2.1. Phonologically-conditioned variation

The realization of coronal stops in American English, which we focus on in this paper, can be quite accurately predicted from syllabic position and identity of adjacent segments (Kahn, 1976; Randolph, 1989). The distribution of allophones can be described as the outcome of an input-output mapping under particular phonological conditions. For example, flapping has sometimes been described by the rule in (1)¹:

This rule may oversimplify things in that there is some evidence that there are degrees of flapping, but it accurately captures a restriction on flaps in running speech—they almost never occur outside of this phonological environment.² Randolph (1989) found in a corpus analysis that out of 953 flaps, only 16 (1.7%) occurred outside of this environment (Randolph, 1989, p. 119–20). In a study of spontaneous speech, Patterson and Connine (2001) found a 93.9% flapping rate for word-internal, intervocalic /t/ following a stressed vowel. Zue and Laferriere (1979) found that 99% of t/d were flapped in the same environment during experimental trials involving read speech.

However, this rule also over-predicts the presence of flapping. Word boundaries after the potential flap significantly impact the likelihood of flapping. Randolph (1989) found that of 1389 alveolar stops preceded by a vowel or glide and followed by a vowel, only 937 (67.5%) were flapped. Under experimental conditions, Fukaya and Byrd (2005) found flapping rates in word-final, phrase-internal contexts as low as 30% for one of their participants (n = 60), though another participant flapped on all 60 trials. In the same experiment, phrase boundaries almost categorically blocked flapping (only 2 out of 179 trials flapped). So although the rule in (1) seems to capture an important generalization about flaps, it doesn’t fully explain their distribution in spontaneous speech.

Voiceless coronal stops have an additional possible realization as a glottal stop. This variant is well-documented in many varieties of English, including American Englishes (Byrd, 1994; Eddington & Channer, 2010; Roberts, 2006). This allophone is restricted to coda position (though glottal stops may independently appear elsewhere), and is sensitive to the identity of the following segment (Huffman, 2005; Seyfarth & Garellek, 2015; Sumner & Samuel, 2005). Glottalization occurs mostly before consonants, and is much more common before sonorants than obstruents. However, it is not excluded from any context: Eddington and Channer (2010) found a 24.8% glottalization rate for /t/ followed by vowels in the Santa Barbara Corpus.

Glottalization is also sensitive to phrasal position: Both Huffman (2005) and Seyfarth and Garellek (2015) report reliably higher rates of glottalization in phrase-final position (i.e., when followed by an intonational phrase boundary). Additionally, Seyfarth and Garellek (2015) report that the effect of the following segment is significantly reduced in phrase final position, and that the rate of glottalization in the absence of any following segment is just over 50%.

Both flap and glottal stop allophones are sensitive to phonological context, but not to the same extent. Flapping is totally blocked in the absence of the right segmental context, or if the context is disrupted by large prosodic breaks, while glottalization can occur in any context but seems to be sensitive to adjacent segments under certain conditions.

How can these two ways of being sensitive to phonological context be reconciled? An early response was to attribute variability to non-phonological ‘performance’ factors, and proceed with categorical descriptions of phonological patterns, in the vein of (1) (Chomsky & Halle, 1968). Another analytical option pursed early in the sociolinguistic tradition (Labov, 1994), and more recently gaining traction in constraint-based frameworks (Coetzee & Pater, 2011), is to incorporate the probabilistic aspect of phonological patterns into the phonological description itself.

Our approach combines two key insights of these approaches, that (1) variability may be introduced into phonological patterns by ‘performance’ factors external to the phonological grammar itself, and (2) modeling variability is crucial to understanding phonological patterns. We propose specifically that systematic variability is introduced during the online processing of speech. Although we do not rule out the possibility that phonological knowledge itself may have a probabilistic component, we suggest that understanding different sources of variability is crucial to the development of parsimonious phonological models. We turn now to a discussion of the speech production system, and how it is influenced by word predictability.

2.2. Speech production

Planning to speak involves several stages. Speaking aloud requires the speaker to first formulate a message at the conceptual level, which then provides the starting point for linguistic processing, and ultimately becomes an articulatory plan ready to be externalized. Current models of spoken word production identify at least two distinct stages of linguistic processing: lexical selection and form encoding (using the terminology of Levelt, 2001; see Dell & O’Seaghdha, 1992; Levelt, Roelofs, & Meyer, 1999) for detailed articulations of two influential models, and Wheeldon & Konopka, 2018 for a recent review).

Lexical selection is the process whereby appropriate linguistic representations are selected to express the concepts in the speakers’ intended message. In the case of single-word production, the result of this process is the selection of a unique lemma, which specifies the syntactic and semantic properties of the word to be spoken. This stage temporally precedes access to any phonological information, which only occurs later, during form encoding.

Form encoding begins with the retrieval of the phonological code associated with the selected lemma. Then, these phonemes are anchored to a metrical structure including at least syllabic and prosodic word levels. This representation guides the assembly of a more detailed phonetic code which can be passed forward for articulatory execution.

In running speech, these processes of selection and encoding must occur multiple times, along with additional processes to integrate these lexical items into the larger syntactic and prosodic context. The relative timing of these processes is not yet well understood. It is broadly agreed that speech is planned incrementally from the beginning of the utterance (Bürki, 2018; Levelt et al., 1999; Meyer, 1991; Shattuck-Hufnagel, 1979). The speaker can initiate articulation as soon as the motor plan for the first word, perhaps even the first syllable, is complete (Kawamoto, Liu, & Kello, 2015). Hence, linguistic planning and articulation occur in parallel, with planning racing just ahead of what is coming out of the speaker’s mouth.

