occur at a higher
rate when the stop or the word containing it is highly predictable from
its immediate phonological context, although it is not clear which measure
will be significant.
Finally, lexical stress on a syllable containing an alveolar stop has
been shown to inhibit deletion of both final and some internal stops, most
likely because of its association with pitch accent in speech. We thus
expect deletion of word-internal alveolar stops to be inhibited in
stressed syllables. There is also the possibility that stress on the
syllable preceding the stop may influence deletion. Deletion was observed
to occur in stressed–unstressed syllable sequences, where flapping
is also possible; however, flapping does not appear to be dependent on
stress on the syllable preceding the stop, which may be the case for
deletion, as well.
We must also keep in mind that word probabilities have not been
investigated in combination with measures of phone likelihood, with which
they are correlated. In addition, sonority, syllable structure, and phone
probabilities are closely interrelated. A determination of which factors
are most explanatory for deletion may depend on a broader theoretical
perspective.
ANALYSIS AND RESULTS
General comments on the analyses
Analysis of data was performed in two steps. The contribution of a
factor to deletion was first assessed in the regression model without
controlling for other factors. For categorical variables, this was
accomplished by comparing the proportion of deleted and undeleted tokens
for all variable values. Deletion proportions were compared using a
chi-square test to identify significant differences in deletion rates. For
continuous variables, means of variable values for deleted and undeleted
tokens were compared using a t-test. Multiple logistic regression
was subsequently used to model internal /t,d/ deletion, allowing
us to identify independent contributions of factors after controlling for
other significant predictors of deletion. In logistic regression, the
outcome of a categorical variable, in this case the presence or absence of
/t/ or /d/, is statistically modeled based on any number
of other explanatory factors. Logistic regression is the basis of
varbrul analysis (Cedergren & Sankoff,
1974) and results in a comparable model of variation.
When factors in the study influenced subsets of the dataset
differently, the subsets were modeled independently, thus allowing us to
establish effects more clearly and avoiding complex interactions of
variables that might complicate interpretation. For example, variables
based on place and manner of articulation applied only to /t,d/
tokens adjacent to a consonant. Rather than model the dataset as a whole,
with factors representing the interaction of segment class and each
consonant class, variables were tested on subsets of tokens with preceding
(or following) consonants or vowels separately. Similarly, inclusion of
one factor in the model may exclude another, as where the speech rate
measure based on a three-word window of fluently produced words centered
on the /t,d/ token word excluded testing simultaneously for the
effects of adjacent silences. In these cases as well, different subsets of
the data were modeled separately. As a result, a variable will be reported
as significant for the applicable subset of the data and after controlling
for previously analyzed variables, except for necessary exclusions. It is
occasionally necessary to anticipate effects of variables not yet
discussed because of an interaction between variables from different
levels. These interactions will be explicitly stated. Notable variables
not significant in subset models are also reported when they relate to
results from the literature.
Extra-linguistic variables
Speech rate
Speech rate was found to be a strong predictor of word-internal
/t,d/ deletion in the complete dataset, with more deletion
occurring in faster speech. The effect of rate reflected the fact that the
mean three-word speech rate for deleted tokens was higher (6.87
syllables/sec) than the mean rate for nondeleted tokens (5.93
syllables/sec). In fact, the mean rate for deleted tokens was higher
than the mean rate for any other variant, as shown in Table 3. As with the overall mean, for each speaker
the mean absolute three-word speech rate for deleted tokens was higher
than the rate for tokens of other variant types.
Mean absolute and relative three-word speech rate for tokens of each
variant (sec)
There was a consistent effect of speech rate, regardless of the rate
measure used; however, more local contexts produced larger contributions
to predicting deletion, with the largest contribution coming from the
three-word rate measure. The three-word measure was used in all subsequent
analyses, and results using other measures will not be reported.
Relative speech rate also predicted deletion. As shown in Table 3, the mean speech rate relative to the
speaker's mean rate was higher for deleted tokens (.707) than for
nondeleted tokens (−.178). In fact, the relative speech rate was a
predictor of deletion for tokens in onsets as well as for those in codas
when these subsets were tested separately. For tokens in syllable onset
position the mean was 1.39 for deleted tokens and −.036 for
nondeleted tokens, so that for onset tokens there was a greater difference
between the relative rate means than between the absolute rate means. For
deleted tokens in codas, the mean relative speech rate is −.160 and
for nondeleted tokens it is −.718.
Dysfluencies
The results show that tokens in words preceded by a silence had a
slightly higher deletion rate (19.3%) than tokens in words not preceded by
silence (17.0%). However, there was no effect of a preceding silence after
controlling other factors. On the other hand, the absence of a following
silence predicted deletion in the dataset, reflecting the fact that words
not followed by a silence had a deletion rate almost twice as high (18.6%)
as words followed by a silence (9.7%). A following silence also inhibited
deletion for tokens in onsets (p < .001; β = −.838)
and codas (p < .001; β = −.633), when these subsets
were tested separately.
No other dysfluencies that we examined individually had an effect on
deletion, although tokens in words adjacent to dysfluent productions were
almost always associated with lower deletion rates than tokens in more
fluent environments. Lexical repetitions had lower deletion rates in both
the first and second occurrences of the word than words that were not
repeated. Words containing /t,d/ tokens that were preceded or
followed by filled pauses also had lower deletion rates compared to words
not adjacent to fillers. A following lexical cutoff inhibited deletion in
words with /t,d/ tokens relative to /t,d/ words not
followed by cutoffs as well, but when the word containing a /t,d/
was preceded by a cutoff, deletion was promoted relative to tokens in
words not preceded by cutoffs.
Age and gender
Age predicted deletion in the dataset, although the contribution to
modeling deletion was not great and the effect was not seen in all subsets
of tokens, suggesting that other factors interact with, or may account
for, the observed effect. The effect of age reflects the fact that younger
speakers had higher rates of deletion (19.0%) than older speakers (15.4%),
consistent with the findings of Guy (1992) for
word-final /t,d/ deletion. Two factors that might provide an
explanation for age differences are the absolute and relative rate
measures discussed earlier, because both measures affected deletion and
may be different in the two age groups. The effect of age was not due to
absolute speech rate differences in the two age groups, because older
talkers spoke, on average, faster than younger speakers. However, younger
talkers' mean speech rate for deleted tokens exceeded their overall
mean speech rate by only .59 syllables/sec, whereas for the older
speakers the difference was .87 syllables/sec. That is, the younger
speakers tended to delete at rates that were on average 30% lower,
relative to their mean rate, than the older speakers. The lower threshold
for deletion for younger speakers combined with their greater range of
rates largely accounts for the higher deletion rate among these
speakers.
There was an additional interaction between age and the syllable
position of the alveolar stop tokens. For tokens in syllable codas the
older speakers had higher rates of deletion (30.9%) than younger speakers
(24.3%), and older speakers produced these tokens at faster rates than
younger speakers did, consistent with the older speakers' overall
higher speech rate. However, after controlling rate, age did not emerge as
a significant factor influencing deletion for tokens in syllables codas.
On the other hand, for tokens in syllable onsets, age predicted deletion
(p < .001; β = −.608), even after controlling speech
rate. Although the older speakers had lower rates of deletion than the
younger speakers, these tokens were, once again, on average produced at
faster rates by older speakers than by younger speakers.
