Effects of phonotactic predictability on sensitivity to phonetic detail

Japanese speakers systematically devoice or delete high vowels [i, u] between two voiceless consonants. Japanese listeners also report perceiving the same high vowels between consonant clusters even in the absence of a vocalic segment. Although perceptual vowel epenthesis has been described primarily as a phonotactic repair strategy, where a phonetically minimal vowel is epenthesized by default, few studies have investigated how the predictability of a vowel in a given context affects the choice of epenthetic vowel. The present study uses a forced-choice labeling task to test how sensitive Japanese listeners are to coarticulatory cues of high vowels [i, u] and non-high vowel [a] in devoicing and non-devoicing contexts. Devoicing contexts were further divided into high-predictability contexts, where the phonotactic distribution strongly favors one of the high vowels, and low-predictability contexts, where both high vowels are allowed, to specifically test for the effects of predictability. Results reveal a strong tendency towards [u] epenthesis as previous studies have found, but the results also reveal a sensitivity to coarticulatory cues that override the default [u] epenthesis, particularly in low-predictability contexts. Previous studies have shown that predictability affects phonetic implementation during production, and this study provides evidence predictability has similar effects during perception.


Introduction
The current study investigates Japanese listeners and the role of phonotactic predictability in how illicit consonant clusters are repaired.While it is commonly thought that Japanese listeners use [u] epenthesis by default because it is the shortest (and thus phonetically minimal) vowel (Dupoux, Kakehi, Hirose, Pallier, & Mehler, 1999;Dupoux, Parlato, Frota, Hirose, & Peperkamp, 2011), the current study proposes that the choice of epenthetic vowel in Japanese listeners relies on a combination of phonotactic predictability and attention to phonetic cues, based on experience with recovering high vowels that are systematically devoiced or deleted (Shaw & Kawahara, 2018;Whang, 2018) in their language.To distinguish the respective roles of phonotactic prediction and phonetic cue perception, participants are presented with conflicting phonotactic and phonetic information, allowing insight into which of the two they prioritize.

Effects of predictability during production
Phonetic cues are often weakened for segments that are predictable from a given context.Exemplar-based approaches to phonology (Bybee, 2006;Ernestus, 2011;Pierrehumbert, 2001) have long noted that it is often the most frequent lexical items that are targeted for reduction due to their predictability.Building on this line of research, Hall, Hume, Jaeger, and Wedel (2016) argue that phonological systems tend to reduce segments in predictable contexts because enhancing the cues would require additional effort while contributing little to successful lexical access (by the listener).For example, word-final coda contrasts are often neutralized cross-linguistically because segments become more predictable as the listener processes more and more of the target item during lexical access.This means that word-final codas contribute less to identifying the target lexical item.Rather than enhancing the weak cues of an already predictable segment, phonological systems choose to enhance cues of segments in unpredictable positions instead, such as in the case of word-initial obstruent aspiration in English (e.g., /p͟ ik/ → [p h ik] vs. /sp͟ ik/ → [sp͟ ik]).Predictability, specifically phonotactic predictability has also been shown to have similar effects on high vowel devoicing in Japanese.Shaw and Kawahara (2018) found that devoiced high vowels in Japanese often deleted completely, leading to consonant clusters, and Whang (2018) found that predictability had a noticeable effect on the likelihood of deletion, where highly predictable vowels deleted while less predictable vowels retained oral gestures to provide coarticulatory cues.

Effects of phonotactic knowledge on speech perception
If it is the case that speakers are varying the amount of phonetic cues depending on the target segment's predictability in a given context, the question that naturally follows is whether listeners similarly vary their attention to phonetic cues based on predictability.Numerous studies have shown that expectations that stem from language experience affect how listeners utilize phonetic cues.For example, listeners are often insensitive to phonetic cues that are not contrastive in their native language.French listeners have difficulty contrasting short versus long vowels (Dupoux et al., 1999), English listeners have difficulty perceiving tonal contrasts (So & Best, 2010), Japanese listeners have difficulty contrasting /l/ versus /r/ because neither are phonemes of the language (Flege, Takagi, & Mann, 1996), and so on.Conversely, listeners are also attuned to cues that are useful in their native language.For example, Korean listeners are more sensitive to V-to-C formant transition cues than English listeners (Hume, Johnson, Seo, Tserdanelis, & Winters, 1999) because coda obstruents are obligatorily unreleased in Korean while they are optionally released in English (Kang 2003), making the transitional cue more useful to Korean listeners for recovery of the coda consonant than to English listeners, who have the option of waiting for the release of the coda obstruent.
Language experience also shapes the perception of phonetic cues.Pitt and McQueen (1998) showed that listeners are biased towards identifying phonetically ambiguous segments as segments with higher phonotactic probability (i.e., phonotactically more predictable).Phonotactic knowledge also seems to play an important role in other domains as well.When processing nonce words, sequences with higher phonotactic probabilities are processed faster (Vitevitch & Luce, 1999;Vitevitch, Luce, Charles-Luce, & Kemmerer, 1997).On the other hand, lexical items with high phonotactic probabilities are processed slower than lexical items with low phonotactic probabilities, presumably because high phonotactic probability in lexical items means that there are also that many more similar lexical items, utimately slowing down lexical access (Marslen- Wilson, 1987;McClelland & Elman, 1986;Norris, 1994;Vitevitch & Luce, 1998).The question then is, which process takes precedence?The answer to this question seems to depend on the task.In general, listeners seem to prioritize the use of lexical knowledge, relying on their phonotactic knowledge only when lexical activation fails (Shademan, 2006;Vitevitch & Luce, 1999; although see Myers, 2015, and related works showing significant contributions of neighborhood density on wordlikeness judgments in non-English languages).Additionally, Mattys, White, and Melhorn (2005) investigated whether participants pay more attention to lexical or sublexical (segmental and prosodic) segmentation cues when they are in conflict.The results showed again that lexical cues are prioritized and that listeners rely on sublexical cues when lexical context or information cannot be accessed due to noise or absence.The current study adds to this line of work on 'top-down' effects by investigating the interaction between two kinds of sublexical information, namely phonotactic predictability and finegrained phonetic cues, using high vowel epenthesis in Japanese as a test case.

