Phonology modulates the illusory vowels in perceptual illusions: Evidence from Mandarin and English

Native speakers perceive illusory vowels when presented with sound sequences that do not respect the phonotactic constraints of their language (Dupoux, Kakehi, Hirose, Pallier, & Mehler, 1999; Kabak & Idsardi, 2007). There is, however, less work on the quality of the illusory vowel. Recently, it has been claimed that the quality of the illusory vowel is also modulated by the phonology of the language, and that the phenomenon of illusory vowels can be understood as a result of the listener reverse inferring the best parse of the underlying representation given their native language phonology and the acoustics of the input stream (Durvasula & Kahng, 2015). The view predicts that listeners are likely to hear different illusory vowels in different phonological contexts. In support of this prediction, we show through two perceptual experiments that Mandarin Chinese speakers (but not American English speakers) perceive different illusory vowels in different phonotactic contexts. Specifically, when presented with phonotactically illegal alveopalatal coda consonants, Mandarin speakers perceived an illusory /i/, but in illegal alveolar stop coda contexts, they perceived a /ə/.


Introduction
The phenomenon of perceptual illusions, particularly that of illusory vowels, has become a very useful probe to understand both loanword patterns and the speech perception mechanism. In this article, we will argue, in line with some recent work, that we can use such perceptual illusions to get a better understanding of not just if, but also how, phonological knowledge is utilized during speech perception. More specifically, we suggest that, during speech perception, the listener attempts to identify the best estimate of the intended underlying representations of the utterance given their phonological/phonetic knowledge and the acoustics of the utterance (Durvasula & Kahng, 2015;Gaskell & Marslen-Wilson, 1998;Gow, 2003;Mitterer, Kim, & Cho, 2013, amongst others); this makes specific predictions about the quality of the illusory vowels; we test and corroborate one such consequence-that there can be different illusory vowels in different illicit phonotactic contexts.
Previous research has unearthed a variety of factors that affect where illusory vowels are perceived. They are perceived in stimuli that contain: (a) consonantal sequences that are phonotactically illicit in the native language of the speaker (Dupoux et al., 1999;Durvasula & Kahng, 2016;Kabak & Idsardi, 2007), (b) consonantal sequences that violate putative language universals such as the Sonority Sequencing Principle (Berent, Lennertz, Jun, Moreno, & Smolensky, 2008;Berent, Lennertz, Smolensky, & Vaknin-Nusbaum, next to coronal consonants (e.g., [t], [d]…) would be the vowel [o]. However, they found that though Japanese speakers seem to perceive more illusory [ɯ] vowels than English speakers in non-coronal contexts, e.g., [egma], the Japanese speakers did not seem to show any increased levels of illusory [o] in coronal contexts, e.g., [etma]. They therefore were not able to get the second illusory vowel that they expected for Japanese speakers.
Fourth, Mattingley et al. (2015) also probed the same issue in Japanese speakers. As Monahan et al. (2009) did before them, they expected different illusory vowels in different consonantal contexts based on the loanword patterns in Japanese, but they explored a larger set of phonotactic contexts than Monahan et al. (2009). Interestingly, though their Japanese participants perceived different vowels in different illicit consonantal sequences, the illusory vowels that they perceived were not always isomorphic with those expected based on the loanword patterns. Specifically, they found that in non-palatal contexts, [ɯ] was by far the most common illusory vowel perceived, and in palatal contexts, [i] was the illusory vowel perceived. One thing to note though is the fact that they did not include a control group; therefore, it is not immediately clear if the Japanese speakers were really hearing illusory vowels, or if there were confounds in the stimuli that encouraged these patterns of perception. This is possible given that the speaker who produced the relevant stimuli was a native speaker of Japanese, who was also fluent in English. In this case, the speaker produced sequences that are illicit in their native language, and might have therefore articulated them with very short excrescent vowels, which could have given rise to the observed patterns in the experiment.
Finally, Durvasula and Kahng (2015) tested Korean speakers' perception of illicit consonantal sequences with English speakers as controls. They showed that, unlike the English speakers presented with the same stimuli, the Korean speakers were perceiving different illusory vowels in different consonantal contexts, and different illusory vowels in some cases for the same context too. More specifically, they showed that in alveolar contexts, Korean speakers perceived the illusory vowel [ɨ]; adjacent to the palatal fricative [ɕ], they perceived the illusory vowel [i]; and adjacent to palatal stops [c], they perceived both [ɨ] and [i] (on different trials of the same stimuli).