There was no effect of gender on deletion rate. However, like Zue and
Laferriere (1979), who reported higher internal
/t,d/ deletion for read speech in the environment
in female speakers than in male speakers, the deletion rate in this
flapping environment in our data was also higher for females (24.4%) than
for males (20.2%).
The results of analyses of extra-linguistic variables on deletion are
summarized in (2).
Lexical variables
Word class
Word class predicted deletion in the complete dataset (p <
.001, β = −.566), reflecting the fact that there was more
deletion in function words (29.2%) than in content words (12.8%). Note
that word class is highly correlated with word length, word frequency, and
predictability from neighboring words, as shown by the differing means and
standard deviations of function and content words shown in Table 4. For all lexical variables, there was a
broader distribution of values in content words than in function words.
Because of the correlation of word class with other lexical factors,
further analyses with lexical variables were performed separately on
tokens in function words and tokens in content words, as well as on the
complete dataset. Correlations between word class and other structural
factors will be considered in the analysis of phonological variables.
Mean values (and standard deviations) of some lexical variables by
word class
Word structure
There was substantially more deletion of tokens in monosyllabic words
(22.2%) than in polysyllabic words (14.7%) in the overall dataset. This
binary length variable proved significant in predicting deletion in the
dataset (p = .009, β = −1.25), as well as for tokens in
content words (p = .001, β = 1.570) and function words
(p = .001, β = 1.570) when these subsets were tested
separately. Interestingly, the binary length effect was in the opposite
direction for content and function words. As seen in Table 5, there was more deletion in monosyllabic
function words and polysyllabic content words than the reverse. This
result appears to be a consequence of structural differences between
function and content words. Function words are fairly evenly distributed
between monosyllabic (43.5%) and polysyllabic words (56.5%). However, in
polysyllabic function words, 99% of all stops occurred in a syllable
onset, whereas in monosyllabic function words, all stops occurred in a
coda. This confound largely explains the higher deletion rate in
monosyllabic as opposed to polysyllabic function words, because (to
anticipate the discussion of phonological variables) there was a much
higher deletion rate for coda tokens than for onset tokens.
Deletion rates by word length and word class
Table 5 shows deletion rate for function
words as a function of two properties, length in syllables and syllable
position of the alveolar stop. When the subset of function word onset
tokens is tested, there is a small effect of length in syllables on
deletion (p = .031, β = .312), reflecting the fact that onset
tokens in longer words are more likely to delete than those in shorter
words. The effect comes from some three- and four-syllable function
words—the indefinite pronouns (e.g., somebody and
anybody)—that have unusually high rates of deletion of
/d/ in onsets, although deletion of /t/ in onsets of
longer words was not unusually high. The high deletion rate in the
indefinite pronouns reflects idiosyncratic reduced pronunciations of these
words (e.g., somebody,
;
nobody,
;
anybody,
). After taking the identity of the class of indefinite
pronouns into account, there was no effect of syllable length on onset
tokens in function words, and hence on function words in general.
Note from Table 5 that as content word
length increased, the deletion rate for tokens in syllable codas also
increased. In contrast, tokens in syllable onsets had about the same
deletion rate in words of all lengths. In fact, length in syllables was
not a significant predictor of deletion before controlling syllable
position of the stop token. When syllable position of the word-internal
stop was taken into account, syllable length predicted deletion in content
words for tokens in codas (p < .001, β = .600), but not
for tokens in onsets. For coda tokens, deletion likelihood increased with
word length.
The effect of syllable length on tokens in onsets and in codas was
reversed when we considered the effect of ordinal syllable position on
deletion. The data showed a correlation between content word length and
the syllable in which the stop occurred. In longer words, stops tended to
occur in later syllables. In content words, the ordinal syllable in which
a stop occurred predicted deletion for onset tokens (p = .002,
β = .225), but not for coda tokens. The effect reflected the fact that
there was more deletion of tokens in syllable onsets when they occurred
later in a word.
The results of analyses involving word structure variables are
summarized in (3).
Word probabilities
Word probabilities investigated included the word frequency of words
containing word-internal /t,d/ tokens, frequencies of adjacent
words, and the predictabilities of /t,d/ token words from adjacent
words. The frequency of the word containing a word-internal alveolar stop
did not predict deletion in the overall dataset, nor did it predict
deletion in function or content words when these were tested separately.
In the complete dataset, the mean word frequency of words with deleted
tokens was higher (3.39) than words with nondeleted tokens (3.13), but
this was the result of the higher mean frequency of function words which,
as we have seen, had higher deletion rates than content words.
In both function and content words there was little difference between
the mean frequency of deleted and nondeleted tokens, as shown in Table 6. If an effect of word frequency is to be found
that is independent of word class, it is more likely to be detectable in
content words than in function words, because of the greater range of
frequencies in content words. Nonetheless, when content words were tested
separately they showed no effect of word frequency. However, as Table 6 also shows, mean frequencies of words with and
without deletions differed by syllable position (onsets and codas) and
preceding phonological context (V_ and C_ environments) of the stop token.
Interestingly, when the four subsets of content words defined by syllable
position and preceding segment class were tested separately, frequency of
the token word did predict deletion, but only for onset tokens in V_
contexts of content words (p < .001, β = .746). This
context largely comprises the V_V0 flapping environment in
words like better and lady, where V0
represents an unstressed syllable; but the context also includes words in
nonflapping environments, in words like attract. Indeed, word
frequency predicted deletion in the V_V0 flapping environment
in content words (p < .001; β = .798) when this subset was
tested separately, and these words appear to be the source of the effect
in the broader class.
Mean log frequency of deleted and nondeleted tokens by word class,
syllable position, and preceding phonological context
In analyses including frequencies of words adjacent to words
containing a /t,d/ token, subsets based on syllable position and
preceding context were tested separately, as they were in token word
frequency analyses. Analyses of following word frequency showed an effect
for content word onset tokens in V_ contexts (p = .01; β =
−.268) and in the V_V0 flapping environment (p =
.008; β = −.303). Note (from the negative beta values) that the
effects reflect less deletion of tokens in words followed by
higher frequency words than of tokens in words followed by lower frequency
words. No effects of preceding word frequency were found in any content or
function word subsets tested, or in function words.
The failure to find word frequency effects more broadly, seems at odds
with previous findings (Jurafsky et al., 2001).
However, note that in the environments in which word frequency effects
were found, deletion is associated with higher frequency token words
(w), but with lower frequency following words (w + 1).
The word frequency results can thus be readily interpreted in terms of
word predictabilities, defined in terms of the probabilities of w
and w + 1 [cf. equation (1)]. Given the effects just
reported, we thus expect an effect of
p(w|w + 1) on tokens in V_ environments
of content words. Higher predictability from the following word results
from two factors: (1) higher word bigram frequency of the pair
〈w, w + 1〉; and (2) lower frequency of
w + 1. This is precisely what the results indicate if we
additionally assume that higher w frequencies will generally be
associated with higher 〈w, w + 1〉
frequencies.
When word predictabilities were substituted for word frequencies in
the regression model, the predictability of a word containing a
word-internal /t,d/ from the word that followed it in speech
p(w|w + 1) emerged as a predictor of
deletion in the overall dataset (p < .001, β = .268),
reflecting the fact that there is more deletion of tokens in more
predictable words than in less predictable words.