Perceptual repair by Japanese listeners
In the now well-known study commonly referred to as the 'ebzo test' (Dupoux et al., 1999), French and Japanese speakers were presented with acoustic stimuli with the high back rounded vowel [u] of varying durations ranging from 0 ms to 90 ms occurring between two consonants (e.g., [ebzo] → [ebuːzo]).The stimuli were designed so that when there is no vowel in the stimuli, the result is a sequence that is phonotactically legal in French but illegal in Japanese.Their results showed that while French speakers could accurately distinguish the vowel-less from vowel-ful tokens, Japanese speakers were essentially 'deaf' to such differences, erring heavily towards misperceiving what the authors call an 'illusory' vowel.On the other hand, French speakers were unable to accurately perceive vowel length, with which the Japanese participants had little trouble perceiving.The authors propose that the results are due to phonotactic differences in French and Japanese, where Japanese listeners perceive a non-existent vowel between two consonants because Japanese phonotactics disallows heterorganic consonant clusters.French listeners, on the other hand, were insensitive to vowel length because it is not contrastive in French.The authors further argue that there is a 'top-down' phonotactic effect on perception, where phonotactically illegal sequences are automatically perceived as the nearest legal sequence rather than repaired at a higher, abstract phonological level.
Dehaene-Lambertz, Dupoux, and Gout (2000) also tested the illusory vowel epenthesis effect in an event-related potential (ERP) study.In this study, Dehaene-Lambertz et al. (2000) carried out experiments similar to that of Näätänen et al. (1997), where electrophysiological responses have been shown to be sensitive to phoneme categories.Dehaene-Lambertz et al. looked at how mismatch negativity (MMN) responses in Japanese and French speakers differ in the absence versus presence of a vowel in the same kind of sequences as those in Dupoux et al. (1999).The experiments followed an oddball paradigm where in one trial a sequence that is legal in both languages was presented as the standard (e.g., [iɡumo]) and one that is illegal only in Japanese as the deviant (e.g., [iɡmo]).The reverse was presented in a separate trial.Although the results reported collapsed the trials, the ERP results generally showed that Japanese speakers are insensitive to the differences between the vowel-ful and vowel-less items, while French speakers are, supporting the behavioral results from the original study by Dupoux et al. (1999).A related fMRI study by Jacquemot, Pallier, LeBihan, Dehaene, and Dupoux (2003), also found similar but slightly weaker results.Jacquemot et al. report that in an AAX task (A-stimulus presented twice before X-stimulus), neural activity increased whenever the X stimulus was different from the A stimulus for both Japanese and French participants.This was true regardless of whether or not the acoustic difference was phonologically contrastive in the language, although neural activation was significantly greater when the acoustic contrasts were also phonologically contrastive.
A more recent study by Dupoux et al. (2011) aimed to further bolster the automatic perceptual repair idea by also investigating European Portuguese, Brazilian Portuguese, and Japanese listeners.The reason for choosing the two dialects of Portuguese was that European Portuguese allows the same types of clusters as French, but Brazilian Portuguese has a strict CVCV phonotactic structure, leading to the expectation that their perception would be similar to that of Japanese listeners.The crucial difference between Brazilian Portuguese and Japanese is that in the former, the default epenthetic vowel is reported to be /i/ as opposed to the Japanese /u/.Since the quality of the epenthetic vowels are different in the two epenthesizing languages, the experiments were modified slightly from the 1999 study to enable identification of the perceived illusory vowels in the results.Like French listeners, the results showed that European Portuguese listeners did not have trouble distinguishing vowel-less from vowel-ful tokens.Japanese listeners, again, showed a tendency towards mistakenly recovering /u/ between consonant clusters.The results, however, additionally showed that Japanese listeners were also sensitive to i-coarticulation in the first consonant (i.e., eb j zo), recovering /i/ rather than /u/.By comparison, Brazilian Portuguese listeners tended to perceptually recover /i/ between illegal consonant clusters by default as expected, but did not show the same degree of sensitivity to u-coarticulation.Although the reasons for the disparity in sensitivity to coarticulatory cues were not discussed, the difference is likely due to Brazilian Portuguese listeners having little experience with a systematic high vowel devoicing process, leading them to underutilize coarticulatory cues relative to Japanese listeners.

Problems and solutions
The series of studies discussed above collectively suggest that there is a top-down imposition of the listeners' native phonotactic grammar during perception.The experiments, however, would benefit from two particular refinements when considering Japanese listeners: using stimuli that are less foreign to Japanese listeners and controlling for the effects of high vowel devoicing in how Japanese listeners perceive certain consonant clusters.
First, the waveform and spectrogram examples of the stimuli used in the studies by Dupoux and colleagues reveal that the burst of C 1 (e.g., [b] in [ebzo, ebuzo]) were rather long, potentially biasing the participants to perceive a vowel.For example, Dupoux et al. (2011) show that in a sequence like [agno], the voiced stop had a burst of at least 50 ms and contained formant-like structures.Japanese voiced stops, however, typically have burst durations of less than 20 ms (Kong, Beckman, & Edwards, 2012).In addition, Japanese high vowels are inherently short, with an average duration of approximately 40 ms, but they can be as short as 20 ms (Beckman, 1982;Han, 1994).Taking the short burst and vowel durations of Japanese together, an atypically long burst with formant structures can be interpreted as containing a vowel, possibly confounding the independent effects of acoustic cues and phonotactic violations (Wilson, Davidson, & Martin, 2014).Furthermore, the closure duration of the voiced stop was also nearly 100 ms, which is closer to the geminate range than the singleton range in Japanese (Kawahara, 2006).Geminates are not known to affect high vowel devoicing in C 1 position, but geminate consonants in C 2 position have been shown to increase the likelihood of preceding vowels being phonated in Japanese regardless of whether the consonants are voiced (Fujimoto, 2015;Maekawa & Kikuchi, 2005).This means that stimuli with geminate-like obstruents in both C 1 and C 2 positions (e.g., [iɡba]) could have further biased Japanese participants towards expecting a vowel in the target context.While this is also a tendency that is phonotactically driven, it is unclear whether the primary driving force behind perceptual epenthesis in the experiments is the heterorganic clusters, the phonetic cues of geminatelike segments, or a combination of both.
Second, the stimuli used in the ebzo tests included a mix of environments in which high vowel devoicing is expected to occur in Japanese as well as non-devoicing environments.The results reported in these studies, however, make no distinction between the two types of environments.Japanese high vowel devoicing is a highly productive process that applies at rates above 80% in most contexts (Maekawa & Kikuchi, 2005), where high vowels lose at the least their phonation and at most delete completely when between voiceless obstruents (e.g., /masutaa/ → [mastaa] 'master'; Shaw & Kawahara, 2018).Given the life-long experience Japanese listeners have in recovering the often-deleted high vowels between consonant clusters, it is very likely that this phonological process has an effect that is independent of phonotactic violations in creating an expectation for a vowel, and the most straightforward remedy to this issue is to test and analyze devoicing and nondevoicing environments separately (e.g., [ezpo] vs. [espo]).
Furthermore, related to the division of devoicing and non-devoicing stimuli, the devoicing stimuli can be divided into low-and high-predictability sub-groups.Varden (2010) states what seems to be a prevalent assumption in the literature on Japanese high vowels, which is that since high vowels trigger allophonic variation for /t, s, h/ in the language (i.e., /t/ → [t∫i, tsu]; /s/ → [∫i, su]; /h/ → [çi, �u]), the high vowels need not be acoustically present in these contexts because they can be predicted with certainty from the allophonic consonant alone.Conversely, this also means that in environments where allophonic variation is not triggered (i.e., /p, k, ∫/ → [pi, pu, ki, ku, ∫i, ∫u]), Japanese listeners would be more inclined to pay closer attention to the phonetic cues because they cannot predict the vowel with certainty.This effect of predictability on the perception of high vowels has long been assumed but never tested systematically.The current study, therefore, presents a perception experiment that specifically controls phonotactic predictability and investigates how it affects Japanese listeners' utilization of coarticulatory cues.
It should be noted that although [t∫, ts], [∫, s], and [ç, Φ] are traditionally analyzed as allophones of /t, s, h/ before /i, u/, respectively as discussed above, the current study regards them as phonemes with extremely skewed phonotactic distributions to more accurately reflect how the sounds are used in Japanese today.For example, minimal loan pairs such as [tiaː] 'tier ' and [t∫iaː] 'cheer' suggest that words like 'cheer' contain an underlying /t∫/ that surfaces faithfully, rather than an underlying /t/ that undergoes allophony.In fact, an analysis of the Corpus of Spontaneous Japanese (Maekawa & Kikuchi, 2005, see Table 2 below for more details) revealed that with the exception of /ts/, which still only precedes /u/, all other 'allophones' can now precede all or most vowels.

