To appear in Charles Cairns & Eric Raimy, eds.

Handbook of the
Syllable. E.J. Brill.

Chapter 12

SYLLABLES IN SPEECH PROCESSING – EVIDENCE FROM PERCEPTUAL

EPENTHESIS

1. Syllables in speech production and processing

Much of phonological theory assumes that speech is both discreet and hierarchical. The speech

stream can be divided into discreet units (segments/sounds), and these units are hierarchically

organized into larger chunks such as syllables, feet, prosodic words, etc. (Nespor and Vogel 1986;

Selkrik 1978, 1986, etc.). Phonological rules are often formalized in terms of these higher order

groupings of sounds. This approach presents many problems to ph`onological theory. Not only is

it difficult to define these higher order categories accurately, but it is probably even harder to

find evidence of their reality.

Speech is transferred between speaker and listener as continuously varying air pressure

fluctuations. The structure of the pressure fluctuation is sometimes such that discontinuities can

be detected easily—for instance at the transition between a plosive and a vowel. This lends some

support to the idea that speech is discreet—the speech stream can be divided into discreet chunks

based on its physical properties. However, the boundaries between segments are not always

demarcated clearly in the physical speech stream, as anyone who has ever tried to find the

boundary between a vowel and glide on a spectrogram or waveform can attest. The fact that it is

not always possible to find correlates to concepts like segments in the physical world has lead

some to doubt the existence of discreet segments (Port and Leary 2005).

Just as difficult as finding physical evidence for the division of the speech stream into

discreet segments, is finding physical evidence for the grouping of these segments into larger
prosodically defined units such as feet, syllables, etc. Although there are temporal and

intonational properties of the speech stream that have been shown to correlate with the

boundaries of some of these larger prosodic units (Beckman and Edwards 1990; Byrd and

Saltzman 2003; Coetzee and Wissing 2007; Selkirk 1981, 2001; Sugahara 2005; Turk and

Shattuck-Hufnagel 2000, etc.), this correlation is not very strong. The observed variation in these

temporal and intonational measures is often large, and the differences between different prosodic

units are often very small (McQueen 1998, 21).

Given the difficulty of finding consistent physical correlates for these higher prosodic units, a

question that needs answering is whether (all of) these units are real. If physical evidence of

these units is lacking, is there evidence that language users parse the speech stream in terms of

these units? If a positive answer can be given to this question, then it can be concluded that some

of these prosodic units are at least psychologically real—that is, even if they do not exist in the

physical properties of the speech stream, the mind of the language user imposes these structures

onto the speech stream in the process of planning speech production and perceiving speech.

These structures may then have mental reality even if they lack physical reality.

In this paper, I provide evidence that listeners do parse the speech stream into higher order

prosodic groupings during the process of speech perception. Specifically, I show that listeners

are sensitive to syllabically defined allophonic distributions. In English, aspirated voiceless stops

only appear in syllable initial position. When presented with an acoustic signal such as [sphika]

with an unambiguous aspirated [ph], listeners impose a percept on this stimulus where [ph]

appears in syllable initial position. This leaves the [s] prosodically stranded—there is no vowel

in the actual acoustic stimulus to which the [s] can affiliate. I show that the listeners “perceive” a

vowel between the [s] and [ph] in this kind of situation, i.e. they perceive [sәphika]. The [ә] that

2
forms part of the percept is not present in the physical stimulus, and is therefore imposed onto

the percept by the perceptual system of the listener. I interpret this as evidence that the

perceptual system does impose higher order structure onto the speech stream, and that this part of

the speech processing process is even robust enough to result in percepts for which there is no

physical evidence.

Before getting into the details of the current study, I briefly review some existing evidence

for the mental reality of such higher order prosodic structures in both production and perception.

It is not my goal to review the extensive literature about this, but rather to discuss just a few

examples of the evidence that language users do use syllables both in production and perception.

Similar evidence for other prosodic constituents also exists (Clifton et al. 2006; Schafer et al.

2000; Wightman et al. 1992, etc.).