The (log) mean predictability from the following word was higher for
deleted tokens (−1.95) than for nondeleted tokens (−2.31).
When word class subsets were tested separately,
p(w|w + 1) also predicted deletion for
tokens in content words (p < .001, β = .283), but not in
function words, so that the overall effect came from the effect in content
words. As expected, the effect in the overall dataset could once again be
isolated to tokens in V_ contexts in syllable onsets in content words
(p < .001; β = .554). Predictability from the following
word in speech also predicted deletion in tokens in the V_V0
flapping environment in content words (p < .001; β =
.662), reflecting the fact that deletion was more likely in more
predictable words than in less predictable words.
There were no effects of predictability from the following word on any
other subset defined in terms of preceding context and syllable position
of the word-internal /t,d/ token. There were also no effects on
deletion of the predictability of a /t,d/ word from the word that
preceded it in speech in the overall dataset in any subsets defined using
preceding context and syllable position in content words.
The results of the effects of word predictability measures are
summarized in (4).
Phonological variables
In analyses of phonological variables, two correlations between
structural variables were controlled by analyzing subsets of the data.
First, syllable position is highly correlated with the major segment class
(consonant or vowel) of the segment following the alveolar stop. The
correlation is reflected in Table 7 by the small
number of /t,d/ tokens in onsets that are followed by a consonant
(N = 294) and the lack of coda tokens followed by vowels. Because
of this confound, the effects of phonological context variables were
separately tested in subsets of onset and coda tokens, as well as in the
overall dataset. Second, the class of a word and its segmental content are
also correlated. Phonological context variables were therefore tested
separately on function and content word subsets, as well as on the overall
dataset. A summary of the effects of phonological factors in deletion is
found at the end of this section.
Deletion rates by syllable prominence on /t,d/ syllable and
the preceding syllable
Syllable position
After controlling for lexical and extra-linguistic variables, syllable
position of the alveolar stop token was a predictor of deletion for the
overall dataset (p < .001, β = −1.174). The effect
reflected the fact that there was a higher rate of deletion for
word-internal /t,d/ tokens in syllable codas (29.1%) than for
tokens in onsets (11.1%). For both tokens in onsets and codas there was
more deletion in function words than in content words. However, there was
a stronger effect of word class on tokens in onsets than on tokens in
codas. For tokens in onsets, the odds of deletion in content words were
about half the odds of deletion in function words (eβ =
.52); in coda tokens, the odds in content words were about two-thirds the
odds in function words (eβ = .67). The stronger effect of
word class on onset tokens was a result of the high deletion rates of
/d/ in indefinite pronouns (cf. the discussion of word length
effects). Taking the identity of indefinite pronouns into account
eliminates the interaction between syllable position and word class.
Prominence
Prominence level predicted deletion in the overall dataset (p
< .001, β = .662) and for tokens in onsets (p < .001,
β = .931). There was only a marginal effect in coda tokens (p
= .06). Note that the prominence variable analyzed all monosyllabic
function words as nonprominent. Including this variable in the model, the
deletion rate for tokens of /t,d/ in nonprominent syllables was
21.2%, and the deletion rate in prominent syllables was 9.5%. Prominence
on the syllable containing the stop token also predicted deletion in
content words (p < .001, β = .807), but not in function
words, when these subsets were tested separately. In all subsets, an
effect of prominence reflects the fact that tokens in nonprominent
syllables were more likely to delete than tokens in prominent syllables.
The failure to find an effect in function words can be explained by the
fact that most function words are monosyllabic, and so unaccented by
definition, leaving few function words to be distinguished by prominence.
When effects were found in function word subsets, they were consistent
with the results for corresponding content word subsets.
When the four subsets of content words defined by syllable position
and preceding segment class were tested separately, all subsets showed
effects of prominence. Deletion was more likely in nonprominent syllables
for onset tokens in postvocalic position (p < .001; β =
.845), for onset tokens in post-consonantal position (p = .002;
β = .637), and for coda tokens occurring after a vowel (p =
.004; β = 1.37). However, deletion was less likely for tokens in codas
of nonprominent syllables in C_ contexts (p = .002; β =
−1.26). The effect also held when the stop occurred intervocalically
in content words (p = .005; β = .876), but not in the
extended flapping environment V{n r l}_V0.
When the syllable preceding the alveolar stop was prominent (e.g., in
supposedly, water), the deletion rate was 17.0%, and
14.0% when nonprominent (e.g., in attempt, validate).
Prominence level of the syllable preceding the /t,d/ token
predicted deletion in the overall dataset, but not for tokens in onsets or
codas when they were tested separately. For onset tokens, there was an
interaction between prominence on the /t,d/ syllable and
prominence on the preceding syllable, reflecting the fact that there were
lower deletion rates when the preceding syllable is nonprominent and the
/t,d/ syllable is prominent, as shown in Table
7. The interaction was not significant for coda tokens, although
the pattern was the same. Thus, as has been proposed for flapping, only
the prominence level of the syllable containing the /t,d/ token
affects deletion. Finally, the prominence of the syllable following the
alveolar stop had no effect on deletion for tokens in codas, in onsets, or
overall.
Alveolar stop identity
After controlling for extra-linguistic and lexical variables, there
was a marginal effect of alveolar stop identity on deletion (p =
.052, β = −.171), with deletion more likely for /d/
tokens than for /t/ tokens. The weakness of the effect is
compatible with the fact that the overall deletion rate was about the same
in /t/ tokens and /d/ tokens, as seen in Table 8. Alveolar stop identity strongly interacted
with word class, and to understand the interaction, function and content
word subsets were tested separately.
Deletion rates for /t/ and /d/ by word class and
syllable position
As shown in Table 8, tokens of /t/
and /d/ are distributed in about the same proportion in function
words (/d/ = 25.3%) and content words (/d/ = 21.8%),
although there was more /d/ deletion in function words and
slightly more /t/ deletion in content words. The content word
effect disappeared, however, when this subset was tested separately. On
the other hand, the effect of stop identity for tokens in function words
remained when function words were tested separately (p = .003,
β = −.629), reflecting the higher deletion rate of /d/
than /t/ in these words. The effect in function words can be
further limited to words with tokens in onsets (p < .001,
β = −1.89), which had an unusually high deletion rate of 33.3%.
This subset of words consisted of only 19 types (430 tokens), which can,
except for three tokens, be further grouped into three subclasses:
indefinite pronouns ending with -body (e.g., somebody,
nobody), contracted forms with 've (e.g.,
could've, would've), and contracted forms with
n't (e.g., wouldn't, shouldn't).
The deletion rate in the subclass of indefinite pronouns was 30.2% and the
deletion rate for the contractions was 42.0%.
As discussed in the section on word length effects, high rates of
deletion in the indefinite pronouns accounted for the effect of word
length in functions words, and it was suggested that the high deletion
rates were a consequence of idiosyncratic pronunciations of these words.