It is also important to note that while Guevara-Rukoz et al. (2017), Monahan et al. (2009), andMattingley et al. (2015) were basing their prior experimental expectations on loanword patterns observed in Japanese, Dupoux et al. (2011) and Durvasula and Kahng (2015) were basing their expectations on predictions made by their particular views of speech perception. As pointed out by many authors (de Jong & Cho, 2012;LaCharité & Paradis, 2000;Peperkamp, Vendelin, & Nakamura, 2008;Smith, 2006;Vendelin & Peperkamp, 2006, amongst others), there can be multiple sources through which loanwords can enter a language, with normal auditory input as just one possible source. Therefore, to use loanwords to understand speech perception, one needs to first factor out the loanword patterns that have entered the language through other channels, and then use the remaining corpus to identify the relevant patterns. However, such historical sources are often not available, and thus, using loanwords to make predictions about illusory vowels is not straightforward. In contrast, the predictions by Dupoux et al. (2011) and Durvasula and Kahng (2015) stem from proposals about the speech perception mechanism, and are therefore more relevant to our understanding of the phenomenon of illusory vowels. It is for this reason that, in what follows, we first lay out a certain view of speech perception and then test out the consequences of the view for illusory vowel patterns.
The phenomenon of illusory vowels generally and the quality of the illusory vowel itself have also received an Optimality-Theoretic analysis by Boersma and Hamann (2009). While they were primarily interested in understanding loanword patterns through the lens of first-language speech perception, they suggest, in a discussion of the illusory vowel phenomenon in Korean speakers, that the illusory vowel is likely to be /ɨ/, because it is involved in optional deletion processes. A similar observation relating particular language-specific processes in Japanese (namely, vowel shortening and devoicing) to the illusory vowel /ɯ/ was made as far back as Dupoux et al. (1999). This is an issue that we will return to below, when we present our own viewpoint and predictions.
Before laying out the expectations for the experiments discussed in this article, it is important to present our conception of the nature of the problem that is being solved during speech perception (following Marr, 1982). We assume that the task of the listener in speech perception is primarily a task of reverse inference-it is to identify the best estimate of the intended underlying (or phonemic) representations of the utterance given the acoustic token (Gaskell & Marslen-Wilson, 1996, 1998Gow, 2003;Mitterer et al., 2013, amongst others). Previously, the viewpoint has largely been employed in understanding perceptual compensation for assimilatory changes, particularly, at the edges of words. For example, the phrase 'garden bench' /gɑːdn bɛntʃ/ is often pronounced as [ɡɑːdm bɛntʃ], 4 where the word-final nasal /n/ has assimilated to the place of articulation of the following segment. It has been shown that listeners are able to compensate for such coarticulatory changes, i.e., when presented with an assimilated variant (e.g., [ɡɑːdm]), listeners are able to recognize the unassimilated word (e.g., 'garden'), but only when the nasal consonant is followed by a word that begins with a bilabial sound (e.g., [ɡɑːdm bɛntʃ]). Durvasula and Kahng (2015) extend such views to understand the phenomenon of illusory vowels.
This view parallels Bayesian models of speech perception (Bever & Poeppel, 2010;Feldman & Griffiths, 2007;Poeppel & Monahan, 2011;Sonderegger & Yu, 2010;Wilson & Davidson, 2013), which typically involve perception as reverse inference from acoustic input to surface representations. As a consequence of this viewpoint, knowledge about both phonological alternations and phonotactic constraints is required to reverse-infer the phonemic/underlying representations from the acoustic tokens. More specifically, in regard to the quality of the illusory vowel, the perceiver's task is an attempt to repair the illicit phonotactic sequence with a (vowel) phoneme that best maps to the phonetic characteristics of the acoustic token. When no relevant phonological alternations bias listeners towards a certain vowel in the particular segmental context, the best vowel guess that repairs the particular phonotactic violation could either be based on co-articulatory cues present in the relevant content or be the phonetically minimal/shortest vowel in the inventory, à la Dupoux et al. (2011). However, when relevant phonological alternations do bias listeners towards particular vowel percepts in specific segmental context, the best guess depends on both the phonetics of the acoustic token and the phonological alternations themselves. The types of phonological processes that are likely to play a role are those that bias the listener's expectations about the quality of the illusory vowel. One such process is a consistent/regular vowel deletion process that targets a particular vowel. The presence of a regular process of vowel deletion that targets a particular vowel (/V 1 / → [∅]) in the phonology of the language supports the reverse inference of the same vowel in the phonemic representation when the surface representation has nothing (reverse inference: [∅] → /V 1 /). For these reasons, in a phonotactically illicit consonantal context, where the condition can be perceptually repaired by a vowel, the best vowel to repair the situation is the phoneme /V 1 / that maps to [∅] in the surface/acoustic representations. As mentioned above, similar or related views, linking language-specific deletion/reduction processes to the quality of the illusory vowel, were presented by Boersma and Hamann (2009);Dupoux et al. (1999).
A second type of process that is likely to bias a listener's expectations about the vowel quality of the illusory vowel is one that involves a phonotactic restriction ([C 1 ] is allowed only next to [V 2 ]). 5 If the listener is auditorily presented with a consonant [C 1 ] in a context where it is not phonotactically licit, and if the sequence can be repaired phonotactically by a vowel, then the best vowel to perceptually repair the sequence is the vowel /V 2 /, as this would also account for the acoustic properties of the illicit consonant; this is so because the consonant [C 1 ] can only appear before a vowel [V 2 ] in the language.