Materials and methods
The stimuli for this study are in the form V 1 C 1 (V T )C 2 V 2 , where V T is the target vowel and C 1 and C 2 are determined based on the stimulus group the token belongs to.The stimuli were divided into three groups: non-devoicing (NoDevoice) where vowel devoicing is not expected, low predictability (LoPredict) where both high vowels can occur and devoice, making coarticulatory cues necessary for recovery of a devoiced vowel, and high predictability (HiPredict) where phonotactic predictability is sufficient for recovery of a devoiced vowel, making coarticulatory cues less important.Below in Table 1 are the stimuli.Note that although it is more accurate to use the IPA symbols for the voiceless alveopalatal fricative [ɕ] and affricate [ʨ] in Japanese, the palatoalveolar symbols [∫, t∫] are used throughout the current study to make [∫] more visually distinct from the palatal fricative [ç] and to make [t∫] consistent in place with [∫].There were 252 stimulus items in total.The stimulus forms shown in Table 1 were first recorded by a trained, non-Japanese-speaking, English-Hungarian bilingual phonetician in a sound-attenuated booth with stress on the initial vowel and with /i, u, a/ as target vowels (V T )./a/ was included as a target vowel because it is a low vowel that typically does not devoice in Japanese, and also to test whether Japanese listeners are sensitive to coarticulatory cues of all vowels or just high vowels.Attempts were made to record the stimuli with two native Japanese speakers, but both speakers had difficulties keeping high vowels voiced in devoicing contexts, and even when they were successful in producing voiced high vowels in devoicing contexts, either the burst durations were too short to manipulate or the target vowel was stressed.
For each recording, the target vowels were manipulated by inserting or removing whole periods to achieve a duration of ∼40 ± 5 ms.From each of the recordings, four additional tokens were created by removing from right to left, half of V T (splice-1), the remaining half of V T (splice-2), half of the C 1 burst/frication noise (splice-3), then the remaining half of the C 1 burst/frication noise leaving only the closure for stops and ∼15 ms for fricatives (splice-4).An example of how the splicing was done is shown in Figure 1 below with the token [ekuto].
The result of the splicing process is a gradual decrease of vowel coarticulatory information available in the burst/frication noise of C 1 .Stop bursts in particular were manipulated to test whether it is phonotactic predictability or interpretation of phonetic information that drive illusory vowel epenthesis, since sensitivity to and interpretation of stop bursts as signaling the presence of a vowel are reported not just in Japanese (Furukawa, 2009;Whang, 2016) but in Korean (Kang, 2003) and English (Davidson & Shaw, 2012;Hsieh, 2013) as well.Naturally produced, vowel-less tokens were also recorded for each stimulus form to test how it differs in perception from the spliced vowel-less stimuli (splice-2), which have traces of coarticulation from the target vowel on the surrounding consonants.

Participants
Twenty-nine monolingual Japanese listeners (16 women, 13 men) were recruited for the perception experiment in Tokyo, Japan.All participants were undergraduate students born and raised in the greater Tokyo area and were between the ages 18 and 24.Although all participants learned English as a second language as part of their compulsory education, none had resided outside of Japan for more than six months or had been overseas within a year prior to the experiment.All participants were compensated for their time.

Procedure
The experiment follows the forced-choice vowel labeling task from Dupoux et al. (2011).The participants were told that they would be listening to foreign words over headphones and that they would have five seconds to choose one of four answer choices that best matches the word they heard.The stimuli were presented through noise-isolating headphones, and answer choices that gave the vowel-less and various vowel-ful spellings of the stimulus that just played were presented on screen simultaneously (e.g., [epuko] → <epko>, <epako>, <epiko>, <epuko>, where the <> notation indicates orthographic representation).
Participants selected their answer choices by using arrow keys on a keyboard (i.e., ↑ ↓ ← →).A typical answer-choice screen is shown below in Figure 2.
While it is true that Japanese orthography is a syllabic system, most Japanese speakers are comfortable with the Latin alphabet, not only because of frequent exposure to loanwords but also because of the keyboards used for word processing.There are currently two main input methods-direct input (one key = one syllabic character) and conversion (QWERTY keyboard used to input CV combinations which are then converted to the corresponding syllabic character)-and the conversion method is commonly more preferred, and thus participants are expected to be comfortable with answer choices presented in the Latin alphabet.The experiment was designed to continue as soon as the participant made an answer choice.