We would like to test whether effects of these subclasses, found there and
again in this analysis, were a result of the phonological properties of
the words or their lexical identities. If we compare deletion rates of
these subclasses to phonologically similar words, the deletion rate of the
function word subclasses is much higher than that of comparable words. For
example, the deletion rate of /d/ in the
VLax_iy0 environment (e.g., idiot,
comedy, which are comparable to the -body words) was
only 8.7%. [The mean frequency of VLax_iy0
context words (3.15) was comparable to the mean frequency of
-body words (3.14).] The deletion rate of /d/ in
the
environment, which is comparable to the contractions, was 8.6%. These
deletion rates are very similar to the deletion rate of /t/ in the
V_V0 environment in function words (6.6%). The greater
likelihood of deletion of /d/ than /t/ in onset tokens in
function words is thus most likely attributable to idiosyncratic
pronunciations in these word subclasses and is not the result of
phonological context or of a difference in deletion between /t/
and /d/.
Recall Rhodes' (1992) claim that the
greater ease of flapping of /d/ over /t/ leads to greater
rates of deletion of /d/ over /t/. To test this
hypothesis, an analysis was performed on tokens in flapping environments.
No effect of stop identity on deletion for tokens in the intervocalic
flapping environment (i.e., V._V0), was found. However, there
was more deletion in VdV0 tokens (21.3%) than in
VtV0 tokens (12.3%), as Rhodes observed for flapping. There was
also no effect of stop identity on deletion for tokens in the extended
flapping environment V{l r n}_V0 when this set was tested
separately. In the extended environment, there was more deletion of
/t/ (23.9%) than of /d/ (19.3%).
Differences in the patterning of /t/ and /d/ were also
proposed in Zue and Laferriere (1979), who
suggested that /t/ is more likely to delete than /d/ in a
nasal release context V′_N (where N represents a syllabic
/n/). Although there was a higher deletion rate for /d/
(34.3%) than for /t/ (14.3%), generalizations are problematic,
because the number of types in the dataset with this context was small
(n = 17; although there were 294 tokens). The high deletion rate
for /d/ can once again be attributed to deletion in some specific
words, in this case, the -n't contractions (e.g.,
wouldn't). If we exclude these function words, then the
deletion rate for /d/ drops to 4%.
Consonant/Vowel: Preceding context
Proportions of deletions for tokens preceded by consonants and vowels
in the overall dataset, and in the onset and coda subsets, are shown in
Table 9. Interestingly, in the complete dataset,
a preceding vowel promoted deletion relative to a preceding consonant
(p < .001, β = −.763). Note that this overall
pattern reflects the pattern in onsets, because 72% of tokens in the
dataset (and consequently 52% of all deletions) were in onsets. In
postvocalic position, deletion was promoted for onset tokens in all words
(p = .003, β = −.346), as well as for onset tokens in
content words (p < .001; β = −.832). A preceding
consonant was shown to strongly promote deletion when the stop was in a
coda for all words (p < .001, β = 1.97), but when it was
in an onset, deletion only proved significant in content words (p
< .001; β = 2.20).
Deletion rates by segmental context
Consonant/Vowel: Following context
Unlike results from studies of word-final /t,d/ deletion, the
class of the following segment as consonant or vowel did not predict
deletion in the overall dataset after controlling for syllable position.
We assume that this is because, word-internally there is a strong
correlation between syllable position and following segment class; recall
that in our dataset, codas cannot be followed by vowels, by definition.
Note that the class of the following segment also did not predict deletion
for tokens in onsets when this subset was tested separately.
Preceding vowel contexts
The analysis of /t,d/ deletion in tokens following a vowel
compared contributions from three variables: (1) the binary classification
of vowels as lax and nonlax; (2) the classification of vowels following
Zue and Laferriere (1979), based on the presence
or absence of a high off-glide; and (3) the phone bigram frequency of the
alveolar stop and the preceding vowel.
For onset tokens occurring after a vowel, in the complete dataset
there was no effect on deletion of the two vowel classes tested or on
preceding phone bigram frequency. Interestingly, when postvocalic onset
tokens were tested separately, an effect of preceding phone bigram
frequency emerged (p = .004, β = .615), but the two vowel
classifications of the preceding vowel remained nonsignificant. The effect
of preceding phone bigram frequency reflected the fact that the mean
bigram frequency of deleted onset tokens was higher (4.64) than for
nondeleted onset tokens (4.57) in postvocalic environments. Preceding
phone bigram frequency was also a predictor of deletion for prevocalic
onset tokens in content words (p = .005, β = 1.813) and in
function words (p = .005, β = 1.813). For coda tokens,
neither preceding phone bigram frequency nor either of the verb
classifications predicted deletion in the overall dataset or in function
or content words.
Consistent with Zue and Laferriere's (1979) findings, we found less deletion in the
intervocalic flapping environment for tokens preceded by a vowel with a
high off-glide (10.0%; e.g., liter, lady), than for
those with no high off-glide (16.9%; e.g., matter,
ladder). However, there was no general effect of this vowel
classification on deletion of tokens in the flapping context. In this
environment there was also no effect on deletion of the distinction
between lax and nonlax vowels or of preceding phone bigram frequency.
Following vowel contexts
Contrary to the findings for prevocalic position, vowel bigram
frequency was not a significant predictor of deletion in onsets for tokens
in prevocalic environments. The mean following vowel bigram frequency
(i.e., the mean bigram frequency of the /t,d/ and the following
vowel) was about the same for deleted tokens (4.86) and nondeleted tokens
(4.81). Deletion rates for tokens in onsets followed by a lax vowel were
lower (10.9%) than deletion rates for nonlax vowels (13.4%), but there was
no contribution to predicting deletion from this vowel class for tokens in
onsets followed by a vowel. There was also no effect for prevocalic onset
tokens of the classification of vowels by off-glide. Following vowel
bigram frequency and classes of following vowels are not applicable to
coda tokens, which cannot be followed by vowels.
Preceding consonant contexts
For /t,d/ onset tokens preceded by a consonant, there was a
robust pattern of deletion when the preceding consonant was a homorganic
sonorant (/l/, /n/, or /r/), as shown in Table 10. Examples of deletion in this environment
from the corpus include children, ended,
wanted, and started. As indicated in Table 10, chi-square tests revealed no significant
difference between the rates for the /n/_ and /r/_
environments, nor between those for /r/_ and /l/_,
although /n/_ differs from /l/_ (p = .03).
Environments with these homorganic sonorants produce more deletion than
the /s/_ environment (p = .009) and all other preceding
segments. The difference between word-internal /t,d/ segments
preceded by homorganic sonorants and other consonants was a predictor of
deletion for /t,d/ tokens in onsets (p < .001; β =
−1.39). After controlling for preceding consonant context using this
binary variable, there was no additional contribution to predicting
deletion in onset tokens from preceding consonant bigram frequency.
Deletion by preceding consonant for tokens in onsets (with vowel
context included for comparison)
Rates of deletion of onset stops preceded by a homorganic sonorant
(16.4%) were similar to flapping rates in these contexts (19.1%). However,
although deletion was highest in /n/_ environments, flapping was
highest in /l/_ and /r/_ environments. It is interesting
to note that the higher rate of deletion after a homorganic sonorant was
similar to that after a vowel (11.8%), also a flapping context. In fact,
the difference between /l/_ and V_ contexts was nonsignificant by
chi-square test, but rates in /n/_ differed from rates in V_
(p = .03).