1.1. Evidence for syllables in speech production

One of the arguments for the use of syllables in speech production comes from studies of speech

errors. In a classic paper, Fromkin (1971, 38-40) shows that naturally occurring speech errors

point to the syllable as a planning unit for speech production. Exchange errors nearly always

respect the syllabic affiliation of the segments involved. Onsets exchange with onsets (‘carp-si-

hord’ for ‘harp-si-chord’) and rhymes with rhymes (‘hunk of jeap’ for ‘heap of junk’). Very

rarely observed are errors where the onset of one syllable exchanges with the coda of another

syllable. Fromkin claims that her corpus of errors contains only one such example (‘whip-ser’ for

‘whis-per’). This suggests that the exchanges are not mere segment exchanges, but exchanges of

segments in the same syllabic position, implying that the segments must be parsed into syllables

at the point in the production planning when the exchange occurs. Similarly, Fromkin notes that

speech errors that omit complete syllables (‘tre-men-ly’ for ‘tre-men-dous-ly’) are rather

3
common, while errors that omit a segment span that includes parts of two syllables are virtually

non-existent (‘tre-men-dy’ for ‘tre-men-dous-ly’).

Speech production models, such as those of Levelt (2001), also often contain a stage where

the phonological information is encoded into syllables. Levelt and his colleagues have conducted

several experiments over the past two plus decades that show evidence for this syllabic stage of

speech planning. I discuss the study of Cholin et al. (2004) as a recent example. They presented

Dutch speakers visually with groups of four morphologically related words. There were two

kinds of such groups: ‘constant’, where the first syllable of each word was always the same, and

‘variable’, where the initial syllable in one word was different. In both conditions, however, all

four words had the same segmental make-up up to the affixes. The difference between the two

lists is thus purely in terms of the syllable structure of one member. Examples are given in (1).

(1) Constant Variable

Meaning ‘lead’ ‘smoke’

Infinitive [leɪ.dәn] leiden [roː.kәn] roken

Gerund [leɪ.dәnt] leidend [roː.kәnt] rokend

Noun [leɪ.dәr] leider [roː.kәr] roker

Past [leɪ.dә] leidde [roːk.tә] rookte

Participants had to read each list out loud as accurately and fast as possible. Cholin et al.

found the response latency to be significantly greater in the variable than the constant condition.

They interpret this as evidence that the participants perform speech planning in terms of syllables.

In the constant condition, only one initial syllable has to be planned while two have to be

planned in the variable condition, accounting for the longer latency in the latter condition.

4
1.2. Evidence for syllables in speech perception

There is just as large a body of evidence that syllables play a role in speech perception, and I

discuss only a few examples from the recent literature. Mattys and Melhorn (2005) conducted a

dichotic listening experiment in which they presented participants simultaneously with two

different disyllabic auditory stimuli in the left and right ear, respectively. Neither of the two

stimuli were actual English words. However, exchanging corresponding parts of the left and

right stimuli did create a word. In one condition, the parts that need to be exchanged to form a

word corresponded to full syllables. For example, participants could be presented with [kir.fɪn] in

the left ear and [dɔl.mæl] in the right ear, where exchanging the initial syllables results in the real

word ‘dolphin’. In another condition, the parts that need to be exchanged consisted only of the

nucleus of a syllable. Participants could be presented with [dil.fɪn] in the left and [kɔr.mæl] in the

right ear. Exchanging the vowel of the first syllables now creates the word ‘dolphin’. In this

experiment, participants would first hear a production of the actual word [dɔl.fɪn] played

simultaneously in both ears. After that, they would hear one of the two dichotic stimuli described

just above. Their task was to decide whether the second presentation contained the word they just

heard. Mattys and Melhorn found that participants were significantly more likely to detect the

word in the dichotic stimulus when full syllables have to be exchanged than when only syllabic

nuclei have to be exchanged—i.e. more YES responses in [kir.fɪn]~[dɔl.mæl] than in

[dil.fɪn]~[kɔr.mæl]. They interpret this as evidence that the acoustic signal from each ear is

mapped onto a syllable sized representation before perceptual blending of the two stimuli

happens. The blending is then performed by manipulating these syllable sized chunks, with the

result that units smaller than a syllable are less accessible to the blending operation.