Defining predictability
Before discussing its effects, predictability should be defined.For the purposes of this paper, predictability is quantified using two Information-Theoretic measures: surprisal and entropy (Shannon, 1948).All calculations presented in this section are based on phonemic transcriptions of all observed biphones in the 'Core' subset of the Corpus of Spontaneous Japanese (hereinafter CSJ-Core; Maekawa, 2003;Maekawa & Kikuchi, 2005).The CSJ- Core consists of approximately 45 hours of recorded spontaneous speech (∼500K words) obtained primarily from academic conferences and provides both orthographic and phonemic transcriptions of the recordings.For example, the word 'mouth' is represented orthographically as <口> and phonemically as <k u cj i>, where <cj> is a 'phonetically palatalized' voiceless alveo-palatal affricate (i.e., /kut∫i/).There is also a 'phone' level transcription that codes the devoicing status of vowels, but this level is not used in this paper under the assumption that both phonetically devoiced and voiced vowels are phonemically the same (e.g., [ku̥ t∫i, kut∫i] → /kut∫i/).Furthermore, the corpus distinguishes 'phonetically palatalized' <Cj> and 'phonologically palatalized' <Cy> segments in its phonemic transcriptions, but <Cj> segments occur exclusively before /i/ while 'phonologically palatalized' <Cy> segments occur before all other vowels, suggesting that <Cj> is really a notation for allophones that result from i-coarticulation.As already discussed above, these 'allophones' behave more phonemically in Japanese today, and thus <Cj> segments were recoded as their phonemic counterparts <Cy> before calculating phonotactic measures.To provide an overview of how often each of the five vowels in Japanese follow the consonants used in the current study, presented first below in Table 2 are Observed/Expected (O/E) ratios for all pertinent C 1 V biphones.O/E ratios compare how often each biphone actually occurs in the data (Observed) to how often each biphone should have occurred if all segments are assumed to have an equal likelihood of combining to form biphones (Expected).The observed number is divided by the expected number, and the resulting O/E value indicates the magnitude of a given biphone's over-/underrepresentation. For example, O/E ratio of 2.0 indicates that the biphone occurred twice as often as expected, while O/E of 0.5 indicates that the biphone occurred half as often as expected.
The O/E values show that /u/ is highly overrepresented in Japanese after most consonants.The exceptions are /ɡ/ after which /a/ is the most common vowel, and /∫, ç/ after which /i/ is the most common vowel.
Surprisal and entropy are both calculated based on the conditional probabilities of vowels after a given consonant (i.e., (Pr(v | C 1 _)).Surprisal is the negative log 2 probability (-log 2 Pr(v | C 1 )) and indicates the amount of information (effort) necessary to predict a vowel after a given C 1 .Entropy is the weighted average of surprisal in a given context and indicates the overall level of uncertainty in a given context due to competition among other possible vowels.If at least one of the consonants is voiced in a given C 1 C 2 sequence, it is not a high vowel devoicing environment, and thus Japanese listeners must consider all five vowels.The entropy values of /b, ɡ, z/ and the respective surprisal values for each vowel in these contexts are provided below in Table 3.There is no upper bound to surprisal, but for a five vowel system the maximum entropy is 2.322 (-log 2 p(0.2)),where all vowels occur with equal probabilities (1/5 = 0.2)./u/ has the lowest surprisal after /b, z/ but /a/ has the lowest surprisal after /ɡ/, reflecting the results of Table 2.If both consonants are voiceless in a given C 1 C 2 sequence, it is unlikely that Japanese listeners would consider all five vowels.Instead, experience with high vowel devoicing informs the Japanese listener that there is most likely a phonemic /i, u/ that was devoiced between the two voiceless consonants.The theoretical maximum of entropy (highest uncertainty) in any given consonantal context with two possible vowels is 1.000 (-log 2 p(0.5)),where both vowels occur with equal probabilities (1/2 = 0.5).Presented below in Table 4 are entropy and surprisal measures recalculated for just high vowels.Generally, surprisal for /u/ is lower for most contexts, with the exceptions of /∫, ç/ after which /i/ had lower surprisal./i/ was also shown to be the most frequent vowel in these contexts in Table 2 above.Additionally, entropy and surprisal never reach zero, showing that all consonants allow both /i, u/ to follow.
Given the surprisal and entropy measures, listeners are most likely to predict vowels with low surprisal overall, but the listener is also more likely to consider other vowels in high entropy (uncertainty) environments.To give a concrete example, given a 'high predictability' context such as after /ç/, the listener can predict the vowel to be /i/ with near-zero effort (surprisal = 0.001) and have near-absolute certainty about the prediction (entropy = 0.012).In 'low predictability' contexts such as after /p/, although /u/ is the more likely vowel, the high level of uncertainty would lead the listener to listen for phonetic cues that either support or contradict the context-based prediction.In other words, phonetic cues for high vowels can be used to counteract uncertainty.

Analysis and predictions
All statistical analyses were performed by fitting linear mixed effects models using the lme4 package (Bates, Mächler, Bolker, & Walker, 2015) for R (R Core Team, 2016).The statistical analyses assess vowel detection and vowel identification.Detection refers to how often participants report perceiving any vowel at all both in the presence and absence of vocalic segments.Identification refers to whether the vowel the participants perceive is in agreement with the acoustic vocalic information contained in the stimuli.
In the case of detection, accuracy is expected to be higher for the NoDevoice group (e.g., [ez_po]) and lower in the LoPredict and HiPredict groups (e.g., [e∫_po] and  [es_po], respectively).Since the phonological process of high vowel devoicing is nearly obligatory (Vance, 1987), it could bias Japanese speakers to expect a high vowel to be present between two voiceless consonants even when it is acoustically absent.Since the current experiment uses nonce-words, there is no underlying or lexical form to access.Devoiced and voiced sequences involving two voiceless obstruents in Japanese would map to the same phonotactically legal underlying form (e.g., [esupo] ≡ [esu̥ po] ≡ [espo] → /esupo/).The devoiced and voiced sequences would all be regarded as legal, and the actual presence or absence of the vowel in the signal readily ignored.While devoicing is possible in the NoDevoice environments, it is extremely rare (Maekawa & Kikuchi, 2005).Since only the vowel-ful token is legal in the language in non-devoicing contexts, a devoiced or vowel-less counterpart is not in an equivalence relationship (e.g., [sude] ≢ *[su̥ de] ≢ *[sde] 'barehand').Thus Japanese listeners are expected to be more sensitive to the presence versus absence of a medial vowel.Furthermore, regardless of the stimulus group, higher accuracy is expected in recognizing that there is no vowel as the burst/frication noise gets shorter, especially when there is no burst present.
In the case of identification, high accuracy is expected for the LoPredict group and lower accuracy for the HiPredict group.Japanese speakers have been shown to be sensitive to high vowel coarticulation in /∫/ (Beckman & Shoji, 1984), but this sensitivity is only useful when the vowel is unpredictable after a given C 1 (i.e., LoPredict group).Japanese listeners, therefore, should be sensitive at least to i/u-coarticulation in the LoPredict group but biased towards a single high vowel that most frequently follows C 1 in the HiPredict group regardless of coarticulation.Since there are four answer choices <i, u, a, ∅>, identification rates are expected to be at least 50% in the LoPredict group and approximately 25% in the HiPredict group.Furthermore, since /a/ rarely devoices in Japanese, <a> responses should be relatively low even for a-coarticulated tokens, defaulting instead to the most phonotactically probable vowel.The NoDevoice group is expected to show some effects of coarticulation, as was the case in Dupoux et al. (2011), but like the HiPredict group, /a/ should show little effect.
These predictions contrast with the account given by Dupoux and colleagues.According to Dupoux and colleagues, there are two mechanisms at play during illusory vowel epenthesis.First, perceptual repair is a one-step process where phonotactically illegal sequences are perceived as their repaired counterparts rather than being perceived accurately first then repaired to their phonotactically legal counterparts.What this means is that listeners do not have access to the source language's underlying form, making heterorganic C 1 C 2 sequences and their repaired C 1 VC 2 sequences equivalent for Japanese listeners.If this is correct, the prediction in terms of detection is that the rate of vowel detection between C 1 C 2 and C 1 VC 2 sequences should be statistically the same since the two sequences are equivalent.
Second, although Dupoux and colleagues argue that perceptual repair is triggered by phonotactic violations, the authors propose a phonetic repair strategy, where Japanese listeners default to epenthesizing /u/ because it is the shortest vowel in the language, whereas Brazilian Portuguese listeners epenthesize /i/ instead for the same reason (Dupoux et al., 2011).If the choice of the epenthetic segment is indeed based on the magnitude of phonetic change rather than phonotactic probability, no observable effect of phonotactic predictability is expected, since phonotactic knowledge merely flags repair sites but is not involved in the repair itself.If the repair strategy argued by Dupoux and colleagues is correct, vowel identification rates are predicted to suffer across all contexts whenever the coarticulated vowel is not /u/.