The findings differ from previous studies of final /t,d/
deletion in which obstruents (stops and fricatives) were found to promote
deletion, although sonorant consonants (nasals and liquids) inhibited it
(Guy 1980, 1997; Labov et al., 1968; Neu,
1980; Wolfram, 1969). Our finding that a
homorganic sonorant promotes deletion is consistent with the findings of
Fasold (1972), Neu (1980), and Zue and Laferriere (1979), who each found that a preceding homorganic
consonant promoted deletion both for final and nonfinal /t,d/ (in
Zue & Laferriere, only /n/ was tested). However, our results
revealed that homorganicity alone is not an effective predictor, because
deletion rates for tokens in onsets after a homorganic obstruent were low.
Rather, homorganicity is relevant when the consonant is also a sonorant.
The results are summarized in (5). The differences between the findings of
this and earlier studies relate to the fact that syllable position of the
alveolar stop was not previously taken into account; as we show next, the
conditioning factors differ for tokens in codas.
As shown in Table 11, deletion of
/t,d/ in coda position was most likely when preceded by a
nonhomorganic obstruent, /f, p, k/. Although the number of tokens
in the environments /f/_ and /p/_ were low, the results
were consistent with the patterning of /k/_. Deletion was next
highest in the contexts /s/_ or /n/_, differing
significantly from the highest group (/k/_ vs. /s/_;
p = .04). Deletion was least frequent when the preceding
consonant was a liquid. For comparison, Table 11
includes the deletion rate of coda /t,d/ after a vowel, which had
a deletion rate between the environments with preceding liquids
(p < .01) and /s, n/_ (p < .001).
Deletion by preceding consonant for tokens in codas (with vowel
context included for comparison)
When we restricted our analysis to content words, vowel contexts did
not differ significantly from liquid contexts. Preceding vowels and
liquids patterned as a class, which we refer to as the class of
approximants (Clements, 1990;
Ladefoged,1982). (Because the glides /w,
y/ did not occur before an alveolar stop in coda position in the data,
/l, r/ are the only consonantal approximants in this context in
the dataset.) Thus, tokens in codas had lower deletion rates when preceded
by an approximant than when preceded by a nonapproximant. The class of
nonapproximants may be divided further, with deletion more likely when the
consonant is nonhomorganic. The results are summarized in (6).
The distinction between approximants and nonapproximants predicted
deletion in the dataset (p < .001; β = 2.50), reflecting
the fact that tokens preceded by nonapproximants had higher rates of
deletion than those preceded by approximants. After controlling for this
distinction, homorganicity of the preceding consonant also predicted
deletion (p = .017; β = 1.03), reflecting the fact that
deletion of a word-internal alveolar stop was more likely when preceded by
a nonhomorganic consonant than when preceded by a homorganic consonant.
After controlling for preceding consonant context using these two
variables, there was no additional contribution to predicting deletion in
coda tokens from preceding consonant bigram frequency.
Recall that previous studies found that the manner of articulation of
a preceding consonant influenced final /t,d/ deletion. Obstruents
(stops, fricatives) favored deletion but sonorant consonants, such as
nasals and liquids, inhibited it. Our findings for coda /t,d/ are
similar in that obstruents generally promoted deletion but liquids
inhibited it. However, the classification of segments into the familiar
manner categories (fricative, stop, nasal, liquid) is less revealing than
the broader classification of sounds in terms of the categories
approximant and homorganic.
Following consonant contexts
There were differences in the rate of /t,d/ deletion depending
on the identity of the following consonant. Chi-square tests revealed a
significant difference in the deletion rates between word-internal stops
followed by /b, m, f, l/ and followed by /s/, as indicated
in Table 12. However, there was no difference
between the contexts _/b, m, f/, and _/l/. Similarly, no
difference was revealed among _/s, n, v/. Deletion rates also
differed significantly between _/s/ and _/z/. Deletion
rates in the overall dataset thus suggested no classification of these
following consonants based on phone features that might be used to predict
deletion. That is, the environments _/b, m, f, l/ do not form a
natural class, nor do _/s, n, v/. The deletion rate difference
between _C onset tokens (or /t,d/ followed by a glide) and _C coda
tokens (followed by a nonglide) were marginally significant by chi-square
test. However, the distinction for following consonant of
glide–nonglide did not predict deletion in the overall dataset or
for tokens in onsets when this subset was tested separately.
Deletion rates by following consonant in onsets and
codas
Our results are at odds with the results of previous studies of final
/t,d/ deletion that found deletion to be negatively correlated
with the sonority of the following segment, with less sonorous segments
promoting deletion. In the present study, some of the highest rates of
deletion occurred before a sonorant consonant (e.g., _m, 66.7%; _l,
50.5%). On the other hand, deletion before some obstruents was
significantly lower (e.g., _s, 31.5%; _v, 23.1%; _z, 14.3%). Our results
of internal /t,d/ deletion indicating that sonority is irrelevant
in predicting deletion is in agreement with Guy (1991) and Labov (1995), who
noted that the sonority of the following consonant in word-final
/t,d/ deletion is insufficient to account for observed deletion
patterns. Their alternative proposals were tested in the dataset.
Guy (1991) proposed that deletion of
word-final /t,d/ is more likely when the stop is unable to
syllabify with a following consonant. Our results for _/r/ and
_/l/ are consistent with Guy's proposal. If the explanation
is generalizable, then we would expect that word-internal alveolar stop
deletion rates would pattern according to the syllable position of the
preceding word-internal alveolar stop. Consonants that cannot be
syllabified with a preceding /t,d/ include /b, f, m, l/,
which are all in the group with the highest deletion rates in Table 12. However, /n, v/ also cannot be
syllabified with a preceding /t,d/, and these two phones had lower
deletion rates than the /b, f, m, l/ group. The phones /s, z,
r, w/ will almost always (except for a very small number of cases with
/s, z/, such as kids'll and whatsoever)
syllabify with a preceding /t,d/. Because the groups /b, f, m,
l, n, v/ and /s, z, r, w/ do not differ in their deletion
rates, there is no relation between deletion rate and the ability of a
consonant to syllabify with a preceding /t,d/.
In an alternative phonotactic account, Labov speculated that deletion
rates may be tied to the frequency with which consonant cluster patterns
are learned. Under the assumption that phone bigram frequencies in the
lexicon reflect exposure of phone clusters to language learners during
acquisition, then phone bigram frequencies of /t,d/ and the
following consonant might be expected to predict deletion rates, with more
deletion in higher frequency bigrams. Following consonant bigram frequency
did predict deletion for /t,d/ for tokens in codas (p
< .001; β = −.781) but not for tokens in onsets. The effect
was limited to tokens in C_ contexts of codas (p = .029; β =
−.600). The effect reflects the fact that alveolar stops in less
frequent stop–consonant bigrams were more likely to delete than in
more frequent bigrams; the mean following bigram frequency in C_ codas of
deletions (5.02) was lower than for nondeleted /t,d/ (5.04). The
direction of the effect is unexpected from Labov's proposal. Although
we found no direct evidence that /t/ and /d/ behave
differently in the contexts examined, the effect of following phone bigram
frequency could provide evidence for a difference based on alveolar stop
identity, because bigram frequencies are computed relative to the identity
of the canonical /t/ or /d/ segment.
Finally, the difference in deletion rates between /s/ and
/z/ shown in Table 12 might suggest that
alveolar stop deletion is promoted by a following voiceless fricative.