5
 Another recent study that points to a role for the syllable in speech processing is the

perceptual epenthesis study reported by Kabak and Idsardi (2007)—see also Berent et al. (2007),

Dupoux et al. (1999, 2001) and Lennertz and Berent (2007). This study is of particular interest,

since the experiments that I report later also investigate the phenomenon of perceptual epenthesis.

Kabak and Idsardi tested the perception by Korean listeners of consonant sequences that are

disallowed in Korean. They use two kinds of disallowed sequences. The first kind violates

restrictions on the syllable structure of Korean. Korean allows only a small set of consonants in

coda position—e.g. [k] is a possible coda while [tʃ] is not. The token [phaktha] is hence a possible

Korean word, while [phatʃtha] is not. The second kind of disallowed sequence violates not

restrictions on syllable structure, but rather restrictions on consonants that are allowed to follow

each other. Although Korean allows [k] in coda position, it does not tolerate a sequence of an

oral consonant followed by a nasal consonant. Like [phatʃtha], [phakma] is hence also not a

possible word. However, the reason for the impossibility of these two tokens is quite different.

[phatʃtha] is impossible because it has a disallowed coda [tʃ]. On the other hand, in [phakma], no

single consonant appears in a syllabic position where it is not allowed. This form is impossible

because of the oral plus nasal sequence.

Dupoux et al. (1999, 2001) demonstrated that Japanese listeners perform perceptual

epenthesis when presented with stimuli that contain disallowed consonantal sequences—i.e. they

perceive a vowel between the two disallowed consonants even if there is no actual vowel present.

When presented with an acoustic stimulus like [ebzo], Japanese listeners perceive [ebuzo], since

the sequence [bz] is disallowed in Japanese. Based on this, Kabak and Idsardi hypothesize that

Korean listeners should do the same when presented with disallowed forms like [phatʃtha] and

[phakma]. They perform a discrimination experiment where participants are presented with pairs

6
of tokens, and required to judge whether the two members in a pair were identical or not. The

crucial pairs consisted of one of the disallowed forms, and a corresponding form with a vowel

intervening between the two consonants—i.e. [phakma]~[phakʊma] and [phatʃtha]~[phatʃɪtha]. If

Korean listeners perform perceptual epenthesis, they should not be able to distinguish the

members of these pairs. Kabak and Idsardi found significantly better discrimination on pairs like

[phakma]~[phakʊma] than on pairs like [phactha]~[phatʃɪtha]. Perceptual epenthesis was therefore

significantly more likely to happen when a token was ill-formed for syllable structure reasons

than if it was ill-formed for reasons of consonant sequencing. Kabak and Idsardi conclude that

this gives evidence that the acoustic signal is processed in syllable-sized chunks during

perception.

2. The use of allophonic distributions to determine syllable structure

In this paper, I follow in the footsteps of earlier research on perceptual epenthesis by showing

that listeners perceive illusory vowels in order to arrive at a final percept that is syllabically well-

formed. However, I also deviate from this research tradition in two ways. First, earlier research

use acoustic stimuli that are phonemically ill-formed. Berent et al. (2007), for instance, present

English listeners with stimuli like [lbɪf, bdɪf, bnɪf], and find that these listeners perceive the

stimuli with an intrusive vowel—i.e. as [lәbɪf, bәdɪf, bәnɪf]. A stimulus like [lbɪf] cannot be

brought into agreement with English phonotactic grammar simply by mapping [l] onto some

other allophone of /l/, [b] onto some other allophone of /b/, or even by a combination of these

two options. No allophone of /l/ can appear in word-initial position followed by any other

consonant in English. These earlier studies therefore do not answer the question of at which level

of perceptual abstraction the acoustic stimulus is parsed into syllables (but see Lennertz and