Some global details of the utterance must be computed before articulation begins, like prosodic phrasing and intonation contours (Keating & Shattuck-Hufnagel, 2002). However, these details can be fixed before form encoding is complete, since they don’t rely on information about the phonemic make-up of words, which could be filled in later. In fact, F. Ferreira (1993) showed that the final slot in a prosodic phrase is assigned a fixed duration by speakers regardless of the length of the word that will be in that slot. As early as 1978, Sternberg, Monsell, Knoll, and Wright proposed that multi-word utterances are encoded as programs in which a number of sub-programs are embedded. This allows utterance-level variables to be set early, while phonetic details of the sub-programs are retrieved only as necessary as the utterance is unfolding. Speech could begin as soon as the first sub-program is ready to be articulated.

There is also variability with respect to the planning scope. Wheeldon and Lahiri (1997, 2002) provide evidence that utterance initiation time can depend on the number of prosodic words, and speakers take longer to initiate speaking for longer sentences when they had time to silently prepare the target sentence. However, when speakers were not given time to prepare in advance, their speaking latencies instead reflected the number of syllables in the initial prosodic word of the sentence. We argue that this variability in planning scope can explain some of the variability observed in sandhi processes.

However, it is common to see segmental interactions across prosodic words, suggesting that it is possible for the forms of multiple words to be encoded in tandem. Anticipatory speech errors are clear evidence for this (Fromkin, 1971): For example, errors like “the mirst of May” instead of “the first of May,” show that the phonological code of /m/ from “May” can be active at least in time to affect the encoding of the intended “first.” The same logic can be applied to context-sensitive allophones such as flaps. Since a flap can only be planned in intervocalic contexts, flapping across the boundary of two prosodic words (e.g., “We upse[ɾ] Andy”) shows that these words must have been encoded within the same window.

Though it must be possible for the form encoding process to span more than one word, it’s not clear how far in advance this planning window extends, or under what conditions it does extend beyond a prosodic word (Bürki, 2018). In fact, it seems that the size of the window may vary, and can depend on several different factors. Wheeldon and Lahiri (1997, 2002) found that depending on the task, utterance initiation time can be driven more by the number of upcoming prosodic words (with delays between presentation of the stimulus and the cue to start speaking), or the internal complexity of the first upcoming prosodic word (in speeded tasks). In other words, how far a speaker plans ahead is task-dependent. The size of the planning window has also been found to depend on syntactic constituency and semantic coherence (Wheeldon, 2013), and on the lexical frequency of the words involved (Konopka, 2012). An increase in cognitive load has been shown to reduce speech rate (Mitchell, Hoit, & Watson, 1996), and been argued to decrease planning scope (F. Ferreira & Swets, 2002; V. Wagner, Jescheniak, & Schriefers, 2010). Individual differences in working memory also correlate with planning scope (Swets, Jacovina, & Gerrig, 2014). Michel Lange and Laganaro (2014) found evidence that speakers who initiate speech more quickly show less sensitivity to phonological details of upcoming words (Experiment 2). Finally, we note that the ‘planning window’ metaphor carries an implicit assumption of discrete units of planning, when in fact it may be more appropriate to think of planning a continuous process that involves gradient levels of activation (Pluymaekers, Ernestus, & Baayen, 2005a). This view of planning is also compatible with the logic of our study: Instead of asking about ‘extent of planning window,’ one would ask instead about what upcoming material has been activated to the degree that it could affect planning of the current word.

One factor which is known to have significant effects on linguistic processing is lexical frequency. Landmark studies from Oldfield and Wingfield (1964, 1965) showed that the time it takes to initiate speech in the task of picture-naming depends on the lexical frequency of the picture’s name. Subsequent work has strongly supported the finding that lexical frequency influences the time it takes to name a given word (Griffin & Bock, 1998; Jescheniak & Levelt, 1994; Schilling, Rayner, & Chumbley, 1998). In multi-word utterances, Konopka (2012) found that sentences beginning with high-frequency words were initiated faster than those beginning with low-frequency words.

Levelt and colleagues’ influential model of spoken word production locates lexical frequency effects at the level of form encoding, when the phonological code of a lemma is being retrieved (Levelt et al., 1999; though see Gahl, 2008 for evidence that frequency effects also arise at prior stages). So, all else being equal, higher frequency words will be retrieved and encoded sooner than lower frequency words. Extrapolating to the planning of multi-word sequences, we hypothesize that the lexical frequency of each word after the first may affect whether or not they are all phonologically encoded within the same window by either speeding or slowing retrieval. This is illustrated in Figure 1. Each box represents the duration of processing for a given word at a given stage. The green leftwards arrow and box represent conditions which are favorable for rapid initiation of form encoding for word 2, allowing it to begin before word 1 form encoding has finished, and allowing for potential interaction between their phonological forms. On the other hand, the red rightwards arrow represents conditions under which form encoding is delayed, and word 2 is prevented from exerting any influence on the form encoding of word 1.

Figure 1

Schema of the time course of select stages of planning a two-word sequence. The vertical dimension represents the distinct stages of linguistic planning, lexical selection and form encoding. The boxes represent the duration of processing at each stage for each word. If conditions facilitate retrieval of the phonological code word 2, form encoding can begin sooner, and may overlap with form encoding of word 1 (green scenario), while a delay could prevent interaction between word 1 and word 2 (red scenario).