Note that generalizing the role of voicing to any following consonant
would be problematic, because the highest rates of deletion are observed
before the voiced segments /b, m, l/. Alternatively, because
adjacent obstruents frequently share the same voicing property, the
observation that more stops deleted before /s/ might be taken as
synonymous with saying that more /t/'s deleted than
/d/'s. Further analysis of the data revealed that neither
explanation provides a straightforward account of the patterns. As shown
in Table 13, although the proportion of deletion
is greater for /d/ before voiceless consonants than voiced
consonants, the observation that /t/ did not occur before voiced
consonants makes it impossible to choose between the two accounts.
Deletion rates for /t/ and /d/ tokens by following
sibilant context
Taking into account whether a consonant or vowel preceded the sequence
is more elucidating. A /t/ or /d/ followed by a coronal
fricative was more likely to delete when preceded by a consonant,
regardless of which stop was at issue, as shown in Table 14. Conversely, when the sequence was preceded
by a vowel, deletion of /d/ was rare, but it was more likely with
/t/. Although this latter observation might suggest a fundamental
difference between /t/ and /d/, on closer scrutiny, the
cases in which /t/ deleted in this context were overwhelmingly
function words, as shown in Table 14.
Deletion for /t/ and /d/ in sibilant contexts by
preceding CV context
From these analyses, we may conclude that the higher deletion rate
before /s/ is a consequence of both the part of speech and the
preceding context. Specifically, when the stop-obstruent sequence is
preceded by a vowel, deletion is higher when the /t,d/ is in a
function word. This higher deletion rate in function words was
particularly striking for /t/; the paucity of tokens with
/d/ in this context made it impossible to determine whether
/d/ deleted at greater rates in function words than in content
words. Furthermore, when a coda /t,d/ was both followed by an
obstruent and preceded by a consonant, deletion was highest. This finding
is consistent with the results of the previous section, in that coda
/t,d/ was more likely to delete when flanked by consonants.
Summary of phonological effects
Our results indicate a significant difference in the patterning of
deletion in onset and coda tokens. When in onset position, an alveolar
stop is more likely to delete when preceded by a vowel than by a
consonant. Conversely, when the stop is in coda position, deletion is more
likely when the stop is preceded by a consonant than by a vowel. After
controlling for syllable position of the word-internal alveolar stop,
whether the following segment was a consonant or a vowel was not
significant, because of the fact that, by definition, stops in onsets are
always followed by vowels (or glides in a small number of cases), and
stops in codas are always followed by consonants. Thus, the overall
finding that deletion was higher for coda stops than for onset stops is
consistent with claims that deletion is more likely when a word-internal
alveolar stop is followed by a consonant than by a vowel.
The effects of prominence were also different for tokens in onsets and
codas. Tokens in the onsets of nonprominent syllables exhibited more
deletion than tokens in prominent syllables. Tokens in the codas of
nonprominent syllables also showed more deletion than those in prominent
syllables, if they were in a V_ environment; however, if they were in a C_
environment, prominence enhanced deletion over nonprominence. These
effects were only significant for tokens in content words, but tokens in
functions words followed the same pattern, although the results were not
significant, because of the small numbers of function words in many
subsets.
When we examined the effects of individual vowels on deletion, we
found that higher frequency vowel–stop bigrams promoted deletion of
/t,d/ tokens in onsets, but not in codas. Stop–vowel bigram
frequencies did not affect deletion in onsets or codas. Furthermore, no
effects of the vowel classifications tested were found for vowels that
preceded or followed word-internal alveolar stops. In particular, vowels
with high off-glides had no effect on deletion of a following
word-internal alveolar stop. However, there was a tendency for less
deletion in flapping environments in which the stop is preceded by a high
off-glide vowel, supporting the notion that high off-glides strengthen
flaps, and thus make deletion in these words less likely.
When we examined individual consonants, deletion in syllable onsets
was promoted when the stop was preceded by a coronal sonorant. This
observation is expected if deletion and flapping are gradient outcomes of
the same process, because this environment is largely the extended
flapping environment V{nrl}_V0. For coda stops in C_ contexts,
on the other hand, rates of deletion were higher when the word-internal
stop was preceded by a nonapproximant than an approximant, and higher when
it was preceded by a noncoronal than a coronal. After controlling for the
quality of the preceding consonant context using these features, there was
no contribution to predicting deletion from consonant–stop bigram
frequencies in onset tokens or in coda tokens, confirming that it was the
quality of a consonant, and not its likelihood in speech, that influenced
deletion of a following alveolar stop. Moreover, deletion was more likely
in contexts where consonant transitions involved changing the place of
articulation, which encouraged cluster simplification.
There were no robust effects indicating that /t/ and
/d/ exhibit different patterns of deletion, either in the overall
dataset or in specific environments about which claims for differences
based on stop identity have been made in the past. Higher rates of
deletion of /d/ than /t/ in function words could be
attributed to the behavior of subclasses of /d/-internal words
(the indefinite pronouns, -n't contractions, and
-'ve contractions) that exhibited higher than expected
deletion rates given their phonological structure. Rhodes' claim that
/d/ deletes more than /t/ in flapping environments was
supported by higher rates of deletion of /d/ than of /t/
in the intervocalic flapping environment, although the difference was not
significant, but in the extended flapping environment there was more
deletion of /t/ than /d/. There was some support for Zue
and Laferriere's (1979) claim that
/t/ deletes more than /d/ in the nasal release
environment, but only if we exclude the subclass of /d/ words with
-n't contractions. With these words omitted, /t/
tends to delete more than /d/ in the nasal release environment,
which may indicate a difference between /t/ and /d/ in
nonlexicalized variation.
No features defining natural classes of consonants that followed
word-internal alveolar stops predicted deletion of the preceding stop,
although different consonants were associated with different rates of
/t,d/ deletion. Deletion could also not be predicted based on the
ability of an alveolar stop to resyllabify with a following consonant.
However, less frequent alveolar stop–consonant bigrams promoted
deletion for tokens in C_ contexts of syllable codas, accounting for some
of the variation in deletion rates across individual following consonants
in this one context. There was also evidence in some contexts that an
effect of the following consonant was the result of an interaction of the
following consonant's identity with the identity of, or class of, the
segment preceding the alveolar stop. In short, no clear conclusions
regarding the effects of different following consonants on word-internal
alveolar stop deletion can be drawn from these results.
GENERAL DISCUSSION: MODELING DELETION
The present study indicates that numerous factors influence the
deletion of alveolar stops in word-internal positions, some phonological,
some lexical, and some extra-linguistic. We argue in this section that the
results support the view that deletion in all environments is a reduction
process. However, evidence will be presented that also suggests that there
are two different reduction processes leading to deletion. One applies to
/t,d/ segments in syllable onsets, and the other to /t,d/
in syllable codas. We will further argue that the two processes can be
understood in terms of a model of speech production (e.g., Levelt, 1989) by proposing that the two processes
arise at different stages of production. If the origins of the two
processes during production are different, then this predicts that the
processes will be differently influenced by some of the factors
investigated; we provide evidence that this was the case.
Deletion was promoted by faster speech in all environments. This
result extends to internal alveolar stops the findings for word-final
alveolar stops in Jurafsky et al. (2001), that
is, a faster utterance rate increased the likelihood of deletion. The
effect of speech rate on deletion argues that the deletion process is
similar to other reduction processes, which have been shown to increase
with speech rate. The fact that higher speech rates resulted in deletion
of word-internal segments also indicates that rate can effect changes in
segmental production across the word and not just at word edges.