Berent 2007). Is each of the individual segments first mapped onto its corresponding phoneme,

7
and then these phonemic percepts are syllabically parsed—i.e. allophonic variation is factored

out before syllabic parsing? Or is the syllabic parse performed on a percept that still contains

allophonic information? In order to address this question, I use stimuli that are allophonically ill-

formed but phonemically well-formed. Specifically, I use stimuli like [sphika]. In English, the

aspirated allophone [ph] of the phoneme /p/ is not allowed in onset position after an [s]. But

another allophone of /p/ is allowed in this position, namely [p]. If listeners first map the segments

in an acoustic signal onto phonemes and then parse the phonemes into syllables, then [sphika]

would have been transformed to /spika/ by the point that the syllabic parse is performed. The

percept /spika/ can be parsed into well-formed English syllables. On the other hand, if the

syllabic parse is performed on a percept that still contains allophonic information, then [sphika]

has to be syllabified, and no well-formed syllabification of such a representation is possible. If

listeners perform perceptual epenthesis when presented with a token like [sphika], it would be

evidence that syllabification happens before allophonic information is abstracted away.

The second way in which this study differs from previous studies is on a methodological

detail. In order to show that the vowels perceived by listeners are actually because of perceptual

epenthesis and not just an artifact of some unforeseen acoustic property of the stimulus, it is

necessary to have some control condition. In earlier studies, the control was always provided by

listeners from a different language in which the stimuli were well-formed. Kabak and Idsardi

(2007) use English listeners, since English allows both [k] and [tʃ] in coda position. They show

that English listeners do not perceive an epenthetic vowel in stimuli like [phatʃtha]. The fact that

the Korean listeners in their experiment do perceive such a vowel therefore cannot originate in

the acoustic properties of their stimuli, but must arise during grammar mediated perceptual

processing. Dupoux et al. (1999) similarly use French listeners as control (since stimuli like

8
[ebzo] are well-formed in French), and Berent et al. (2007) use Russian controls (since [bd, lb,

bn] are possible onset clusters in Russian). In the study that I report here, I use the exact same

listeners both in the perceptual epenthesis condition and in the control condition. This is possible

because I use stimuli that are ill-formed in English simply because of the context in which they

appear. The exact same stimuli are well-formed in a different acoustic context—that is, no well-

formed syllabic parse of [sphika] is possible, but splicing [la-] onto this exact same stimulus

gives [lasphika] for which a well-formed syllabic parse is possible. If English listeners perceive a

vowel between [s] and [ph] in [sphika] but not in [lasphika], then we can conclude that the vowel

percept in [sphika] does not originate in some unforeseen acoustic properties of the [s] and [ph].

This methodological difference between the study reported here and earlier studies is important

for two reasons. First, it is logistically easier in that listeners of only one language need to be

recruited. Secondly, by using the same listeners in both the experimental and the control

conditions the effect of perceptual epenthesis can be tested within individuals rather than across

different listeners with different native languages.

As already mentioned in the previous paragraph, I exploit the distribution of aspirated stops

in English in the design of the experiments reported later in this paper. To the best of my

knowledge, aspirated stops are only observed in absolute syllable initial position in all dialects of

English. It is not the case that all voiceless stops in syllable initial position are aspirated—in

American English, inter-vocalic /t/ is often flapped in the onset of an unstressed syllable, and in

some British dialects such /t/’s are often glottalized. But what is true is that all aspirated stops

appear in syllable initial position. If listeners pay attention to allophonic details such as

aspiration, and if they use this allophonic information when they parse the acoustic signal into

9
syllables, then they should insert a syllable boundary before every aspirated stop in the syllabic

parse imposed on the stimulus.

In the design of the experiments reported below, I assume that perceptual processing consists

of at least two stages. During the first stage of acoustic encoding a faithful acoustic copy of the

percept is created. No higher order prosodic structure, such as syllable structure, is present in the

acoustic representation formed during this stage, implying that the illusory epenthetic vowels are

also still absent at this stage of processing. In the second stage, phonological interpretation, the

listener parses the acoustic representation into higher order prosodic units, and tests these

structures for well-formedness against the grammar of his/her language. If the prosodic parse of

the acoustic representation is well-formed, the final linguistic percept is equal to this parse.