Relevant phonological patterns and predictions
In what follows, we briefly describe some phonological patterns in Mandarin and English that are relevant for the phonological contexts tested in this paper. We then present specific predictions that stem from the relevant language-specific facts and the viewpoint of speech perception presented above.
In Mandarin, there are no obstruent coda consonants. Therefore, consonantal sequences where the first consonant is an obstruent consonant are not possible [*atɕʰma, *atʰma]. This constraint is true both within words (so across syllables) and across higher phonological domains: i.e., a Mandarin speaker is not likely to hear such sequences at all, in a random speech stream. Following the prior work on the topic of illusory vowels, word-medial obstruent-nasal clusters are thus contexts that are ripe for illusory vowel perception in Mandarin Chinese.
Some of the phonological patterns in Mandarin exhibit exactly the characteristics needed to bias the perception of the illusory vowels in the relevant contexts. Weinberger (1996) discussed one such phonological process of vowel deletion that targets the vowel /ə/ in certain fast conversational-speech contexts: An unstressed word-final /ə/ is deleted when preceded by a nasal consonant (1). 6 This allows /ə/ to be a good vowel for perceptual repairs in most contexts because it already varies with ∅ (nothing) in the surface phonetic representations. Furthermore, /ə/ is often toneless (and stressless) in Mandarin (Weinberger, 1996). Therefore, inferring such a vowel phoneme allows for the listener to straightforwardly account for the lack of tone (or stress) between two consonants that form an illicit sequence. For these reasons, /ə/ is an excellent candidate for an illusory vowel-we call it illusory vowel 1. 7 This can also be thought of as the default illusory vowel in Mandarin, as this is the illusory vowel expected in contexts where there are no other influencing phonological processes.
1. Fast-speech vowel elision in Mandarin Chinese (Weinberger, 1996) • A second pattern in Mandarin that can bias the perception of illusory vowels has to do with consonant-vowel phonotactics. While alveolar stops as a group can precede all vowels [✓tʰi, ✓tʰa, ✓tʰu, …], alveo-palatal consonants can only appear before high front vowels or glides [✓tɕʰi, *tɕʰa, *tɕʰu, …] (Note: /tʰ/ specifically seems to be absent before /e/ based on the corpus results presented in Tables 4-5). These facts allow front vowels, particularly /i/, to be a good vowel for perceptual repairs in illicit alveo-palatal coda perceptual illusions Art. 7, page 6 of 27 contexts (but not in illicit alveolar coda contexts). We call the high front vowel /i/ illusory vowel 2. Now, we turn to the relevant phonological and phonetic facts in English. In English, there is no alveo-palatal consonant such as [tɕʰ], but there are palato-alveolar consonants such as [tʃʰ]. 8 So, when an English speaker is presented with [tɕʰ], they are likely to perceive it as /tʃʰ/-this is expected based on the results of non-native perception discussed by Best, Hallé, Bohn, and Faber (2003). Furthermore, both alveolar and palatoalveolar consonants are possible in coda positions in English. Therefore, both [tʃʰm] and [tʰm] are phonotactically valid sequences. One might object that /tʰ/ in coda positions typically glottalize, but they do appear variably as [tʰ] (Kreidler, 1989;Vaux, 2002), particularly across compound members and across words in careful or enunciated speech in our own experience. So, [tʰm] is a possible sequence/pronunciation in English. As a consequence, sequences such as [atɕʰma], [atʰma] have perfectly licit phonemic parses in the language. Though not pursued here, the viewpoint presented in this article predicts that /ə/ should be the main illusory vowel for English speakers in illicit phonotactic contexts, as it regularly deletes in fast/casual-speech processes (Hooper, 1978).
In light of the above phonological facts from Mandarin and English, we present the predictions that the viewpoint presented earlier makes for particular illicit sequences (Figure 1). We expect that when Mandarin speakers are presented with an illicit consonantal sequence involving an alveo-palatal affricate (e.g., [tɕʰm]), the illusory vowel perceived by the listener should be biased by the phonology; as in this context, only high front vowels are possible. While nothing directly stemming from the viewpoint presented above motivates a more specific prediction than a high front vowel, one could argue based on markedness considerations that that high front vowel must be /i/, illusory vowel 2 (Note: Apart from /i/, there is a second high front vowel in Mandarin, namely, the rounded counterpart /y/. It is possible that /i/ is preferred for acoustic reasons; if so, that is consistent with the reverse inference view that we suggest here. However, without precise acoustic recordings and models, it is difficult to sustain such a claim currently, so we haven't explicitly argued for it.). Furthermore, when Mandarin speakers are presented with an illicit consonantal sequence involving an alveolar stop (e.g., [tʰm]), no particular phonological pattern in Mandarin has a direct influence on the identification of the illusory vowel. In such a context, they are expected to perceive the more general illusory vowel 1, which happens to be /ə/.
In contrast, when an English speaker is presented with either of the above sequences, i.e., [tɕʰm] or [tʰm], no illusory vowel is expected due to the phonology of the language, as both sequences are possible (albeit, across higher prosodic domains than just a syllable boundary). However, the presence of confounding acoustic factors in the stimuli, such as extremely strong burst release or aspiration, might induce some illusory vowels even for English speakers (along with Mandarin speakers). There are two important caveats to bear in mind about the above predictions. First, the prediction that Mandarin speakers should hear an illusory /i/ in alveo-palatal contexts, but a /ə/ in alveolar contexts is a prediction that holds true if only phonological knowledge were deployed (essentially, it is a mutatis mutandis prediction). However, all naturalistic stimuli are likely to carry subtle artifacts. For example, it is possible that the Mandarin speakers will have some /ə/ illusions in illicit alveo-palatal contexts for reasons related purely to the acoustic properties of the particular stimuli in the experiment. Therefore, the above prediction can also be seen as saying that, for Mandarin speakers, the /i/ illusions are going to be higher in illicit alveo-palatal contexts than illicit alveolar contexts; or conversely, the /ə/ illusions are going to be higher in illicit alveolar contexts than illicit alveo-palatal contexts. 9 Second, the possibility that artifactual acoustic effects in the stimuli could trigger illusory vowels even for English speakers entails that the appropriate predictions are not within-language predictions, but between-language predictions. Such predictions would control for any effects due to acoustic artifacts in the stimuli. Therefore, the appropriate predictions for us are that Mandarin speakers, compared to English speakers, should perceive more /i/ in [tɕʰm]  In the following sections, we test these predictions out using two perceptual experiments, an ABX task (Experiment 1), and an identification task (Experiment 2).