Results
Shown in Figure 3 below are the overall results of the experiment.Figure 3A shows results for all C 1 and Figure 3B for stop C 1 only, which consequently also results in the exclusion of all high-predictability tokens, since /Φ, s, ç/ are all fricatives.The colors indicate the target vowels, and the solid and dashed lines indicate vowel detection and successful vowel identification rates, respectively.Vowel detection rates simply collapse all nonzero responses, whereas vowel identification rates only include cases where participant responses matched the coarticulated vowels in the stimuli (e.g., respond <epuko> for [epu̥ ko]).The smaller the distance between two lines of the same color, the higher the proportion of successful vowel identification.
Figure 3A and B are qualitatively similar, where detection and identification rates fall as more of the C 1 V T information is spliced, and the most noticeable effect of including fricatives in 3A is that identification rates are driven lower.In both figures, there are three things that stand out.First, detection rates for /u/ never reach 100% even when there is a full vowel of 40 ms present in the stimuli, suggesting that there is confusion between the presence and absence of /u/.Second, vowel detection rates never quite reach 0%, remaining above 20% even in the absence of any C 1 burst noise (Figure 3B, splice-4), suggesting an overall confusion between vowel-fulness and vowel-lessness.Third, /a/ identification rates (blue dashed line) fall the most dramatically and are the lowest in tokens where the medial target vowel is spliced out, suggesting that only high vowels are potentially available for recovery.
Because the results of splice-1 and splice-3 tokens show no surprising trends, the rest of this paper will focus on the splice-0 (full-vowel), splice-2 (no vowel), and splice-4 (no vowel and no C 1 burst/frication) results.The splice-2 results will also be compared against naturally produced vowel-less tokens to test how the presence of coarticulatory cues affect the responses.

Tokens with full medial vowel
Shown below in Figure 4 are vowel identification rates for tokens with a full target vowel of 40 ms, broken down by context and by C 1 .The figure shows that the identification rates for /i, a/ are essentially at ceiling, but identification rates for /u/ are surprisingly low at below 90%.
The most common wrong response by the participants for [u] identification was ∅ for all C 1 as shown in Figure 5, meaning that the participants either heard the vowel accurately or confused [u] with ∅, but rarely confused the vowel with another vowel.The confusion specifically between /C 1 C 2 / and /C 1 uC 2 / sequences suggests two things.First, [C 1 uC 2 ] can be mapped to both /C 1 uC 2 / and /C 1 C 2 /, although there is a bias towards the former.Second, the fact that there is confusion between /C 1 uC 2 / and /C 1 C 2 / even when a vowel of 40 ms is fully present suggests that the distinction between the two sequences is weak, and that C 1 C 2 and C 1 uC 2 sequences are treated as more or less equivalent by Japanese listeners.While this provides some support for the account presented by Dupoux and colleagues, the participants also exhibit some confusion between /i/ and ∅ after /∫, ç/, which most likely stems from /i/ being the most common vowel after these consonants, which was shown in Table 2.
The results of the full-vowel tokens suggest that stimuli such as [epko, epu̥ ko, epuko] are possibly all being treated as equivalent to /epuko/.Because they all map to the same phonotactically legal structure, there is bidirectional repair, although with a bias towards vowel recovery.The fact that there is confusion for /u/ across the board, even for /ɡ/ despite /a/ being the most common vowel to follow, provides some support to the phonetically minimal repair hypothesis presented by Dupoux and colleagues.However, the fact that there is also confusion for /i/ after /∫, ç/ additionally suggests that phonotactic probability affects perception as well.

Tokens with no medial vowel
This section compares the results of naturally vowel-less tokens and the splice-2 tokens where the medial, phonated vocalic material has been completely removed but C 1 burst/frication noise fully remains.Acoustically, the difference between these tokens is that the naturally vowel-less tokens contain no obvious coarticulatory information, unlike the spliced tokens.

Naturally vowel-less tokens
The prediction in terms of vowel detection was that the rate of <∅> responses should be highest for non-devoicing contexts since high vowel devoicing is rare in these contexts  making Japanese listeners more sensitive to the presence versus absence of a medial vowel.Conversely, the rate of <∅> responses was expected to be low in contexts where high vowel devoicing is expected, since high vowels can delete in these contexts, leading to a bias towards recovery.This bias should be especially high in HiPredict contexts because one of two high vowels can be predicted with near-absolute certainty, leading Japanese listeners to disregard phonetic cues that may contradict their contextual vowel prediction.
Presented first below in Table 5 are the responses for naturally produced VCCV tokens.Bold numbers indicate the most frequent responses for a given C 1 .A chi-square test was performed using the chisq.test()function in R to test whether the observed response rates were significantly different from chance./a/ responses were excluded under the assumption that /a/ is not a candidate for recovery and also because /a/ responses were at or near 0% in most contexts.The results showed that the observed responses were significantly different from chance at p < 0.01 with the exception of /p/ (p = 0.4909).
Overall, the results show that <∅> responses are 50% or lower across all contexts, revealing an overall bias towards perceptual repair.However, the rate of <∅> responses is highest for NoDevoice environments as predicted, suggesting that there indeed is an effect of high vowel devoicing.Additionally, <∅> responses are lowest for HiPredict environments, confirming the prediction that predictability has an effect on the rate of repair as well.
The responses to naturally vowel-less tokens also suggest that there is an effect of phonotactics that drives the choice of vowel that is recovered by Japanese listeners.The vowel recovered after [∫, ç] is again /i/ rather than /u/, further strengthening the account that the choice of the vowel used for phonotactic repair is not just merely a default, minimal vowel but rather chosen based on phonotactic probability.This is also in line with a recent finding by Durvasula and Kahng (2015), who also found in Korean listeners that the choice of recovered vowel is better predicted by the phonological alternations observed in the language rather than a phonetically minimal repair strategy.

Spliced vowel-less tokens (Splice-2)
Shown below in Figure 6 are vowel identification rates for tokens with all of the vowels spliced out, broken down by context and C 1 .The figure shows that the identification rates for high vowels are highest for low-predictability contexts but remain above 40% in the other two contexts.Identification rates for /a/ are generally lower than for high vowels across all contexts, but are clearly lowest in high-predictability contexts.
Two other predictions going into the current study were that participants should be more sensitive to the coarticulatory cues of high vowels than non-high vowels, and that this sensitivity should be highest in low-predictability devoicing contexts.A mixed logit model was fit using the glmer function of the lme4 package of R, with successful vowel identification rates as the dependent variable to test these predictions.The statistical analysis compares the rate of correct identification of spliced vowels from coarticulatory cues, so naturally produced VCCV tokens, which should contain no vowel coarticulatory cues, are not included in the analysis.The fixed effects structure of the model consisted of target vowel, context, and their interaction.The target vowel factor was coded with treatment contrast coding with the low vowel /a/ as the baseline, to test whether identification rates for the high vowels /i, u/ indeed were significantly higher than the non-high vowel.
The context factor was coded with sum contrast coding, so that each level is compared to the grand mean of vowel identification rates across all contexts.Two models were fit for this analysis (and all subsequent analyses) because regression models only return N-1 coefficients, where N is the number of levels in a factor group.Although the coefficient of a missing level can be calculated by hand, its significance level is not examined, making a second model with a different reference level necessary to thoroughly explore the missing coefficient from the first model (Clopper, 2013).Tables 6 and 7 show the results of a model with 'HiPredict' and 'NoDevoice' as reference (and thus missing), respectively.Aside from the changed reference levels, the models show nearly identical results.The model with a fully-crossed, maximal random effects structure failed to converge, hence the   final random effects structure included by-participant and by-stimulus random intercepts as well as by-participant random slopes for target vowel.The interaction was shown to be a non-significant contributor to the fit of the models (p = 0.4546), and thus was excluded from the final models.Note that the diacritic for devoicing (i.e., V̥ ) is used throughout to indicate the vowels that have been spliced out.
The results confirm both predictions.First, the rate of successful vowel identification is significantly higher for both high vowels [i ̥, u̥ ] than the [ḁ] baseline.Second, the rate of successful vowel identification was significantly higher than the grand mean in LoPredict contexts but significantly lower in HiPredict contexts.NoDevoice contexts on the other hand did not significantly differ from the grand mean.