The processing of word 1 may also itself be facilitated or delayed by frequency effects, but how this might affect cross-word interactions is less clear. Miozzo and Caramazza (2003) found that in single word utterances, high-frequency distractors affect production latency less than low-frequency distractors. They conclude that frequent words are planned earlier relative to subsequent words, and therefore interfere less with following words. This would suggest that a high-frequency first word might make it less likely that it is planned together with the following word. Given the diagram in Figure 1, this would make sense if frequency effects apply at the level of the form encoding (as argued in Levelt et al., 1999), but leave the relative timing of lexical selection intact. In this case, two words should be less likely to be encoded at the same time when the first word is frequent compared to when it is less frequent, and word 1 frequency should have a negative effect on flapping rate. Konopka (2012), on the other hand, found that a high-frequency first word leads to greater semantic interference with the following word, suggesting a high frequency of the first word makes it more likely for two words to be planned together, at least with respect to their semantics. This would make sense if frequency effects apply at the lexical selection stage and thus retrieval of the second word happens earlier relative to the phonological encoding of the first word, as argued in Alario, Costa, and Caramazza (2002). We would then expect that phonological planning of the second word would be more likely to happen while the first word is being planned, and flapping rate should increase with the frequency of the first word. Kittredge, Dell, Verkuilen, and Schwartz (2008) tried to arbitrate between these views of the level at which frequency effects apply. They found frequency effects at both stages, though age of acquisition effects were only found at the phonological level (see also the discussion Tanner et al., 2017). To conclude, it’s less clear what to expect with respect to the effect of the frequency of the first word, but a higher frequency of the second word should make it more likely for two words to be planned together.

While lexical frequency reflects the general likelihood of encountering a word in any context, it’s also clear that language users are sensitive to the contextual predictability of words. Griffin and Bock (1998) showed that speakers are much quicker to name objects when they had just heard a semantically congruent sentence. Beattie and Butterworth (1979) found that in spontaneous speech, words that are less predictable from context are more likely to be preceded by a hesitation. This effect remained even when the lexical frequency of the words was controlled, but only for low-frequency words. Konopka (2012) found that the scope of planning, indexed by a phonological priming effect, was expanded when the first word in a sentence was one that subjects had recently produced (in an earlier, unrelated experimental task).

Measures of predictability have also been shown to have an effect on words’ phonetic realization. Gregory, Raymond, Bell, Fosler-Lussier, and Jurafsky (1999), analyzing a subset of monosyllabic t/d final words from a corpus of telephone conversations, found that the highest frequency words were 22% shorter than the lowest frequency words. They also found that word duration was correlated with semantic relatedness and discourse repetition, with word duration decreasing for words that had been previously mentioned and which were semantically related to the words in the preceding conversation. Jurafsky, Bell, Gregory, and Raymond (2001) found that several measures of predictability are correlated with rates of final t/d deletion. Gregory et al. (1999) found similar results, and also that flapping of t/d is more likely between pairs of words with high mutual information. Gahl and Garnsey (2004) found that speakers were more likely to delete a verb-final t/d when the verb was in a syntactic frame that matched its usual syntactic complement. Torreira and Ernestus (2009) found an effect of bigram frequency with the following word on the acoustic realization of /t/ in French. Pluymaekers, Ernestus, and Baayen (2005b) showed that for seven high-frequency words in Dutch, mutual information with the following word was predictive of reduction, with fewer segments realized when mutual information was high. Raymond, Brown, and Healy (2016) found that the predictability of the following phonological environment (i.e., whether a given word is typically followed by a consonant or vowel) had a significant effect on rates of t/d deletion. In the flapping environment specifically, they found that the likelihood of deletion was positively correlated with the likelihood that the target would be followed by a vowel. These empirical results clearly establish that predictability affects the phonetic realization of specific segments within words.

This evidence suggests that predictability, in both the sense of prior probability and contextual probability, have an important effect on the time course of linguistic processing in general and form encoding in particular. We propose that the variability of the planning window, and its interaction with form encoding, is crucial to understanding how predictability affects allophonic variation.

2.3. The Production Planning Hypothesis

How does predictability modulate the selection of context sensitive allophones? We draw on a recent line of investigation which relates intra-speaker variability to the dynamics of speech planning (Bürki, 2018; Kilbourn-Ceron, 2017b; MacKenzie, 2012, 2016; Tamminga, 2018; Tamminga et al., 2016; Tanner et al., 2017; M. Wagner, 2011, 2012). Our proposal, in brief, is that predictability affects the size of the form encoding window. The form encoding window, in turn, restricts the size of the input to the phonological input-output mapping. Information that falls outside this window cannot affect allophone selection—even if that information is found in the very next word. This is the case illustrated in the lower portion of Figure 1 in red. If the trigger of a phonological process is not planned soon enough, the process cannot apply, a situation that Tamminga (2018) aptly names a co-presence failure. In the case of t/d, a co-presence failure with a following vowel would remove the opportunity for the flapping rule to apply, therefore the rate of flapping should be modulated by the likelihood that the conditioning environment (i.e., the following vowel) has been planned early enough.

For example, consider a two-word sequence like cat attack. Of the several possible phonological encodings of the word cat, the flapping rule in (1) predicts a flap when the following word is attack. However, this assumes that the segmental information of attack has been (at least partially) retrieved and is available at the time that the encoding of cat is taking place. If cat must be encoded in the absence of information about the following word, then the flapping rule could not come into play at all, and some other variant should be selected. We propose that both of these scenarios are possible, as illustrated in the lower half of Figure 1, and that this is the source of some of the variability of context-sensitive cross-word interactions. Thus the Production Planning Hypothesis predicts that factors that affect speech planning also affect phonological interactions between words.