Participants
Twenty native Mandarin speakers from Beijing (mean age = 20.5 years, 7 men and 13 women) and 19 native English speakers from Michigan (mean age = 21 years, 6 men and 13 women) participated in the experiment. All the participants were recruited at Michigan State University. The Mandarin participants spent an average of 2.4 years (SD = 1.6 years) 10 in the US before the experiment, primarily as undergraduate students at Michigan State University. Furthermore, the Mandarin speakers also reported an average of 14.1 years (SD = 2.6 years) of exposure to English as a second language in a classroom environment in China. The Mandarin participants received no compensation for their participation, but the English participants received extra credit for their participation.

Materials
All the test items followed the template aC₁V₁ma, in which C₁ was an alveolar or alveopalatal consonant [tʰ/tɕʰ]; and V₁ was [i / ə / ∅ (Null)]. None of the stimuli were words in either Mandarin or in English. All the stimuli had stress on the first vowel with a highhigh-low tone sequence on the 3-syllable nonce words and a high-low tone sequence on the 2-syllable nonce words. They were natural recordings by a trained male phonetician (the first author), who is a native speaker of Indian English and Telugu, and a secondlanguage speaker of standard Hindi. The speaker was chosen to avoid providing either linguistic group with stimuli from their native dialect/language. There were two reasons for the use of this particular speaker. Firstly, he could naturally produce all the stimuli, as they are phonotactically licit in his dialects of both Hindi and Telugu (particularly, across words). The use of a native Mandarin speaker to record the stimuli would have only been possible if the speaker had neutralized their own linguistic biases, as many of the sequences are not licit in the language. We strongly suspect that the use of Mandarin speakers to record stimuli would have introduced biases into the stimuli (in the form of very short excrescent vowels), especially for those sequences that are not licit in the relevant language, thereby making the interpretation of the results much more challenging. Secondly, the use of an American English speaker to record the stimuli was also avoided, because those that we tried had difficulty in producing unstressed medial vowels that were unreduced (i.e., they could not block the vowel-reduction process in their dialect), and it would be a challenge for native American English speakers to produce Mandarin alveo-palatals. Furthermore, we did not want to introduce a bias that would help the control group, as the overall phonetic patterns would have been more natural for the American English speakers than for the Mandarin speakers. The interpretation of the results would therefore have been confounded by this. For these reasons, we used the first author's voice for recording stimuli. (Note: If our objective were trying to understand American English loanwords in Mandarin, then using an American English speaker would have been necessary. However, this is not our focus. Instead, we are interested more generally in speech perception, and particularly, in how Mandarin speakers, but not American English speakers, are expected to hear more illusory vowels under different acoustic circumstances. In this latter case, the use of an American English speaker to record the stimuli is not obvious and has to be considered in the context of what confounds are raised. It is also not our intention to probe any language-neutral perception. However, the use of an American English speaker would have asymmetrically made the task easier for the English participants, as the acoustics in the non-crucial portion of the tokens would have been easier to interpret, thereby allowing them to potentially focus more on the crucial environment. Such a confound makes it difficult to interpret any differences between the English and Mandarin speakers' perceptual responses).
There are three more issues worth discussing in detail with respect to the stimuli and the speaker. 11 First, the alveo-palatal pronuciations by the speaker were reasonably close to native Mandarin pronunciations. One of the speaker's native languages, Telugu, has sounds which can be reasonably close, acoustically, to alveo-palatal stops. However, we depended on three of the co-authors (who are native Mandarin speakers), and four more native Mandarin speaking members of the phonology-phonetics lab in our department to first vet all the stimuli for naturalness, particularly with respect to the crucial consonants' place and manner of articulations. And, only after the relevant members and co-authors were satisfied with the stimuli did we proceed with experimentation. Second, a similar issue arises with the naturalness of the tones, particularly that of the low tone used. Here, it is worth noting that some phonological analyses consider T3 in Mandarin, traditionally transcribed as a low-dipping-rising tone, to be underlyingly low (Duanmu, 1999(Duanmu, , 2007Yip, 1980). In fact, Duanmu (1999, p. 14) suggests that this tone largely has a low pitch contour, and is best described as either 211 or 11 in the Chao system (where, 5 is the highest pitch, and 1 is the lowest pitch). Furthermore, T3 is consistently pronounced as a low tone before another tone (Chao, 1968;Duanmu, 2007;Lin, 2007). Finally, in many Mandarin speakers' speech, the final rise of T3 is absent even in final position, hence there is just a low tone in final position (Duanmu, 2007;Lin, 2007). Therefore, both the low tone and high tone used in this experiment could very well be reasonably natural for Mandarin speakers. Having said the above, our original intention was not to use Mandarin tones, but to use tones consistent with both previous/future experiments in our lab. It is possible that the use of non-Mandarin tones might have resulted in confounds for the Mandarin speakers; however, for reasons discussed towards the end of the Introduction, the important comparisons to control for acoustic artifacts in the stimuli are betweenlanguage comparisons. As a consequence, using perfectly phonetically matched Mandarin tones that have no correspondents in American English, might well have introduced confounds into the control group's (English speakers') responses, given that pitch height difference and contours do play a role in the English stress/intonational system. To us, there is no immediately obvious way of solving this problem, and therefore the best way to proceed is to use tones in a consistent fashion, and look for tonal interactions in future experiments that are focussed on such effects; we particularly chose high-low and highhigh-low tone sequences in order to mirror a natural declination in pitch. Here again, we made sure that the stimuli sounded reasonably natural to the Mandarin-speaking co-authors and other Mandarin (and English speaking) members of our lab before we proceeded to running experiments. Third, it is also reasonable to ask how natural the medial vowel in the stimuli were for two language groups. Again, we initially depended on impressions of the group of Mandarin linguists we consulted (mentioned above) to establish the naturalness of the vowel quality. Anticipating the results of Experiment 2 here, we note that the results presented in Figure 5 suggest that both the English and Mandarin speakers were at reasonably high levels (and not much different from each other) at identifying the medial vowels.
Each item was recorded several times using the software Praat (Boersma & Weenink, 2016) with a microphone (Logitech USB Desktop Microphone; Frequency Response -100 Hz-16 KHz) at a 44 KHz sampling rate (16-bit resolution; 1-channel). The stimuli were normalized in Praat to have a mean intensity of 70 dB SPL. From these recordings, two tokens were selected for each item and they were each presented twice; therefore, there were 12 tokens in the experiment.  ([atɕʰma, atɕʰəma, atɕʰima]). This amounted to a total of 96 trials in the experiment, presented in pseudo-randomized order with the added constraint that there be no identical triplets in succession. (Note: We included the pairs [atʰima~atʰəma] and [atɕʰima~atɕʰəma] for experimental reasons to ensure a reasonable number of clearly different stimuli, and as a sanity check. With respect to the latter concern, if the Mandarin speakers confused the vowel pairs at much higher rates than the English speakers, then it would be difficult to see if the Mandarin speakers really heard a particular vowel in a certain context. As will be obvious later in Figure 2, the Mandarin speakers were at least as good as the English speakers with these pairs).
The experiment was conducted in a quiet room with a group of 4-6 participants per session. The stimuli were presented with a low-noise headset (Koss R80 headphones) to each participant through an ABX task scripted in Praat (Boersma & Weenink, 2016). The participants were asked to listen to the triplets of stimuli and determine whether the last sound was more similar to the first or the second and click on the corresponding box (1 or 2) on the screen with a mouse. All the instructions were in English for the English speakers ("Is the last sound more similar to the first or the second?") and in Mandarin for the Mandarin speakers ("第三段音和第一段音还是第二段音更相似?"). The experiment started with a practice session to ensure familiarity with the task. The practice session had 12 trials with another set of nonce words (where the C₁ was [m]). The inter-stimulus interval was 500 ms and the inter-trial interval was 1500 ms. All 96 trials were randomized for each participant. The experiment took about 7-8 minutes.