Comparison of naturally vowel-less and spliced vowel-less tokens
Naturally vowel-less tokens and spliced tokens by themselves tell only part of the story.Another prediction was that Japanese listeners should be able to recover high vowels from the coarticulatory information in spliced tokens, leading to differences between splice-2 and naturally vowel-less tokens.If it is the case that phonotactic violation alone is responsible for vowel epenthesis and that the choice of vowel is the phonetically minimal segment, namely /u/, then the presence of vowel coarticulatory information should do little to affect the choice of vowel.
Shown in Table 8 below are the results of two mixed logit models comparing the detection rates for spliced tokens to the grand mean of naturally vowel-less tokens across all contexts.As was the case with the analysis of spliced tokens presented in Tables 6  and 7 above, the results of the two models returned nearly identical coefficients and identical associated significance testing results aside from the changed reference levels.Since most of the second model's results simply repeat those of the first model, Table 8 combines the two models' results into a single table, reporting all the coefficients of the The results show that compared to the naturally vowel-less baseline, i-coarticulated tokens but not u-and a-coarticulated tokens drive up the vowel responses significantly.Additionally, the significant effects of context suggest that Japanese listeners are less likely to perceptually epenthesize a vowel in NoDevoice contexts and more likely in HiPredict contexts.
For identification rates, spliced tokens are compared separately to naturally vowel-less tokens to make the effects of coarticulation for each vowel clearer.Presented first in Table 9 below are the responses for spliced [u] tokens, and Figure 7 shows how the responses compare to naturally vowel-less tokens.
Two mixed logit models were constructed with the rate of <u> responses as the dependent variable.The predictors were stimulus type (i.e., spliced vs. natural), C 1 , and their interactions.C 1 was used as a predictor rather than context because the epenthetic vowel does not seem to be uniform across all contexts but rather seems to depend on C 1 .The C 1 factor was coded with sum contrast coding.By-participant and by-stimulus random intercepts were included.By-participant random slopes for target vowel and C 1 were also included.All predictors were significant contributors to the fit of the model.The model results are shown below in Table 10, with the grand mean of ∅ tokens (i.e., naturally vowel-less) across all C 1 as the baseline.As with Table 8, the results of two models are combined into a single table with coefficients obtained from the second model in italics.
The model results show that the presence of a coarticulated vowel does indeed significantly raise the overall rate of <u> responses compared to naturally vowel-less tokens.The model also shows clearly that none of the consonants from the NoDevoice context nor their interaction with u-coarticulation have a significant effect on the overall rate of <u> responses.Just the NoDevoice consonants were analyzed separately, and the results  showed that u-coarticulation raised the overall rate of <u> responses (p = 0.00565), but this raising effect was driven by [b] tokens, where [bu̥ ] tokens had significantly higher <u> responses than [b∅] tokens (p = 0.0036).The other two consonants showed no significant effect of u-coarticulation.
In the case of LoPredict consonants, [p, ∫] drive down the rate of <u> responses for naturally vowel-less tokens, and the interaction terms show that this consonantal effect is significantly mitigated in [u] spliced tokens.This means that the high rate of <u> responses for spliced tokens remains high even when the preceding consonants are [p, ∫].LoPredict consonants were analyzed separately, and the results showed that u-coarticulation raised the overall rate of <u> responses (p = 0.0011).[pu̥ , ∫u̥ ] tokens had significantly higher <u> responses than [p∅ ∫∅] tokens (p = 0.0435, 0.0445, respectively), but [k] tokens showed no significant difference.
The overall model also shows that the HiPredict consonants [Φ, s] drive up the rate of <u> responses in naturally vowel-less tokens and also that this raising effect is mitigated for [u] spliced tokens.HiPredict consonants were also analyzed separately, and the results showed that vowel coarticulation did not have a significant effect overall (p = 0.55256).
Only [ç] tokens showed a significant effect, where [çu̥ ] tokens had significantly higher <u> responses than [ç∅] tokens (p = 0.0051).<i> responses were also driven up by i-coarticulation compared to the baseline of naturally vowel-less tokens.Shown below in Table 11 is a summary of the responses for spliced [i] tokens, and Figure 8 shows how the rate of <i> responses compare to naturally vowel-less tokens.As was the case with [u] tokens, vowel coarticulation seems to affect which vowel participants report to hearing.Similar models as in Table 10 were fit, with the same predictors and random effects structure.The dependent variable was <i> responses with the grand mean of naturally vowel-less tokens across all C 1 as the baseline.The results are shown in Table 12 below.
Again, i-coarticulation has a raising effect on the rate of <i> responses, but none of the NoDevoice consonants show significant effects.The NoDevoice consonants were analyzed separately, and the results showed that i-coarticulation raised the overall rate of <i> responses (p < 0.0001).All spliced tokens had significant higher <i> responses than naturally vowel-less tokens as well (p < 0.0001).
The overall results also show that [∫, ç] drive up the rate of <i> responses in naturally vowel-less tokens, which is unsurprising, since /i/ is the most probable vowel after these consonants.Separate models for LoPredict and HiPredict contexts were fit, and the LoPredict results showed that the difference between spliced and naturally vowel-less tokens were significant overall (p < 0.001).The raising effect of [i], however, was not significant for [∫] (p = 0.9998).The HiPredict results mirrored that of the LoPredict model in that the raising effect of [ç] in naturally vowel-less tokens was large enough to make the difference between spliced and naturally vowel-less tokens non-significant (p = 0.5608).
Thus far, the results suggest that the choice of epenthetic vowel for Japanese listeners is not simply a default /u/, but rather that the choice of vowel is sensitive to the acoustic cues in the signal./u, i/ are both high vowels that are targeted for devoicing in Japanese, so this is perhaps not surprising.Japanese listeners have had a lifetime of practice attending to subtle coarticulatory cues to recover devoiced and deleted high vowels.Then what about a vowel like /a/, which rarely undergoes devoicing?The responses to spliced [a] tokens are shown in Table 13 below, and Figure 9 shows how the responses compare between naturally vowel-less and spliced tokens.Although limited to post-stop environments (i.e., [b, ɡ, p, k]), the results show that participants can recover the spliced [a] vowel at relatively high rates.Bilabial place also seems to have a facilitatory effect.Given that <a> responses were generally low in naturally vowel-less tokens, the raising effect even in the limited environments is surprising.The fact that /a/ identification is limited to stops may be due to the articulatory differences between stops and fricatives.Because stops have a portion in which there is no airflow, coarticulation with the following vowel can be more complete by the time the stop burst/aspiration occurs.This is also true of bilabial place, where the lack of lingual gesture allows the following vowel to be coarticulated earlier.This is less true of fricatives where the transition into a fricative is more gradual, and coarticulation with the following vowel occurs towards the end of the segment rather than throughout.Since /a/ is a low vowel that a Japanese listener does not often have to recover, it may be that the beginning of the fricative already leads to the listener anticipating a high vowel and ignoring the low vowel cue towards the end.
Two mixed logit models with the same predictors and random effects structure as in Tables 10 and 12 were fit.Responses to [∫] tokens were removed from the model because /a/ responses were at 0% for both the naturally vowel-less and spliced [a] tokens, resulting in no meaningful difference.When included in the model, [∫] tokens had an extremely low intercept of -17, but an absurdly high standard error of 6,999, both of which are most likely errors stemming from an absolute lack of difference between participants.The interaction between target vowel and C 1 was not a significant contributor to the fit of the models and thus was excluded in the final models.The results are shown below in Table 14.The dependent variable was <a> responses with the grand mean of naturally vowel-less tokens across all C 1 as the baseline.
The results confirm that a-coarticulation indeed has a significant raising effect on the rate of <a> responses.The bilabial stops [b, p] have a significant raising effect on the rate of <a> responses, while the alveolar fricatives [z, s] have a significant lowering effect.
The models above all compared spliced tokens to naturally vowel-less tokens to investigate whether perceptual repairs by Japanese listeners are automatic as previously claimed and thus treated as the same.First, the vowel detection results reveal that Japanese listeners are more likely to perceive a vowel in devoicing contexts than in non-devoicing contexts.In high predictability contexts, where vowel detection is highest, the perceived vowel is the most probable high vowel (i.e., /Φu, su/ and /çi/).Because the detection rates are already high for naturally vowel-less tokens, the presence of coarticulatory cues did not significantly change the rate of vowel responses.In low predictability contexts, coarticulation effects on the rate of vowel identification were significant for both high vowels with the exception of /∫i/ tokens.There was also an unexpected effect of a-coarticulation, but this was limited to stop contexts.Together, the detection and identification results suggest that there is indeed a bias towards repairing phonotactically illegal consonant clusters, but the epenthetic vowel is chosen due to a combination of phonotactic predictability and sensitivity to phonetic cues.