Furthermore, the Production Planning Hypothesis makes the prediction that any phonological alternation which relies on phonological information from an upcoming word must be variable, since phonological processes cannot apply if the conditioning phonological environment in the next word has not yet been retrieved, and we know that speakers do not reliably retrieve the phonological detail of more than one word ahead of time.³

Tanner et al. (2017) present evidence that the rate of coronal stop deletion in British English is affected by the following phonological context, and that production planning modulates this relationship. They found that longer pauses between words and higher word frequency reduced the effect of the following context on deletion. They hypothesized that longer pauses between words and higher word frequency both reduce the chances that a word is planned at the same time as the following word. They did not however find strong evidence that the effect of following context was modulated by the conditional probability of the two words, so the relationship between predictability and phonological context effects is still not entirely clear. The present study seeks to find clearer evidence by examining a process that is more closely dependent on the segmental context, flapping, and comparing to another which is much less dependent, glottalization.

2.4. Alternative accounts of predictability effects

Most previous work on the relationship between predictability and pronunciation variation has focused on phonetic reduction in particular. Several types of proposals have emerged from this research, many of which are mutually compatible with each other and with the proposal of this paper. However, few address the question of how the distribution of phonologically-sensitive pronunciation variants is affected by predictability. We briefly review some existing proposals and outline their predictions in the context of our study.

Words and segments that are more predictable have been consistently found to be phonetically reduced, especially when considering duration (see end of Section 2.2). Production ease accounts consider phonetic reduction a reflex of easier planning conditions, though accounts differ on whether the planning difficulty of previous, current, or future material is the focal point (see for example V. S. Ferreira & Dell, 2000; Pluymaekers et al., 2005a; Watson, Buxó-Lugo, & Simmons, 2015). Under the assumption that lexical frequency eases planning, higher lexical frequency should be associated with a realization more dissimilar from the citation form. Communicative accounts propose that phonetic reduction is driven by the speaker’s desire to efficiently and accurately transmit their intended message (Aylett & Turk, 2004; Hall, Hume, Jaeger, & Wedel, 2018; Jaeger, 2010; Turk, 2010, among others). Use of a pronunciation variant like a flap, which neutralizes the phonemic t/d distinction, or a glottal stop, which removes place of articulation cues, presumably decreases intelligibility. On this assumption, a communicative account would predict that these variants should be used when message predictability is relatively high, as would be the case for higher frequency words.

Finally, representational accounts attribute predictability effects to accumulation of stored experiences of specific words and phrases (Bybee, 2001, 2007; Pierrehumbert, 2001). Frequent repetition leads to lenition, so under this account, both the frequency of the variable word and its frequency of co-occurrence with the following word would have a positive correlation with use of the context-sensitive variant.

What is clear from this work is that there is a strong negative correlation between predictability and duration. This presents the possibility that many of the apparently qualitative changes associated with high predictability, including the change from coronal stop to a flap or glottal stop, could in reality be a gradient process that arises from temporal compression of gestures, rather than mappings like the rule in (1). There is evidence that flapping is (or at least can be) a gradient process that involves degrees of flapping, (e.g., Fox & Terbeek, 1977), even if the acoustic consequences of flapping often appear to be rather categorical (De Jong, 1998). A gradient account is also made plausible by the fact that consonants other than t/d are subject to similar temporal reductions in flapping environments (Browman & Goldstein, 1992; Turk, 1992), and by findings that flapping does not neutralize the distinction between an underlying /t/ and /d/, which remains detectable in small but consistent phonetic differences in the length of the preceding vowel (Braver, 2011; Herd, Jongman, & Sereno, 2010; Malécot & Lloyd, 1968). This pattern is unexpected if flapping involves a categorical phonological change (though see Bermudez-Otero, 2011).

Our proposal is compatible with many aspects of this account. In particular, we emphasize that although our discussion of flapping and glottalization implies a categorical alternation, our discussion and conclusions are compatible with a gradient analysis of these variants. The transcription in the Buckeye corpus on which this study is based is just a coarse proxy measure based on perception. Though the articulatory reality is undoubtedly more complex (Fukaya & Byrd, 2005; Purse, 2019), perceptual annotation is a reasonable starting point for investigation given the finding of De Jong (1998) that even gradient articulatory overlap can lead to categorical perceptual results.

A major point of difference is the assumption that phonological representations are invariant and that qualitative variation is attributable solely to temporal compression. The PPH relies on the assumption that contextual changes in phonetic form are encoded during the speech planning process. Studies on coarticulation have found evidence that anticipatory coarticulation is planned, rather than being an automatic articulatory process. Whalen (1990) investigated anticipatory coarticulation between a word-initial /a/ and a consonant or vowel in the next syllable. The F1 of /a/ was lower if a /b/ followed compared to a /p/, and F2 was lower if an /u/ followed, compared to an /i/. However, when participants were asked to initiate speech with part of the word missing (e.g., A_I or AB_), coarticulatory effects disappeared for the missing segment, even though the missing segment was immediately revealed and integrated into the utterance. Liu, Kawamoto, Payne, and Dorsey (2018) used a similar paradigm, and tested a greater number of participants. They found that several participants showed the same pattern of anticipatory coarticulation as the three participants in Whalen (1990), though others showed no differences between conditions. This provides evidence that articulatory plans are actively adjusted during the planning process as a function of what information is available about upcoming segments.