Fluency is known to affect reduction processes, with reduction less
likely in dysfluent speech. The tendency for less fluent speech to inhibit
deletion was seen in the behavior of some of the categorical measures of
fluency that we tested. Deletion in all contexts was inhibited in words
followed by a pause in speech, consistent with general findings that words
followed by a pause tend to have fuller, less reduced final segments, as
well as specific findings of lower deletion rates in words followed by a
pause of word-final alveolar stops (Jurafsky et al.,
1998; Wolfram, 1969). Although no
significant results were found for other categorical fluency factors that
were examined, such as adjacent fillers and repetitions, words in
dysfluent environments defined by these categories had lower proportions
of deletions than words in fluent contexts. We suspect that a larger data
sample would show a broader effect of dysfluency on the inhibition of
deletion within the word.
Although speech rate and the fluency factors that we examined
similarly influenced tokens in both onsets and codas, this was not the
case for other factors, which affected alveolar stops in onsets and codas
differently. One explanation for the differences is that reduction through
deletion results from different processes in the two structural
environments. In fact, deletion in the two environments has previously
been described in two different ways. Deletion in codas (especially in
consonant contexts) has been attributed to cluster simplification, and
deletion in onsets (especially in flapping contexts) has been
characterized as resulting from segment lenition. Our results for onsets
and codas support this interpretation, because of important differing
results for alveolar stops in onsets and in codas.
There were substantially more deletions in codas than in onsets.
Recall that a /t,d/ in coda position was necessarily followed by a
consonant. Deletion patterns for coda tokens were thus consistent with
many previous studies of final /t,d/ deletion (cf. Guy, 1980), which have found that following consonants
promote deletion. For coda tokens, deletion was higher when the stop was
preceded by a consonant than when preceded by a vowel. A cluster of three
consonants is the context in which an intermediate stop has been observed
to be most susceptible to deletion (cf. e.g., Wright,
1996) and supports the conclusion that deletion in coda position
can be viewed as a consequence of consonant cluster simplification. It is
thus not surprising that coda stops preceded by approximant consonants
showed rates of deletion similar to stops preceded by vowels, given the
vowel-like nature of approximants. Among the nonapproximants, cluster
simplification resulting in deletion was more likely when the consonant
differed in place of articulation from the alveolar stop, increasing the
articulatory difficulty of the cluster.
Contrary to coda stops, word-internal onset stops were more likely to
delete when preceded by a vowel than a consonant, although deletion rates
were not as disparate in the two contexts as they were in codas. This
result is not anticipated by previous studies for two reasons. First,
onset stops are not typically assumed to delete; and second, previous
studies have not generally found post-vocalic deletion word-internally
(Zue & Laferriere, 1979) or word-finally
(Guy, 1992). The reason for deletion in this
context, we would suggest, is that 65.3% of postvocalic onsets were in a
flapping environment. Post-vocalic deletion of stops in onsets is thus
consistent with the proposal of Rhodes (1992)
that deletion in these environments is simply a more extreme version of
lenition than flapping. A relation between flapping and deletion is also
supported by the fact that there tended to be more deletion of /d/
than /t/ in post-vocalic flapping contexts, consistent with
Rhodes' observation that /d/ flaps in a wider range of
phonological environments than /t/. In addition, higher preceding
vowel bigram frequencies promoted deletion, indicating that deletion in
this context is a gradient process that is sensitive to usage factors. A
further indication of the relation between deletion and flapping can be
seen in post-consonantal environments, in which deletion was more likely
to occur following consonants that allow for flapping (V{l r
n}_V0) than when a medial alveolar stop followed some other
(nonhomorganic, nonsonorant) consonant. Moreover, when the
post-consonantal onset deletion rate was highest (Vn_V0
environment), flapping rates were lower, compared to environments where
flapping rates were higher (V{lr}_V0 environments), in which
deletion was less common. Deletion of alveolar stops in onset position
thus appears to result predominantly from gradient lenition that includes
flapping, and so was promoted in environments that also encouraged
flapping.
Some differences between our results and results from earlier studies
of word-final /t,d/ deletion may be attributed to the fact that
syllable structure was distinguished in the current work. Alveolar stops
analyzed in studies of final /t,d/ deletion, particularly those of
connected speech, may have included stops that occurred in coda position
as well as in onset position. This is because of the fact that in
connected speech, a word-final coda stop will generally resyllabify as the
onset of the following syllable if it creates a licit onset with the
following segments. We may estimate an upper bound on the rate of
resyllabification across word boundaries by determining how often
word-final alveolar stops are followed by a vowel or glide in spontaneous
speech, cases in which resyllabification typically occurs. Using the
transcribed subset of the Buckeye corpus (approximately 100,000
words), proportions of word-final alveolar stops in onsets and codas were
estimated. Of the words not followed by a silence, 43.9% of final
/t,d/ tokens were followed by a word whose initial segment would
allow for resyllabification of the final stop and may thus have been
produced in an onset. Therefore, it is reasonable to assume that the data
in previous studies of word-final /t,d/ deletion were likely
dominated by alveolar stops in codas, though were not exclusively coda
tokens. By comparison, in the current dataset, 80.6% of all word-internal
/t,d/ tokens in the transcribed subset (a superset of the dataset
used for the analysis) were in onsets, because phonotactics prohibits many
consonant sequences that can occur at word boundaries. Deletion of
alveolar stops in onsets with vowel contexts would thus have been masked
in previous studies by the preponderance of alveolar stops in codas.
Moreover, given the differences in the effects of context variables on
onset and coda tokens found in the current study, it is not surprising
that there are discrepancies across reported results from word-final
alveolar stop studies. Previous results crucially depended on the mix of
onset and coda productions in the data sample analyzed.
The lower rate of deletion in faster and less fluent speech indicates
that deletion is the result of reduction. Moreover, there is evidence for
two deletion processes, lenition in syllable onsets and cluster
simplification in syllable codas. Can the two processes be distinguished
in a model of speech production? We can examine this question by
considering how reductions may occur in a production model.
Reductions can arise at different levels of speech production. First,
the deletion of a word-internal alveolar stop may be part of a variant
pronunciation that is stored in the lexicon. Previous studies have
proposed that this is the case for the words and (Neu, 1980) and just (Labov, 1975). We have found evidence that some
frequently occurring subclasses of functions words—the indefinite
pronouns and -n't and -'ve
contractions—exhibit deletion rates that are unexpectedly high, when
compared with words with similar phonological structures. That these
frequently used words have lexicalized pronunciation variants is
consistent with models of lexicalization in which change occurs in more
frequent environments (cf. Bybee, 2000).
Current production models propose that after lexical access, word
segments are linearized within metrical frames in the process of segment
planning (e.g., Levelt, 1989). During segment
planning, regular but nonobligatory pronunciation variants are introduced.