Tokens with no vowel and no burst/short frication noise
The results discussed in Section 3.2 for spliced vowel-less but burst-ful tokens (splice-2) show that Japanese listeners are biased towards perceiving a vowel between heterorganic consonant clusters, and that the choice of vowel is sensitive to the coarticulatory cues present in the C 1 burst/frication noise.Numerous studies have shown that the presence of a stop burst or frication noise in phonotactically illegal sequences is often interpreted as signaling the presence of a vowel (see Davidson & Shaw, 2012, Hsieh, 2013 for English;Furukawa, 2009, Whang, 2016 for Japanese; Kang, 2003 for Korean).This section therefore discusses the results of splice-4 tokens, where the target vowel has been spliced out completely and C 1 also has been spliced out leaving just the closure for stop C 1 and <15 ms of frication noise for fricative C 1 .
The responses to all splice-4 tokens are summarized in Table 15 below.A mixed logit model was fit to test whether the rates of <∅, i, u, a> responses were significantly affected by the identify of the vowel that was spliced out.Stop C 1 and fricative C 1 were analyzed separately.The results revealed that the responses were not significantly different regardless of the target vowel, with the exception of spliced [u̥ ] tokens where C 1 was /b/, which drove up <u> responses (p = 0.002333).Because the effect was limited to a single consonant, this section collapses the responses across all target vowels and focuses more on vowel detection.
The results show first and foremost that the rate of <∅> responses never reaches 100%.This is perhaps expected for fricative C 1 , since there was ∼15 ms of frication remaining in the tokens.Factors contributing to the results for stop C 1 , on the other hand, are less obvious.A mixed logit model was fit separately for the stops and fricatives since the fricative tokens had a short frication noise remaining whereas the stop tokens had no burst at all.The full model for both data subsets had the following structures.The fixed effects included context, V 1 , and their interaction.All stimuli used in the experiment had the form V 1 C 1 (V)C 2 V 2 , where the order of V 1 -V 2 was always either [e-o] or [o-e].V 1 was included as a predictor to test whether the ordering of the initial and final vowels had a significant effect on vowel detection, which would suggest that there might be V-to-V coarticulatory cues that the participants are picking up on.The random effects included by-participant and by-stimulus random intercepts as well as by-participant random slopes for context, V 1 , and their interaction.Shown first below in Table 16 is the result of the final model for the stop-only subset.Since the HiPredict context had no stops, the subset only includes NoDevoice and LoPredict contexts with the latter as the baseline.The interaction was shown to be a non-significant contributor to the fit of the model (p = 0.5463) and thus was removed.
The results show that V 1 did not have a significant effect, but the rate of vowel detection was significantly higher for NoDevoice tokens than LoPredict tokens.A possible explanation for this effect is that the C 1 in NoDevoice tokens had consistent phonation during closure, as shown in Figure 10 below.
The mixed logit model for the fricative-only subset also showed that the vowel detection rate for the NoDevoice fricative [z] is significantly higher than for HiPredict fricatives although not higher than the LoPredict fricative [∫] (p = 0.658).For the fricatives, only context was a significant contributor to the fit of the model, and thus V 1 (p = 0.81919) and V 1 :Context (p = 0.82666) were excluded from the fixed effects structure of the final model.The results are shown below in Table 17.
Although the fact that vowel detection rates never fall to 0% can be easily explained by the presence of prevoicing for NoDevoice tokens and the 15 ms frication noise for the   Nevertheless, participants report perceiving a vowel some of the time.It is possible that some participants are picking up on the mismatch between the transitional cues out of V 1 and into V 2 .This seems unlikely, however, in that transitional cues into a vowel often outweigh transitional cues out of a vowel for Japanese listeners (Fujimura, Macchi, & Streeter, 1978) and that Japanese listeners rely more on centroid spectral cues than on formant transitions (Hirai, Yasu, Arai, & Iitaka, 2005).Perhaps a more likely explanation is one of task effect.Although the stimuli sounded as though they contain a geminate obstruent, there was no geminate option given as a possible answer.This might have kept the participants from fully eliminating the vowel-ful answer choices, and having been exposed to numerous vowel-ful tokens (both acoustically and perceptually) during the task, the participants might have assumed that a vowel should be present at least some of the time.