We can make a similar point based on the data used in this study by looking for evidence that speakers make choices about the articulatory plan, rather than simply compressing the existing articulatory plan differently depending on temporal factors. We will look at two qualitatively different outcomes of reduction, glottalization and flapping, which are not degrees of the same gradient reduction process, but involve different planning choices by the speaker.

2.5. Summary and predictions

The relationship between predictability and phonetic reduction is clear: Many measures of predictability are positively correlated with reduction. However, there has not yet been much research on how predictability affects allophones, which may also be considered reductions, but differ in important ways. In particular, some allophones like flaps require information about the phonological context to be available during their planning.

Our first research aim is to establish a clearer empirical picture of how predictability and allophone distribution relate to each other. Some previous work has found that higher predictability is associated with a higher probability of phonological interactions like flapping (Gregory et al., 1999) and voicing assimilation (Ernestus et al., 2006).

Given this previous work, we expect to find a significant positive correlation between flapping and our measures of predictability. As for glottal stops, we are not aware of any previous studies that have reported rate of glottalization in relation to measures of predictability. If glottalization of /t/ is considered a general reductive process, being a reduction of the tongue tip gesture, it would be expected that lexical frequency of the target word should be positively correlated with glottalization.

Secondly, we put our account of predictability effects to the test. The mechanism proposed by the Production Planning Hypothesis makes different predictions about how different allophonic processes will be affected by predictability based on whether they are sensitive to phonological context. Increased predictability facilitates planning, potentially widening the advance planning window, and therefore increases the rate at which a context-sensitive process like flapping applies in North American English.

In contrast, glottalization of /t/ in North American English does not strictly require a particular phonological context to be realized. It is not excluded from any context: Eddington and Channer (2010) found a 24.8% glottalization rate for /t/ followed by vowels in the Santa Barbara Corpus, and Seyfarth and Garellek (2015) found rates between around 25% and 90% for different types of following consonants in the Buckeye corpus. Before a pause, i.e., in the absence of a following segment, the likelihood of a glottal stop was just over 50%. Since the present study is restricted to intervocalic context, we expect that contexts which promote the inclusion of a following vowel in the same planning window should slightly decrease the likelihood of glottalization, since the ‘default’ rate of glottalization when no segment follows is relatively high (i.e., around 50%), while pre-vocalic glottalization is relatively low.

3.1. Dataset

The set of observations used for our analysis were pairs of words in which the first ended in /t/ or /d/ immediately preceded by a vowel (hereafter target words), and followed by a vowel-initial next word (hereafter trigger words), e.g., “ended up,” “quite easy.” The were collected from the Buckeye corpus of conversational speech (Pitt et al., 2007), a corpus of sociolinguistic interviews with 40 speakers native to central Ohio, totaling about 300000 words. The speakers were balanced by age (over/under 40), gender of speaker, and gender of interviewer (Kiesling, Dilley, & Raymond, 2006).

The corpus contained 11863 qualifying word pairs, which were extracted along with existing time-aligned phonetic transcription using the Montreal Corpus Tools software (McAuliffe, Stengel-Eskin, Socolof, & Sonderegger, 2017).⁴ The Pitt et al. (2007) transcriptions were prepared automatically and subsequently hand-corrected by phonetically trained research assistants. For flaps [dx], annotators were instructed to only include segments with sustained voicing throughout the phone. For glottal stops [tq], transcribers were instructed to “label all /t/ or /d/ phones which show glottalization the phoneme label /tq/” (Kiesling et al., 2006). A test of labeling consistency using four transcribers and four one-minute samples from the corpus yielded an inter-transcriber agreement of 92.9% for stop consonants (Pitt, Johnson, Hume, Kiesling, & Raymond, 2005, 80.3% overall). For our analysis, we grouped the observations into four categories based on the realization of the underlying coronal stop in the target word: “full” if the surface transcription matched the underlying (21.46% of tokens), “flapped” if transcribed with [dx] (54.83%), “glottalized” if transcribed with [tq] (14.24%), and “other” for any other transcribed segment.

We removed observations in which the trigger word was a disfluency marker like “um”⁵ (20% of tokens), and those in which the trigger word was reduced to a syllabic sonorant on the surface (0.09% of tokens). This left 8428 tokens for analysis.

We enriched the dataset with information about the probabilities of the observed words. The prior probability for each word was estimated by retrieving its lexical frequency from SUBTLEX-US, a database of word frequencies based on a 51 million-word corpus of film and television subtitles (Brysbaert & New, 2009). Frequencies ranged from 39971 per million to 0.2 per million for words that occurred only once in the corpus. The range of values for target and trigger words was comparable, with median values of 15.44 and 16.41 words per million respectively. The empirical correlations between each of these measures and flapping and glottalization are shown in Figure 2. Distributions are plotted in Appendix A.

Figure 2

Relationship between SUBTLEX-US word frequency (per million words) and proportion of tokens transcribed as flaps [dx] (left panel, blue), glottal stops [tq] (right panel, red) in the Buckeye corpus. Solid lines show trigger word frequency, dashed lines show target word frequency, with shading showing 95% confidence intervals of a linear smooth (GLM, logit-link). Bubbles (trigger word frequency) and triangles (target word frequency) show density of observations, with size reflecting the number of observations at a given value on the x-axis.