Results
A visual inspection of the mean percentage of correct responses to the stimuli by both the Mandarin and English speakers (Figure 2)  In order to confirm the observations made by visual inspection of the results, we followed up with statistical analysis in R (R Development Core Team, 2014). A three-way mixed ANOVA was run, using the ez package (Lawrence, 2015), to model the percentage of correct responses as the dependent variable with Language (English, Mandarin) as a between-subjects factor and two within-subjects factors, Consonant (tɕʰ, tʰ) and Vowelpair (i~nothing, ə~nothing It is important to recognize that the between-language comparisons are indeed the crucial comparisons. For example, it is possible to observe Figure 2 and infer that there is a tendency for Mandarin speakers to confuse [atɕʰəma~atɕʰma] more than [atɕʰima~atɕʰma]; 14 however, to make sure that this difference doesn't arise from stimulus artifacts, it is important to compare the results directly with the English speakers. Looking at the English speakers' responses in Figure 2 suggests they have a similar increase in the degree of confusion for [atɕʰəma~atɕʰma] compared to [atɕʰima~atɕʰma], though they were not expected to show any such difference. This suggests the increased Mandarin confusion for [atɕʰəma~atɕʰma] over [atɕʰima~atɕʰma] should not be interpreted as how the Mandarin speakers' phonological knowledge affects perception. More generally, we want to reiterate that, in order to probe truly phonological knowledge in speech perception, one needs to guard against stimulus artifacts, which practically all naturallyrecorded stimuli are bound to have, by conducting between-language comparisons. Within-language comparisons, in our opinion, cannot guard against such issues. 13 We excluded the comparison pairs involving the two vowels based on the recommendation of a reviewer. Originally, we ran a two-way mixed ANOVA, where Language and Comparisons (all 6 comparison pairs) were two separate factors. The analysis crucially revealed an interaction of Language*Comparison [F(5,185) = 15.36, p < 0.0001], which again suggests that the Mandarin and English speakers had different responses to different comparison pairs. 14 Thanks to a reviewer for making this observation in our results.

Discussion
In Experiment 1, Mandarin speakers confused [atɕʰima~atɕʰma] and [atʰəma~atʰma] more than English speakers. This is in line with our expectation that Mandarin speakers, but not English speakers, are biased towards some illusory vowels due to the phonology of their native language. There were also a few unexpected results. First, Mandarin speakers also confused [atɕʰəma~atɕʰma] and [atʰima~atʰma] more than English speakers. In understanding these unexpected results, it is important to note that the ABX task is not a direct task looking at illusory vowels in the crucial test stimuli, but is instead fundamentally comparative. For example, it is possible that for Mandarin speakers [atɕʰəma] might have been perceived as /atɕʰima/ due to the phonotactic patterns of Mandarin; [atɕʰəma] was therefore confusable with [atɕʰma], which was also perceived as /atɕʰima/ due to an illusory vowel. Such a possibility, while not initially considered by us, is consistent with the overall focus of the explanation residing in the listener's knowledge of Mandarin phonology. The [atʰima~atʰma] difference on the other hand is not amenable to such an explanation. Here, it is worth highlighting that the Mandarin speakers' correct responses are not as dramatically different from the English speakers. Furthermore, the percentage of correct responses is quite high for Mandarin speakers compared to other stimulus pairs including the stimulus pairs with different vowels, so it is possible that the Mandarin responses have hit some sort of ceiling effect related to the non-nativeness of the stimuli. However, we acknowledge that this is not entirely clear. Crucially, Experiment 2, to be discussed below, confirms that Mandarin listeners do not hear an /i/ in [atʰma]. Therefore, Mandarin listeners are confusing [atʰima~atʰma] because of some other perceptual reason than perceiving an /i/ in [atʰma]. It is important to note that our prediction of Mandarin listeners not confusing [atʰima~atʰma] was based on the prediction that, if no other perceptual changes are there, then [atʰima~atʰma] will not be confused because Mandarin listeners do not infer an /i/ in [atʰma]. Experiment 2's results confirm this crucial aspect of the prediction.
Quite unrelated to our immediate question, there was also a tendency for the English speakers to confuse the [atʰəma~atʰima] more than the Mandarin speakers. This again might be due to phonological reasons; in this case, reasons related to English phonology. The phenomenon of unstressed vowel neutralization that has been extensively observed in English (Burzio, 2007;Chomsky & Halle, 1968) might have negatively influenced the English speakers' performance. As with the first unexpected result, it is worth pointing out that such an explanation is consistent with our broad view of the role of phonological knowledge, and deserves further scrutiny of its own in future work.
The main issue raised by the unexpected results in Experiment 1 is that ABX tasks are fundamentally indirect (or comparative), and perceptual illusions not only in the expected token, but also in the comparison tokens, might lead to incorrect responses. It is therefore important to attempt an experiment that focusses directly on the illusory vowels in the crucial test items, i.e., the items with the illicit consonantal sequences [atʰma] and [atɕʰma]. Experiment 2 addresses this need with a more direct identification task. Experiment 2 addresses this need with a more direct identification task.

Participants
The participants in Experiment 2 were the same as in Experiment 1. Experiment 2 was an identification task that drew participants' attention to the medial vowel in the stimuli. Therefore, it was conducted after Experiment 1 (after a short break) so as not to have the participants focus on only the vowel in Experiment 1.