Such variants include assimilations (e.g., nasal place assimilation) and
also segment deletions. Segment planning reductions are largely computed
based on phonological context, but whether they are used will depend on
production parameters (like speech rate) and on speech context (i.e.,
register, style, or fluency). Planning variants are thus more likely to be
selected to meet the requirements of faster, more fluent speech or a more
casual speech style. They will be dispreferred in slower or more careful
speech, or when there are planning problems, as signalled by dysfluencies
in production. When planning is complete, the segment plan is then
incrementally used to generate a gestural score that can drive
articulation. During articulation, further reduction may occur in fluent
speech, in part in response to extreme parameter settings for rate and
prominence, through gestural overlap and gestural reduction. These
articulatory adjustments result in gradient productions, and gestures may,
in the extreme, be omitted completely (or substantially enough), so that
an associated segment target is not perceived; that is, the segment is
deleted.
The differing behavior of medial alveolar stops in syllable codas and
in syllable onsets suggest that they can be interpreted in terms of these
different methods of introducing reduction during production. We propose
that coda deletions, which result from cluster simplification, can be
explained largely as the result of segment plans that omit alveolar stop
targets in consonant environments for greater gestural economy. Gestural
economy is achieved at the expense of alveolar stops that are not final in
a consonant sequence, through planned assimilations with neighboring
consonants, or complete target omission. Onset deletions, on the other
hand, are the outcome of a gradient process of gestural lenition of
planned targets during articulation that can result in shortened stops,
flaps, target approximation, or even nearly complete gestural
omission.
Once again, the goal of economy of gestural sequencing is to meet rate
and style requirements in fluent speech. However, note that when gestural
omission of coda stops in planning is not exercised, the resulting stop
segment is subsequently subject to lenition during articulation. This
lenition may lead in some productions to gestural omission, just as it
does in some onset stops, so that coda stops may also be deleted by
lenition, and thus may show a low degree of sensitivity to lenition
factors. We have seen that this is the case for rate and fluency
factors.
Having two sources of reduction, articulatory lenition and planning
simplification, means that there will be further differences between the
two reduction types that are reflected in the behavior of alveolar stops
in onset and coda tokens. Evidence for the separate origins of deletions
through gestural lenition or gestural economy in the proposed model comes
from the interaction between the onset and coda medial alveolar stop
positions with speaker age, word and phone probabilities, and syllabic
prominence.
Deletion was more likely in the speech of younger talkers than older
talkers, consistent with the findings of Guy (1992). The effect of age could not be attributed to
speech rate, because in our data older speakers spoke, on average, faster
than younger speakers. The difference between the two age groups resulted
from the tendency for younger speakers to delete at lower rates (relative
to their mean rate) than older speakers do. This tendency among the
younger speakers is perhaps best explained by a difference in articulatory
practice between the two age groups that allows older, more practiced
speakers to produce gestures at faster (absolute or relative) rates than
younger speakers. If the difference is indeed articulatory, then we would
expect different behaviors in the two age groups, depending on whether we
look at word-internal alveolar stops in syllable onsets or syllable codas,
and an interaction was found. For words with word-internal alveolar stops
in syllable codas, the older speakers' faster rate of speech resulted
in more deletions, but there was no difference between the two age groups
after controlling for speech rate, which increases all reductions. For
tokens in syllable onsets, older speakers showed fewer deletions even at
their higher speech rates. The older speakers' greater practice in
articulating apparently allows them to produce stops in onsets without
deletion at higher speech rates, but articulatory practice had no effect
on deletion in syllable codas. To the arguments based on age, we can add
that, despite some previous evidence for gender differences in deletion
rates, no overall differences were found in deletion rates of males and
females. This result would be expected if the source of deletion is
articulatory reduction and not differences in the planning choices that
may be made by different speaker groups.
A further difference in the patterning of deletion in syllable onsets
and codas emerged and was seen in effects of word probabilities. Word
predictability only affected deletion of alveolar stops in syllable
onsets. Deletion for word-internal alveolar stops in V_ onsets in content
words was higher when a word containing the stop was more predictable from
the following word. The effect is consistent with results from studies of
final alveolar stops that have shown that words that are more predictable
from the following word are more reduced, that is, they show more
word-final alveolar stop deletion (Jurafsky et al.,
2001) and have shorter durations (Raymond et
al., 2001) than less predictable words. In our study, more
predictable words were also produced at a significantly faster rate (6.19
syll/sec) than less predictable words (6.02 syll/sec) (F
= 10.125; p = .001). However, there was an additional effect of
predictability on deletion of alveolar stops in onsets, even after rate
was controlled for. The effect is perhaps because words that are more
expected in context will be produced more fluently, resulting in more
reduction in words containing stops in syllable onsets during articulation
through gestural lenition. This word predictability measure had no effect
on deletion of stops in codas, presumably because predictability does not
influence the earlier stage of phonological planning. However, the
proportion of coda deletions was higher in more predictable words than in
less predictable words, indicating that fluency factors may minimally
affect coda deletion as well, as predicted for fluency factors.
Word-internal alveolar stops in onsets and codas were also affected
differently by prominence. In onsets, alveolar stops were more likely to
delete if the syllable was nonprominent. Nonprominent syllables are
produced with weaker consonant articulations, and this lenition promotes
deletion of onset stops. In codas, on the other hand, stops in C_C
consonant contexts were more likely to delete if the syllable was
prominent. This result is unexpected if coda deletion is also the result
of gestural adjustments during articulation, because prominence generally
strengthens segments. However, it may be explained if the strengthening of
clusters can be enhanced by the gestural omission of cluster-internal
(alveolar) consonants during planning. If there is no cluster-internal
stop, as in V_C environments, then strengthening due to prominence is
possible, resulting in lower deletion rates, like for tokens in onsets,
which is what was observed.
Finally, phone bigram frequency measures affected onsets and codas
differently. In onsets, deletion was more likely in postvocalic contexts
when the vowel–stop bigram frequency was high. Not only were high
bigram frequency pairs more likely to delete than low frequency pairs, but
they also deleted at a lower mean rate (6.62 syll/sec) than low bigram
frequency pairs (7.20 syll/sec), indicating a fluency effect. As with
more predictable words, words with higher frequency bigrams will be
produced more fluently, encouraging more articulatory lenition. In C_C
context codas, on the other hand, there was an effect of
stop–consonant bigram frequency, but unlike the bigram effect in
onsets, deletion was more likely when the frequency was low than when it
was high. An increase in deletion for low frequency sequences can be
explained as perhaps a planning effect, in which less commonly used
articulatory plans within words are more likely to be simplified during
segment planning than more common plans, leading to increased deletion in
low bigram frequency environments.
In conclusion, we have found that deletions of word-internal alveolar
stops is the result of two reduction processes. In syllable onsets,
deletions can be understood as a consequence of gestural lenition in less
practiced, more fluent, or less prominent productions. Deletions of
word-internal alveolar stops in syllable codas, on the other hand, are the
result of a different process, because there were only weak effects of
lenition factors on this environment. Coda deletions were, rather,
predominantly sensitive to segmental context, and so are better
characterized as the result of meeting speech requirements through
simplification of gestural sequences during speech planning, with only
weak effects of articulatory fluency.
Word-internal alveolar stop deletion was widespread in the spontaneous
speech that was analyzed. It occurred in all contexts examined in the
speech of our corpus. The word-internal data that we employed in the study
and a methodological approach that acknowledges the probabilistic nature
of variation in spontaneous speech encouraged understanding of variation
as a complex phenomenon that can be explained within the broader context
of speech production.