We also calculated a measure of probability which takes into account the likelihood of each word pair as a collocation. There are many ways pairwise likelihood could be calculated. The bigram frequency is simply the likelihood of two words occurring together. This value is highly dependent on the frequencies of the individual words in the bigram, since the bigram cannot be more frequent than either of the words individually. It also cannot distinguish between pairs where the first word is infrequent and the second is highly frequent, or vice versa.

We chose to focus on the conditional bigram probability of the trigger word given the target word (hereafter conditional probability), which controls for the base frequency of the target word. For example, “out of” and “instead of” are sequences with similar relative bigram frequency, occurring equally often, but very different conditional probabilities: “instead” is highly likely to be followed by “of” (about 90% likely), while “out” is only followed by “of” about 20% of the time.

In order to estimate the conditional probabilities for the words in our dataset, we fitted a bigram language model to the SWITCHBOARD corpus (Godfrey, Holliman, & McDaniel, 1992), a corpus of spontaneous telephone conversations comprising about 3 million words. Using this larger corpus as the basis for the language model allows for more accurate estimates, especially for two-word sequences. The language model was fitted using the lmplz function from the KenLM language model toolkit (Heafield, 2011), which uses modified Kneser-Ney smoothing without pruning. The conditional probabilities calculated by the language model were matched orthographically to the two-word sequences from the Buckeye dataset. The empirical relationships of conditional probability to flapping and to glottalization are shown in Figure 3, and the empirical distribution is presented in Appendix A.

Figure 3

Relationship between Conditional Probability (of trigger word given target word) and proportion of tokens transcribed as flaps [dx] (blue, left panel) or glottal stops [tq] (red, right panel) in the Buckeye corpus. Solid lines and shading are linear smooths (GLM, logit-link) with 95% confidence intervals. Bubbles show density of observations, with size reflecting the number of observations at a given value on the x-axis.

Several variables were also included in the dataset to act as controls in the statistical analysis. The underlying voicing of the target words’ final segment was recorded; /t/ was flapped in 56.1% of cases, while /d/ was flapped slightly less at 52%. Number of syllables is highly correlated with word frequency, as the most frequent words are monosyllabic. Target and trigger words were labeled as either monosyllabic or polysyllabic, with each syllabic segment in the Buckeye surface transcription counting as one syllable. Most of the observations consisted of two monosyllabic words (77.8%), followed by monosyllabic-polysyllabic pairs (13.5%), polysyllabic-monosyllabic pairs (7.6%), and 86 polysyllabic pairs (1%). Flapping rates were comparable within these groups (53%–57%), and glottalization rates showed some variation (7%–17%).

We also included duration as a control measure for several reasons. The first relates to an articulatory account of flapping that views it as an automatic result of durational compression, as it is entertained for example in the literature on articulatory phonology (Browman & Goldstein, 1992; Byrd & Saltzman, 2003).

The PPH is very compatible with a gradient account of flapping, and with proposals about gradient gestural overlap. What distinguishes the PPH from prior articulatory overlap accounts is that it does not view flapping as an automatic consequence of temporal compression. By including duration as a control measure, we want to ensure that temporal compression alone is not sufficient to explain the observed patterns of variability.

A second reason to include a durational control is that the phonological literature on flapping has related the process to prosodic phrasing, and holds that the effect of other factors such as syntax is mediated by the presence or absence of certain prosodic junctures (Nespor & Vogel, 1986). The Production Planning Hypothesis is very compatible with the idea that prosodic phrasing will modulate flapping rate, but it also predicts that it should not be the only factor affecting the likelihood of flapping. A durational measure can serve as a proxy measure for prosodic boundary strength (Wightman, Shattuck-Hufnagel, Ostendorf, & Price, 1992), and including it in the model will help with the argument that the observed variability is not purely a result of variability in the prosodic phrasing of utterances.

The durational measure we chose for this study was the ratio of observed/expected duration of the target word. The expected duration was calculated by adding together the mean durations of each phone in the surface transcription. The mean phone durations were calculated over the entire Buckeye Corpus. A value below 1 indicates that the target word is shorter than expected based on the average durations of its component phones, and a value greater than 1 means it is longer than expected. In addition to any predictability-induced durational effects, this variable also captures compression due to faster speech rate, or expansion due to boundary-induced final lengthening (Wightman et al., 1992). Its distribution is illustrated by the plots in Appendix A.

Another control measure we included is the presence or absence of pauses, clearly a factor in coronal stop realization. The rate of glottalization before a pause is much higher than when no pause follows, 47.4% versus 8.4% (n = 1143 with pause, 7285 without pause), and flapping before a pause was rare (1.3%). Therefore we included a variable tracking whether or not a pause was annotated in the Buckeye transcription between the target and trigger words.

3.2. Analysis

The occurrences of flapping and glottalization were analyzed in separate logistic regression models with elastic net regularization. This technique penalizes large coefficient estimates, which allows (1) the shrinkage and/or removal of the least predictive variables, and (2) mitigation of collinearity-induced estimate inflation. This is of particular concern since the probability-based predictors are correlated by definition. Figure 4 shows that Trigger Word Frequency and Conditional Probability have a strong positive correlation, as expected since the latter is calculated using the former. Using the penalized regression technique may lead to dropping whichever one of these variables is less predictive. However, this does not necessarily constitute evidence against an independent effect of the less predictive variable – we return to this issue in the discussion.