Materials
The stimuli were the same 6 test items used in Experiment 1 as described in Section 2.1.2. There were two recordings used for each item, as in Experiment 1, and they were each presented thrice; therefore, there were 6 tokens of each test item, and a total of 36 tokens in the experiment, presented in pseudo-randomized order with the added constraint that there be no identical test items in succession.
An issue that is important to acknowledge here is that of experimental power, in both Experiments 1 and 2, related to the number of repetitions of test items (or pairs, in Experiment 1). 15 The issue of power is always a concern in any experiment, especially when null results are involved, and it is possible that the power in the current experiments might have increased if we had increased the repetitions. But, when we planned the experiments, given that the participants would take part in both the identification task and the ABX task, we were concerned about learning effects of many presentations, whereby participants might adopt complex response strategies in response to the stimuli in the experiment. To us, this is still a largely underexplored topic, and we wanted to err on the side of caution with respect to learning effects. Furthermore, given that the responses to the repetitions were very likely not independent of each other, it is not necessary that adding more repetitions would have increased the power of the experiments -if anything, a much more substantial increase in power in such experiments is likely to come from increasing the number of participants, we think.

Procedure
In Experiment 2, we used an identification task to investigate a perceptual epenthesis effect. The experiment was conducted in a quiet room with a group of 4-6 participants per session. The stimuli were presented with a low-noise headset (Koss R80 headphones) to each participant through an identification task scripted in Praat. The participants were asked to listen to a stimulus and determine whether the medial vowel was [i], another vowel, or nothing and click on the corresponding box on the screen with a mouse ( . We chose to provide the participants with the option [other vowel] because both in English and Mandarin, there is no unique letter to identify a schwa. The letter "e" could have been used, but it is still minimally ambiguous between [ə] and [e]. Furthermore, though not immediately relevant to us, the Pinyin character "i" can also stand for the [ɹ̩ ] after some fricatives/affricates. All the instructions were in English for the English speakers ("What's between the two consonants?") and in Mandarin for the Mandarin speakers ("两个声母中间有什么?"). Before the actual experiment, each participant completed a practice session to ensure familiarity with the task. The practice session had 12 trials with another set of nonce words, where C₁ was [m/r]. The inter-trial interval was 1500ms. All 36 trials were randomized for each participant.

Results
A visual inspection of the mean percentage of responses to the stimuli by both the Mandarin and English speakers suggests that all and only the expected differences were found between the two language groups (Figure 3)  gave them for [atɕʰma]). We also tried to probe if there was any common element to the 4 Mandarin speakers who didn't give any vowel responses, and didn't find them to be exceptional in either the amount of time spent in the US, or in the duration of exposure to English as a second language in a classroom environment in China. Finally, we tried to see if the [no vowel] responses by the Mandarin speakers were related to the duration of their English exposure in China or the time they spent in the US. Figure 4 shows there is no observable trend. As mentioned in the Introduction, the crucial planned comparisons are the ones that compare the responses by the two language groups to each type of stimulus without a medial vowel, as such a between-language comparison controls for any subtle artifacts in the stimuli. Therefore, we conducted pairwise Mann Whitney U tests (   We also include further pairwise analyses looking at a different question: Is it the case that Mandarin speakers perceive more /i/ than other vowels in the alveo-palatal context, and more other vowels in the alveolar context, while there are no such differences for English speakers? 17 However, it is important to note that this question, which we include for the sake of completeness, requires within-language comparisons; therefore, unlike the results in Table 2, this way of asking the question does not control for stimulus artifacts directly. The within-language comparisons ( Finally, just as a final sanity check, we also looked at the participants' responses to the stimuli with medial vowels. 18 The results are shown below in Figure 5. We conducted statistical analyses, paralleling the crucial between-language comparisons for the test items without medial vowels. We ran a three-way mixed ANOVA where the dependent variable was the percentage of responses, and the independent variables were Language (English, Mandarin) as a between-subjects factor; and Stimulus ( Most importantly, there was no three-way interaction effect; this suggests that the Mandarin speakers' responses did not differ substantially from the English speakers differentially depending on the stimuli. As with the crucial comparisons above, we followed up by 17 Thanks to a reviewer for suggesting this. 18 Thanks to a reviewer who requested we include the analysis in the article. The results of the analysis on the stimuli with medial vowels suggest that the Mandarin speakers' differing responses to the crucial alveolar and alveopalatal stimuli without medial vowels (as observed in Figure 3) were not due to differing amounts of confusion with the relevant vowels in the same consonantal contexts.