However, if the prosodic parse of the acoustic representation is ill-formed, the perceptual system

performs an extra step in which the representation is altered in some fashion so that it is brought

into agreement with the requirements of the grammar of the listener. This is the step during

which, for instance, perceptual epenthesis occurs. The listener then settles on a linguistic percept

that differs from the original acoustic representation created. This model is represented in Figure

1. See Kingston (2005) and Poeppel et al. (2008) for support of such a multi stage model.

10
 Linguistic Percept ≠
Acoustic Representation
Phonological Interpretation

Linguistic Percept =
Acoustic Representation changed
Acoustic Representation
to be well-formed

YES NO
Is this parse well-formed?

Impose prosodic parse on acoustic

representation

Create faithful Acoustic

Acoustic Encoding

Representation of stimulus

Figure 1 Stages in perceptual processing

Some of the implications of this model are tested in experiments discussed in the rest of this

paper. Specifically: (i) In this model, prosodic parsing is performed on the acoustic

representation—i.e. before allophonic information is abstracted away. Allophonic information

will therefore influence the prosodic structure imposed on the stimulus, so that a token like

[sphika] will receive a different prosodic parse from a token like [spika]. (ii) If a well-formed

prosodic parse exists that is faithful to the actual acoustic properties of the stimulus, then this is

the percept on which the listener will settle – [spika] will be perceived as [spi.ka]. On the other

11
hand, if no such a parse exists, the acoustic representation will be altered in some way to create a

representation that can be parsed into a well-formed prosodic structure – [sphika] will be

perceived [sә.phi.ka]. (iii) Since the perception of a stimulus like [sphika] involveds an extra

processing step, listeners should be slower in arriving at a perceptual decision to this kind of

stimulus. (iv) During the early stages of processing (acoustic encoding) stimuli like [sәphika] and

[sphika] will be represented differently, and this difference will only disappear in later processing

stages (phonological interpretation). In the next three sections, I discuss three perception

experiments designed to test these hypotheses.

3. Experiment 1: Same or Different?

In this experiment, participants are presented with pairs of stimuli, and their task is to decide

whether the members of a pair is the same or a different “word”. This task taps into the later

stages of perceptual processing. Not only do listeners have to listen to at least the first member of

a stimulus pair in full, but the task requires of participants to treat the stimuli as “words”, hence

increasing the likelihood that participants will rely on grammatical processing of the stimulus.

The expectation is hence that the participants’ responses will reflect the output of phonological

processing in the model represented in Figure 1. If the hypotheses stated above are correct,

participants should have difficulty distinguishing the members of pairs such as [sәphika]~[sphika].

Because no well-formed syllabic parse exists for the second member of this pair, participants are

expected to perform perceptual epenthesis on this stimulus and therefore to perceive the stimulus

as [sәphika]—i.e. identical to the other member of the pair. On the other hand, when presented

with a pair of stimuli that are identical to [sәphika]~[sphika] except that each member is preceded

by [la-], participants should more easily distinguish the members of the pair. Well-formed

12
syllabic parses are possible both for [lasәphika] and for [lasphika], and the participants should

therefore arrive at different perceptual decisions for the members of such a pair.

3.1. Methods

Participants. Fifteen undergraduate students from the University of Michigan participated in this

experiment. They were all native speakers of American English, reporting no hearing or

speaking deficits. Participants were paid for their participation.

Token selection and stimuli creation. The stimuli used in this experiment were all based on

the non-words in (2). All of these non-words have the structure [CV.sә.Chv́.CV] where [Ch] is

one of the voiceless aspirated stops of English, and [v́] is a stressed vowel. I recorded a native,

phonetically trained, female speaker of American English reading each of these tokens ten times.