Summary of main findings
There were five main findings in the perceptual experiment.First, Japanese listeners seem to sometimes confuse the high vowel that is phonotactically the most likely after a given C 1 with ∅ even when the high vowel is 40 ms long and fully phonated.This sort of confusion was not observed with the low vowel /a/, which is typically not devoiced.Second, results from naturally vowel-less tokens revealed that the vowel most often perceptually epenthesized between illicit clusters is /u/, largely due to the fact that it is the vowel with lowest surprisal after most obstruents in Japanese.This is further supported by the finding that after /∫, ç/, which is most often followed by /i/ rather than /u/, the choice of epenthetic vowel is in fact /i/.Third, participants successfully identified spliced high vowels in splice-2 tokens (full C 1 with target vowel completely spliced out) at rates significantly higher than the baseline rates observed in naturally vowel-less tokens.Identification rates of spliced /a/ were significantly lower and limited to after stops.Fourth, related to the third finding, identification rates of high vowels were lowest in HiPredict contexts, suggesting that listeners are less sensitive to low-level coarticulatory cues in contexts where the phonotactics typically is sufficient for identifying the target vowel.Lastly, <∅> responses never quite reach 100% even for splice-4 tokens where both C 1 and target vowel were fully spliced out suggesting a bias towards CV structure, but this may have also been due to task effects.

Discussion
The aim of the current study was to test whether Japanese listeners utilize coarticulatory cues more in contexts where the listener is expected to have less certainty regarding their predicted vowel (low predictability) than in contexts where the vowel can be predicted with high certainty (high predictability).Broadly speaking, overt consonant clusters were shown to be mapped to a phonotactically legal CVC sequence, neutralizing the contrast between CC and CVC sequences as Dupoux and colleagues have shown.However, the specific vowel recovered was shown to be modulated by CV co-occurrence probabilities in Japanese and that CC and CVC were not perceived as equivalent.This difference in CC and CVC processing in Japanese listeners was also found in Cutler, Otake, and McQueen (2009).First, the perception of full-vowel tokens showed that there is confusion between /u/ and ∅, even when there is a 40 ms-long, phonated [u].It is possible that this confusion arises because /u/ is indeed the default epenthetic vowel in Japanese, making it equivalent to ∅.However, a survey of biphone co-occurrence probabilities in the Corpus of Spontaneous Japanese revealed that /u/ also happens to be the most common vowel after most consonants, making it difficult to attribute the default status of /u/ as stemming simply from its shortness (contra Dupoux et al., 1999Dupoux et al., , 2011)).Furthermore, similar confusion with ∅ is observed for /i/ after /∫, ç/, suggesting that the choice of epenthetic vowel must be conditioned by the phonotactic probabilities of the language.
Second, the perception of vowel-less tokens further suggests that Japanese listeners confuse vowel-ful and vowel-less tokens with a tendency towards vowel-fulness.The results for splice-4 (vowel-less and burst-less) tokens in particular showed that Japanese listeners interpret even the most minute acoustic cues such as prevoicing of stops as signaling the presence of a vowel (Section 3.3).However, participants do not seem to simply perceive a default vowel.A comparison between naturally vowel-less and spliced vowel-less tokens showed that spliced tokens drive up the rate of target vowel responses significantly.This suggests that while heterorganic C 1 C 2 sequences are perceived as being equivalent to C 1 VC 2 as Dupoux and colleagues argue, the particular vowel is again not simply the 'default' but the result of sensitivity to the acoustic information in the signal as dictated by the listener's native language.The participants, therefore, are recovering the vowel that is the most probable based on the phonetic cues contained in the burst/frication noise of C 1 .
Third, the rate of high vowel identification was above chance at 40% across all contexts in spliced vowel-less tokens.Specifically, recovery rates were the highest in LoPredict contexts as predicted, and the recovery rates were significantly lower for HiPredict contexts, also as predicted.Recovery rates in NoDevoice contexts fell somewhere between the two devoicing contexts.The high rates of recovery suggest that Japanese listeners are hypersensitive to vowel coarticulatory cues, and the lower rate of recovery in HiPredict contexts additionally suggests that sensitivity to coarticulatory cues are conditioned by phonotactic predictability.
Lastly, sensitivity to coarticulatory cues in Japanese listeners is limited primarily to high vowels.The participants were worst at identifying /a/.Non-high vowels are typically not devoiced in Japanese, and thus Japanese listeners have relatively little experience recovering them.

Conclusion
Based on the results discussed above, perhaps the terms perceptual epenthesis and illusory vowel epenthesis should be not used interchangeably.The default epenthetic vowel is /u/ in Japanese not simply because it is the shortest, but because it is the most common high vowel that Japanese listeners have been trained to recover all their lives.Phonotactic repair in Japanese listeners, therefore, is more akin to perceptual repair, where they use phonotactic and phonetic processes to choose the most probable vowel.In contrast, Brazilian Portuguese lacks a similar systematic devoicing process, and thus phonotactic repairs by Brazilian Portuguese listeners as reported by Dupoux et al. (2011) might be more illusory in nature, triggered primarily by phonotactic violations.

Figure 4 :
Figure 4: Successful vowel identification in VC 1 VC 2 V tokens with full medial vowel.

Figure 6 :
Figure 6: Successful identification rate of target vowel for spliced VCVCV tokens.

Table 2 :
Observed/Expected (O/E) ratio of C 1 V from CSJ. Highest O/E in bold.

Table 3 :
Overall entropy for non-devoicing C 1 and surprisal of all vowels.

Table 4 :
Overall entropy for devoicing C 1 and surprisal of /i, u/.

Table 5 :
Responses for naturally produced VC 1 C 2 V tokens.Most frequent responses in bold.

Table 6 :
Mixed logit model results comparing successful vowel identification rates across difference predictability contexts.Reference context level = HiPredict.

Table 7 :
Mixed logit model results comparing successful vowel identification rates across difference predictability contexts.Reference context level = NoDevoice.

Table 8 :
Mixed logit model results comparing vowel detection between VCCV and spliced VCVCV tokens.firstmodel and just the coefficients of the missing level as obtained from the second model added in italics.

Table 9 :
Responses for VC 1 (u)C 2 V tokens with medial vowel spliced out.Most frequent response in bold.

Table 10 :
Mixed logit model results comparing <u> responses between VCCV and spliced VC(u) CV tokens.

Table 11 :
Responses for VC 1 (i)C 2 V tokens with medial vowel spliced out.

Table 12 :
Mixed logit model results comparing <i> responses between VCCV and spliced VC(i)CV tokens.

Table 13 :
Responses for VC 1 (a)C 2 V tokens with medial vowel spliced out.

Table 15 :
Responses for VC 1 ˺C 2 V tokens with medial vowel and C 1 burst/frication noise spliced out.
Mixed logit model result for vowel detection in spliced vowel-less and burst-less stop tokens.fricatives, the 10+% of vowel detection for the LoPredict stops [p˺, k˺] is still somewhat puzzling.Without a vowel and without a burst between C 1 and C 2 , a token such as [ep˺ko] contains a doubly long stop closure, much like a geminate medial consonant as in[ekko].Geminate consonants are phonotactically legal in Japanese and require no repair.

Table 17 :
Mixed logit model result for vowel detection in splice-4 fricative tokens.