Figure 4

Correlations between variables to be included in model. Opacity and ellipse size indicate magnitude, and no ellipse indicates non-significant correlations.

Following the procedure outlined in Tomaschek, Hendrix, and Baayen (2018), the models were fit using the glmnet (Friedman, Hastie, & Tibshirani, 2010) package for R (R Core Team, 2013). We used the cv.glmnet function which performs k-fold cross-validation and returns possible values for lambda, the penalty imposed on non-zero coefficients. The value for alpha was 1, equivalent to the lasso model, which yields the model stringent penalty on non-zero coefficients (but may sacrifice accuracy). We selected a value of lambda within one standard error of the minimum mean squared error (MSE) of the cross-validated models, as per the recommendation of Tomaschek et al. (2018, and references cited therein). This resulted in a model with the smallest number of non-zero coefficients while maintaining a reasonable MSE. If a coefficient remains in the model, it can be taken as evidence that it is an important part of explaining the variance in the dataset. Reliable standard errors are not available for regularized regression models, so we have refrained from reporting p-values in the tables below. Results of non-penalized logistic regressions, including standard errors and p-values, are reported in Appendix B, and were qualitatively similar.

Each model predicted the log-likelihood of a particular variant (either flap or glottal stop) as a function of the variables described in the previous section. Both models included the following fixed effects: Target Word Frequency, which was standardized by subtracting the mean and dividing by two standard deviations; Duration, log-transformed to approach normality and also standardized; Pause, a categorical variable with “no pause” as the reference level (0) and “pause” set to 1; Target # of Syllables, Trigger # of Syllables (monosyllabic or polysyllabic) were binary variables, which were centered around 0 by subtracting the mean value. Additionally, the model for flapping included Underlying t/d, tracking the underlying voicing of the target word’s final segment (also a centered, binary variable). For glottalization, the model excluded all data with /d/-final words, since these segments are very rarely realized as glottal stops (11 of 1561 /d/-final tokens in the current dataset), and therefore also excluded the Underlying t/d variable.

The glmnet package does not support inclusion of random effects in the model. However, we report non-penalized regressions in Appendix B with random intercepts by-speaker and by-target word, with qualitatively similar results. Additional models which included all variables and maximum identifiable random-effects structure were also fitted, again with qualitatively similar results.

Tables 1 and 2 show the model estimates for the fixed effects coefficients in the fitted models. Each coefficient represents the estimated change in log-odds of the outcome when other predictors are held at their mean observed values, except Pause which is held at 0 (no pause).

3.3. Results

4.1. Flapping

Word-final coronal stops are more likely to be flapped when the word that follows is highly predictable, according to at least one of the variables that we investigated. Conditional Probability, the probability of the trigger word given the target word, had a significant positive effect on flapping. The frequency of the trigger word itself also appeared to have a positive effect, but the estimate was not significant in the statistical. This may have been due to issues with suppression because of a high correlation between Conditional Probability and Trigger Word Frequency (Tomaschek et al., 2018) – further work in more controlled paradigms is necessary to ascertain whether these are two independent effects (see Kilbourn-Ceron & Goldrick, 2019, for recent findings on this question).

These results are compatible with the predictions of the Production Planning Hypothesis. The predictability of the trigger word, whether global or local, affects how quickly the word is accessed and encoded during speech planning. The more predictable the trigger word is, the more likely it is to be planned within the same window as the previous (target) word. Since simultaneous availability of the following vowel is a necessary condition for flapping, increased availability of the trigger word is predicted to have a positive effect on the likelihood of flapping. The fact that the Conditional Probability effect is larger than, and possibly masks, the Trigger Word Frequency effect may make sense based on previous speech production findings. Beattie and Butterworth (1979) found that hesitations were consistently correlated with contextual probability, but lexical frequency effects were no longer significant when contextual probability was held constant. Konopka (2012) found that the extension of planning scope based on the frequency of the first word in a sentence only took place if the structure of the sentence had been primed, making it easier to plan. It may be that lexical frequency effects in running speech are too subtle to be detected or only come into play under certain planning conditions. Controlled studies are needed to discover the contribution of lexical frequency to planning of allophonic variants.

In terms of probabilistic reduction accounts, the effect of Conditional Probability supports the idea that ease of planning of upcoming material leads to reduction of words being currently planned, if flapping is considered a reduction. On the other hand, we failed to detect any effect of Target Word Frequency on flapping, similar to the results of Gregory et al. (1999), suggesting that the effects of predictability on durational reduction and segmental deletion may be qualitatively different from how predictability interacts with flapping. Under a representational account of probabilistic reduction, it might be possible to argue that the underlying driver of reduction is the frequency of the word pair, which is correlated with the Conditional Probability measure. Under a communicative account, it could be argued that a more complex relationship is at play. For example, Turnbull, Seyfarth, Hume, and Jaeger (2018) proposed that there is a trade-off of inferrability between the target and trigger words in nasal place assimilation. Their results supported this trade-off idea, with target words showing a higher degree of coarticulatory effects on F2 when target word predictability was high, but also when trigger word predictability was low. This is the opposite of what we found for flapping, which also encodes information about the upcoming word, namely that it begins with a vowel. The PPH would predict that the increased predictability of the trigger word should also facilitate nasal assimilation. These conflicting results suggest that gradient coarticulatory effects could be an interesting future testing ground for these two types of accounts. It could be interesting to compare nasal place assimilation and coronal stops realizations directly, since nasal assimilation neutralizes to another phoneme, while flapping and glottalization do not, and so may have different informational consequences.