Discussion
While the results of Experiment 1 were in line with the predictions, the experiment itself was not a direct test of illusory vowels in the crucial test items, as ABX tasks are inherently comparative. This issue was addressed in Experiment 2, where an identification task was employed. The results of Experiment 2 suggest that: (a) [atɕʰma] triggered only /i/ illusions and no other vowel, (b) [atʰma] did not trigger an /i/ illusions; instead the Mandarin participants heard another vowel, presumably /ə/. The responses of the English participants suggests that the Mandarin speakers' performance was not due to acoustic artifacts in the recordings.
An interesting issue that arises from the results of both Experiment 1 and Experiment 2 is related to the somewhat small differences in accuracy rates between Mandarin and English speakers even for the crucial pairs (Experiment 1), and the somewhat low illusory vowel responses by Mandarin speakers in the crucial stimuli without medial vowels (Experiment 2). 19 For example, in Experiment 2, the [other vowel] illusion rate (which, note, is the measure of illusory /ə/) in the [atʰma] context for the Mandarin speakers was about 36%. It is important to note that there is quite a large variation in illusory vowel responses in the previous literature; compare the above results to those of similar experiments in previous published work: 20 (a) There was a 65-70% identification 19 Thanks to the Associate Editor and a reviewer for raising this issue. 20 Note, the list is representative, not comprehensive. of illusory /ɯ/ responses by Japanese speakers (Dupoux et al., 1999), (b) there was a roughly 58% identification of /ɯ/ for Japanese speakers, and roughly 65% identification of /i/ by Brazilian Portuguese speakers (Dupoux et al., 2011), (c) there was a varying number of illusory /ɨ/ perception for Korean speakers based on the consonantal sequence involved (range: 25%-50%), (d) finally, Davidson and Shaw (2012), in a speeded AX task, report different levels of accuracy (proportions of hits) for English speakers for stimulus pairs (that differed in presence or absence of medial vowels) for different types of consonantal sequences (range: 0.51-0.88). While the differences in illusory vowel rates (and accuracy rates) between experiments appear quite sizeable, we don't yet have a complete understanding of what causes such differences. There are of course many possible reasons for the differences. First, as has been very well documented, illusory vowels are not the only way to 'repair' incoming phonotactically illicit sequences; for example, there can be consonant deletions, consonant changes too (Davidson, 2007;Davidson & Shaw, 2012;Hallé, Segui, Frauenfelder, & Meunier, 1998). Note, such repairs are also generally consistent with the reverse inference view presented in this article: Listeners might have inferred that the underlying consonants are different (but, still those that can account for a lot of the acoustic properties of the incoming stimuli), leading to what might be termed a consonant change in the percept; or listeners might have inferred that some of the acoustic characteristics perceived might be 'noise' of some sort, and therefore didn't infer any underlying consonantal counterpart, leading to what might be termed consonant deletion in the percept. Given that there are other perceptual repairs, it is quite unclear generally how and why participants infer different possibilities. Therefore, it is difficult to predict the exact proportion of illusory vowels listeners of a language should hear in relevant contexts. Second, even if the above were disregarded, though there is some understanding of how aspects of the acoustic input proximate to the illicit consonant cluster have an effect on illusory vowels (Davidson & Shaw, 2012;Wilson & Davidson, 2013, amongst others), there are still many other aspects that researchers, to our knowledge, have not yet studied systematically. For example, what is the influence of the durational (and more generally, phonetic) properties of the other vowels in the stimuli and do illusory vowel percepts increase/decrease depending on whether the surrounding vowels are shorter/longer? There is some reason to believe this should happen based on speech-rate effects on the parsing of ambiguous stimuli (Dilley, Morrill, & Banzina, 2013;Dilley & Pitt, 2010;Heffner, Dilley, McAuley, & Pitt, 2013). As Dilley et al., have consistently found, whole syllables can 'vanish' in ambiguous stimuli due to the distal/proximal speech rates. As a very reasonable extension, one can also conjecture that there should be similar speech rate effects on illusory vowel effects, which would modulate the rates of perception of illusory vowels. Another reasonable source of differences in such experiments is the acoustic properties of the stimuli with the medial vowels. Conducting an identification experiment, where such stimuli with medial vowels have longer (or shorter) vowels should affect the inference of a [∅] in the acoustic input as a vowel. 21 Third, many of the previous results on illusory vowels are based on non-tonal languages. But, in Mandarin, there is a general relationship between tone and syllable count, and the stimuli with no medial vowels only had two tones. So it is possible that the presence of just two tones that already corresponded to other acoustically present vowels/syllables in the stimuli negatively affected the Mandarin speakers perception of an illusory vowel in the medial position, thereby increasing their [no vowel] responses.
Finally, one can never discount that listeners can also use the general auditory perception system (not just the speech perception system) in these tasks. This is particularly true given that the stimuli in the experiment, though reasonably natural, still might have sounded non-native to the Mandarin speakers. Therefore, it is possible that the Mandarin speakers were influenced by general auditory processing and therefore more pre-disposed (than if only the speech-perception mechanism had been in play) to recognizing that there were no vowel-like elements in the medial position for the stimuli without medial vowels.
For the above reasons, we have repeated throughout that the clearest test of explanations of illusory vowels, based on a particular view of speech perception that we advocate here, is in the direction of such effects in between-language comparisons; but expectations of the magnitude of the effect are still somewhat premature. However, we raise the above possibilities as both interesting and necessary avenues for further research.
Before concluding this discussion, it is important to mention that an aspect of the illusory vowel phenomenon that Experiment 2 could not clearly isolate was the exact quality of the illusory vowel for Mandarin speakers in [atʰma]. As mentioned in Section 3.1.3, one of the possible responses given to the participants was [other vowel] because both English orthography and Pinyin, which is the official romanization for Standard Chinese, use the same letter "e" appears in graphs for [ə], [e], and [ɤ] (Note: In Pinyin, the "e" in the orthographic sequence "ei" represents [e], but the letter "e" by itself can represent [ə] before a nasal coda or [ɤ] in an open syllable). Therefore, giving the participants the orthographic option of "e" would not have been completely sufficient to identify the exact quality of the vowel. Despite the orthographic confound, the experiment, by itself, does allow us to eliminate some potential counter-hypotheses to account for the patterns. It is not possible to maintain that it is simple phonotactics that determine the vowel. As mentioned in the Introduction, alveolar stops are possible before all vowels in Mandarin except [*tʰe], which means [tʰi] is a phonotactically possible sequence. Since the [i] response was at floor levels, it is not possible to say that the illegal sequence present in [atʰma] is fixed by just any phonotactic sequence allowed in the language. The consistent selection of the [other vowel] response by the Mandarin speakers suggests that only one (or a few phonotactically legal vowels) are possible illusory vowels in that context. Again, this cannot be purely due to the acoustic factors present in the stimuli either, since that would predict the English speakers should also have a similar rate of [other vowel] responses. Instead, the English speakers largely chose [no vowel]. If the Mandarin speakers' responses for [other vowel] were purely due to the presence of acoustic factors, then it would remain unexplained why the English speakers did not use the same cues to respond in a similar fashion.
For a similar set of reasons, it is not possible to maintain a straightforward frequency based account for the illusory vowels, particularly those that are adjacent to the alveolar context. In Tables 4-5, we present the bigram and trigram frequencies for [tʰV] or [tʰVm] [atʰma] items. This suggests that the illusory vowels that the Mandarin speakers were perceiving were not primarily driven by such frequency information.
As an additional note for those interested in further pursuing a frequency-based account, we would also like to mention that even if another frequency measure were found to correlate with the identification rates in these experiments, it would only be meaningful if there was a clear (theoretical) explanation of why that particular frequency measure is the appropriate measure, and not bigram and trigram frequencies.
In short, what is needed to accept a frequency-based explanation is a clear theoretical viewpoint that clearly explains why some types of frequencies, and not others, are relevant to the process.