From the ten recordings, I selected one to use in the creation of the stimuli. I selected as token to

use a repetition in which all of the sounds, and specifically the [ә], were clearly articulated. All

of the tokens were then equalized in intensity using the Scale intensity… function in Praat

(Boersma and Weenink 2008).

(2) Labial Coronal Dorsal

vɛsәphímu masәpháli kusәthíla lusәthápi basәkhífa fisәkhána

lasәphíka visәpháno fasәthímo misәtháku masәkhílu pisәkhámi

I discuss the further processing of [lusәthápi] here, but each of the other tokens was processed

in exactly the same way. All acoustic manipulation was done in Praat. From [lusәthápi], I created

two additional stimuli by splicing out first only the [ә], and then both the [ә] and the aspiration

associated with [th]. From each of these three stimuli (the original [lusәthápi] and the two created

13
by splicing out [ә] and [h]), I then created yet another stimulus by splicing off the initial [lu]. This

gave the six stimuli shown in (3).

(3) [s] word-medial [s] initial

[ә] and [h] present lusәthápi sәthápi

[ә] spliced out lusthápi sthápi

[ә] and [h] both spliced out lustápi stápi

These six stimuli were used to create stimuli pairs. Four ‘identical’ pairs were created by

matching up each of the four stimuli with [h] with itself. Four ‘different’ pairs were created by

matching a stimulus with both [ә] and [h] with stimuli from which [ә], and both [ә] and [h] have

been deleted. Two different instantiations of each of the different pairs were created—with the

stimuli appearing in each of the two possible orders. Stimuli in a pair were separated from each

other by 500 ms of silence. These stimuli pairs were presented to participants in a same/different

task. Given the hypotheses set out in the previous section, the expected responses to the different

pairs are given in (4)—only one of the two orders between the stimuli is given, but the same

response is expected for both orders. When the stop in Token 2 is unaspirated, [s]-initial and [s]-

medial pairs are expected to be equally discriminable. Legal syllable parses faithful to the

acoustic percept is possible for all four stimuli in these pairs, and it is expected that the listener

will arrive at percepts that are acoustically identical to the stimuli. However, when Token 2 is

aspirated, it is expected that the [s]-medial pair will be more discriminable than the [s]-initial pair.

There is no legal syllabic parse of [sthápi], and the listener is expected to perform perceptual

epenthesis, and hence to arrive at the percept [sәthápi] for this token, identical to the expected

percept for Token 1.

14
(4) Expected responses to different pairs

Token 2 Position Expected

Token 1 Token 2 Reason
[+asp]? of [s] response
Legal syllabic parse possible for
Medial lusәthápi lustápi Different
both tokens.
No
Legal syllabic parse possible for
Initial sәthápi stápi Different
both tokens.
Legal syllabic parse possible for
Medial lusәthápi lusthápi Different
both tokens.
Yes
No legal parse for Token 2 possible,
Initial sәthápi sthápi Same
perceptual epenthesis expected.

Procedure. Stimuli presentation and response collection were controlled with the SuperLab

software package. Participants were run in groups of up to three. Data collection was done in a

sound attenuated room. Participants were seated in front of individual laptop computers, each

with a response button box attached. Stimuli were presented via headphones, and responses were

collected via the button boxes. In each experimental block, each of the 12 pairs (4 identical, 4

different in two orders) was presented once. Since stimuli pairs were created from all 12 tokens

in (2), each block contained 144 stimuli pairs. Before a stimulus pair was presented, a visual cue

was presented in the middle of the computer screen. A token pair was then presented, and

participants indicated their choice by pushing one of two buttons marked as “same” and

“different”. Participants were instructed to respond as quickly and accurately as possible. The

block was presented four times, with participants allowed a self paced break between repetitions.

Stimuli pairs were differently randomized for each participant in each repetition. Before the first

block, participants were given 10 practice trials on stimuli pairs created in the same manner as

described above, but from tokens not used in the actual data collection part. Participants received

no feedback during the experiment.