We also found that the presence of a pause was a significant predictor of flapping. Pauses are associated with larger prosodic boundaries, which have been found in earlier work to block flapping (Patterson & Connine, 2001; Scott & Cutler, 1984). In addition to acting as proxies for prosodic boundaries, pauses may have also been indications of hesitations, disfluent speech, and/or planning difficulties. It would be interesting in future work to try to disentangle the effect of prosodic boundaries from those of planning-induced pauses and lengthening (F. Ferreira, 1993). A preliminary look at the observations in our corpus that were followed by filled pause words shows that flapping is much lower than average for these words, with 20.3% for “um” (n = 133) and 15.2% for “uh” (n = 303), even though they technically fit the segmental description to trigger flapping.

For Duration, we found no significant effect. This suggests that for a given sequence of phones, its duration compared to the mean of that same sequence does not predict whether or not the target word contains a flap. Flapping was more likely when the underlying phone was /d/ rather than /t/, which is unsurprising since /t/ has the additional possible realization of glottal stop competing with flapping. There was no detectable effect of number of syllables for either the target or trigger word.

4.2. Glottalization

Our analysis of glottal stops revealed that Conditional Probability had a negative correlation with the likelihood of glottalization. This is in the direction predicted by the Production Planning Hypothesis: An extension of the planning scope to include the following vowel initial word would make it more likely that glottalization should be suppressed, since the flap variant will be chosen instead. We note that since our analysis only includes intervocalic contexts, the canonical flapping environment, we are cautious in interpreting this result as evidence that glottalization is highly sensitive to segmental context. It could be that the lack of glottalization is mirroring the increase in flapping in intervocalic contexts. However, previous work such as Seyfarth and Garellek (2015) has shown that glottal stops are still sensitive to segmental context in pre-consonantal environments, so production planning effects should still be in force. If the analysis were to be repeated with only pre-consonantal contexts, the opposite effect would be predicted, with the highest correlation between glottalization and Conditional Probability in pre-sonorant contexts.

This type of pattern is unexpected under accounts of probabilistic reduction, assuming that glottalization is a reduction relative to the full /t/ gesture. It could be possible to argue, as an anonymous reviewer suggested, that glottalization involves reinforcement through addition of a glottal gesture. If this is so, then a production ease account could explain the negative correlation as the addition of a gesture in difficult-to-plan contexts, though it is unclear why the tongue tip gesture would simultaneously be dropped. Under a communicative account, glottal reinforcement could be considered a way to strengthen a cue to word boundaries in environments where the upcoming word may be difficult for the listener to retrieve. However, there is evidence that realization of a /t/ as a glottal stop hinders recognition of the target word. Garellek (2013) found that subjects are significantly less accurate in recognizing minimal pairs like “dent-den” when the final /t/ is realized as a glottal stop (though see Chong & Garellek, 2018 for further results on recognition of glottalized vowel-consonant sequences). Hence, there might have to be a trade off between intelligibility of the target and trigger words under this account, as suggested by Turnbull et al. (2018). Furthermore, an explanation based on glottal reinforcement does not predict an opposite effect of conditional probability before sonorants.

There were no significant effects of lexical frequency on glottalization. However, exploratory analyses based on model comparison suggest that frequency may play a role in explaning some of the variation in glottlization; further research is needed to clarify this issue.

The effects of Duration and Pause were significant and in the expected positive directions. Glottal stops are common before pauses (Seyfarth & Garellek, 2015), and glottal voice quality in general is highly associated with intonational phrase boundaries (Redi & Shattuck-Hufnagel, 2001), which are typically lengthened. This may indicate that in addition to being sensitive to upcoming segments, the planning of the glottal stop itself may be related to details of its position in larger prosodic structure, and glottalization may serve as a cue for a boundary, or increase the perceived strength of a boundary. The negative effect of frequency could then be a reflection that boundary strength correlates negatively with the predictability of a following constituent (Turk, 2010). There was no detectable effect of number of syllables for either the target or trigger word.

4.3. Theoretical interpretation

Our analysis of flapping and glottalization reveal that the distribution of these variants is not straightforwardly explained by assuming that they are predictability-motivated reductions. Target word frequency, previously found to be a reliable predictor of durational compression and segmental deletion (Aylett & Turk, 2006; Jurafsky et al., 2001), was not predictive of either allophone. The main finding of our analysis is that conditional probability of the trigger word given the target word had significant and opposite effects on flapping and glottalization. This is consistent with the predictions of the Production Planning Hypothesis, and adds to recent studies in support of speech production effects on phonological variability (Kilbourn-Ceron, 2017a; Lamontagne & Torreira, 2017; Tamminga, 2018; Tanner et al., 2017).

Obviously a much greater range of processes will have to be looked at closely in order to tease apart which mechanism(s) are responsible for the observed effects. Our main goal here was to show that the Production Planning Hypothesis makes very concrete predictions in this regard, which differ from the predictions of alternative hypotheses, and that our data support these predictions.

A relevant question that remains open is at what stage allophonic variants are selected. For example, syllable-initial aspiration of voiceless stops in English could be implemented during phonological encoding, as long as syllabification is done first. Or, it could arise during phonetic encoding if an aspirated stop motor program is selected directly from a phonological representation which is unspecified for aspiration. It may well be that both are possible mechanisms for contextual variation. Although the answer to this question does not change the logic of our study, we find this to be an interesting avenue for future research that might allow us to make even more detailed predictions about variable sound patterns.