Conclusion
In this paper, we set out to corroborate a particular view that states that the task of the perceiver during speech perception is to identify the best estimate of the intended underlying representations of the utterance given the acoustic token. A consequence of this viewpoint is that both phonological knowledge (therefore, knowledge of phonological alternations and phonotactic constraints) and the acoustics of the relevant tokens make contributions during perception, i.e., perception cannot be solely based on one or the other. It is important to note that the recruitment of phonological knowledge is a necessary aspect of speech perception, if the perceiver is trying to identify the best estimate of the intended underlying representations. This is because, except in trivial cases where the underlying representations 'match' the surface representations and are directly inferable from the acoustics, there are many cases where the underlying representation is not directly inferable from the acoustics, without knowledge of the phonology, e.g., a language where word-final /t/ becomes a glottal stop-here, inference of the correct intended underlying representation has to use the phonology of the language. The above view makes testable predictions about the quality of the illusory vowels that a native speaker of a particular language might perceive in different illicit phonotactic contexts. Based on Mandarin phonological and phonetic patterns, the view predicts that, compared to native English speakers, native Mandarin speakers should perceive more illusory vowel /i/ in illicit consonantal sequences where the first consonant is alveopalatal, but more illusory /ə/ in illicit consonantal sequences where the first consonant is alveolar. The results of Experiments 1 & 2 were in line with these expectations. We further argued that the illusory vowels perceived cannot simply be due to phonotactics or (bigram/trigram) frequencies of segments.
Finally, we would like to note that, while the research is in line with previous work that shows a general modulation of speech perception by phonological knowledge (Boomershine, Hall, Hume, & Johnson, 2008;Huang, 2001;Hume & Johnson, 2003;Johnson & Babel, 2010), the article presents results that suggest a particular viewpoint on how exactly that modulation takes place that can naturally account for such phonological sensitivity in speech perception.

Implications for loanword phonology
As pointed out in the Introduction, while loanword patterns cannot be assumed to be purely based on perceptual processes (de Jong & Cho, 2012;Kang, 2011;LaCharité & Paradis, 2000;Peperkamp et al., 2008;Smith, 2006;Vendelin & Peperkamp, 2006), a better understanding of speech perception lends itself naturally to a better understanding of loanword patterns.
The current viewpoint naturally accounts for at least some of the loanword patterns observed in Mandarin Chinese. A few typical examples of loanword borrowings in Mandarin Chinese are presented in (2), where the relevant vowel is boldfaced and underlined (Note: Lexical tone is not marked in the transcriptions). For the words in (2), the source language is English, and the original words contain alveolar/palato-alveolar consonants in coda positions. As can be seen, when the original sourceword contains an alveolar consonant in a coda position, it is borrowed with an alveolar consonant followed by a [ɤ] vowel (22a); and when the original sourceword contains a palato-alveolar consonant in a coda position, it is largely borrowed with an alveo-palatal consonant followed by an [i] vowel (22b).

2.
Typical loanword patterns in Mandarin Chinese (Some of the data is from Lin (2007) In a corpus study consisting of 2423 borrowings into Mandarin from English (N = 1177), German (N = 977) and Italian (N = 269), Miao (2005) showed that the epenthetic vowels in loanwords containing illicit coda consonants show very specific patterns (Note: By 'epenthetic,' we do not imply an analysis that the vowel is not present in the underlying representation and surfaces in the surface representation. Instead, we just mean it descriptively to indicate it is not present in the sourceword). In alveolar contexts [t, tʰ], the epenthetic vowel is [ɤ] in 98% of the cases; and in alveo-palatal contexts, the vowel [i] is the epenthetic vowel in 83% of the cases. Given that the [ɤ] is an allophone of /ə/ in Mandarin Chinese (Lin, 2007), both these patterns are consistent with the perceptual claims presented in this paper. As can be observed, the results have an immediate bearing on the theoretical literature of loan-word adaptations, where there is a debate on the factors affecting loanword adaptations (Davidson, 2007;Kang, 2011;Peperkamp, 2005;Smith, 2006, inter alia). Whereas some claim that perceptual factors are perhaps the primary factor influencing loanword adaptation patterns (Peperkamp, 2005;Peperkamp & Dupoux, 2003), others claim that phonological factors play a big role, and perception is at best a minor factor in such patterns (Jacobs & Gussenhoven, 2000;LaCharité & Paradis, 2005;Paradis & LaCharité, 1997;Uffman, 2006). The proposed account in the current article suggests that the dichotomy between phonology and speech perception is perhaps a false one, given that phonological knowledge is argued to be crucially recruited during speech perception. The question, as we see it, is not whether a particular loanword pattern is due to phonology or speech perception, but instead how phonological knowledge is used during speech perception to infer the relevant underlying representations for the incoming acoustic input from the loanword source language. This is of course not to deny the role of nonauditory sources of loanwords.
A second issue that is brought to the fore is due to the claim that the illusory vowel is related, not to any vowel insertion pattern in the native phonology, but to vowel deletion patterns that exist in the native language. This is so, because as per the current viewpoint a native phonological process such as (/V 1 / → [∅]) can be used to infer a vowel in the underlying representation when no vowel is present in the acoustic signal, but a native phonological process such as (/∅/ → [V 1 ]) is not helpful in identifying the underlying representation when no vowel is present in the acoustic signal. This prediction does in fact bear out in many of the languages probed for illusory vowels-in Japanese, the illusory vowel in neutral contexts is /ɯ/ 22 (Dupoux et al., 2011); in Korean, in similar contexts, it is /ɨ/ (Durvasula & Kahng, 2015); and as seen above, in Mandarin Chinese, the illusory vowel in neutral contexts is /ə/. All of these vowels are involved in native vowel deletion processes in the respective languages, but not necessarily in any native vowel insertion processes. These are all also vowels that are the typical (or default) epenthetic vowels in the loanwords in the respective languages. Therefore, if a loanword pattern can be traced back to perceptual factors, then it is likely that the epenthetic vowel found in such loanwords is actually a vowel that is deleted, and not inserted, in the native phonology.