15
3.2. Results and discussion

The response patterns were transformed to d'-scores as in Signal Detection Theory (MacMillan

and Creelman 2005) before they were statistically analyzed. Higher d'-scores correspond to

higher discriminability. The d'-scores were submitted to 2 × 2 ANOVA’s, both by participants

and by items, with factors aspiration (both members aspirated, one member unaspirated) and [s]-

position (word-initial, word-medial). Both main effects were found to be significant: aspiration

(by participant F(1, 14) = 139.0, p < .001; by item F(1, 11) = 575.9, p < .001), and [s]-position

(by participant F(1, 14) = 15.7, p < .002; by item F(1, 11) = 73.4, p < .001). Also the interaction

between the factors was significant (by participant F(1, 14) = 12.0, p < .005; by item F(1, 11) =

12.8, p < .005). The average by participant d'-scores are represented graphically in Figure 2.

[lusәthápi]~
[lustápi]

[lusәthápi]~
[lusthápi] [sәthápi]~
[stápi]

[sәthápi]~
[sthápi]

Figure 2 Mean d'-scores by participant for Experiment 1. Error bars show 95% confidence

intervals.

The first thing to note is that the discrimination is better for the pairs where one of the tokens

in the pair does not contain aspiration. This is a direct result of the actual size of the physical

16
acoustic difference between the members of a pair. Discrimination is better when the members

differ both in terms of aspiration and the presence/absence of a schwa, than if they differ only in

terms of the presence/absence of a schwa. Discrimination is also better for the pairs where [s]

appeared in word-medial position than the pairs where [s] appeared in word-initial position.

However, as indicated by the significant interaction between factors, and as is also clear from

Figure 2, this difference comes primarily from the pairs where both tokens contain an aspirated

stop. The d'-scores for [s]-initial and [s]-medial pairs with one unaspirated pair member do not

differ significantly from each other, as indicated by the overlapping confidence intervals. On the

other hand, d'-scores for pairs where both pair members are aspirated do differ significantly.

Specifically, discrimination is significantly better for pairs like [lusәthápi]~[lusthápi] than for

pairs like [sәthápi]~[sthápi]. Put differently, having aspiration on both pair members is more

detrimental to discriminability in the [s]-initial condition than in the [s]-medial condition.

It is important to remember that the actual size of the acoustic difference between [lusәthápi]

and [lusthápi], and between [sәthápi] and [sthápi] is exactly the same. In fact, since all these

stimuli were created from the same token, these two pairs are acoustically identical except for the

presence/absence of the initial syllable [lu-]. The difference in discriminability between these

two pairs therefore cannot originate in the acoustic differences between pair members. The

perceptual systems of the participants in this experiment are more likely to map [sәthápi] and

[sthápi], than [lusәthápi] and [lusthápi], onto the same percept. I interpret this as evidence that the

participants in this experiment tended to perceive the acoustic stimulus [sthápi] with an intrusive

schwa between the [s] and [th].

I interpret the results of this experiment as evidence that listeners parse acoustic percepts into

prosodic structure, and that these parses are evaluated against the grammar of the listener.

17
However, there is another possible explanation for these results that should be considered. It is

true that in the mental lexicon of an English listener, the sequence [#sәth…] has higher

probability than the sequence [#sth…]. In fact, [#sth…] probably has zero probability. It is hence

also possible that the effect observed in this experiment follows from such frequency statistics

calculated over the lexicon rather than from prosodic parsing and grammatical evaluation.

Listeners could posit several perceptual hypotheses that are partially consistent with the acoustic

stimulus, and then settle on the percept that has a higher statistical likelihood (Connine et al.

1993; Newman et al. 1997). When presented with a stimulus like [sthápi], the listener could

therefore posit both [sthápi] and [sәthápi] as perceptual hypotheses and consult his/her mental

lexicon to determine the probability of each of these percepts. Since the probability of [sth…] is

lower than that of [sәth…], the listener then settles on [sәthápi].

There are reasons to doubt this interpretation of the results. In all of the previous studies that

investigated the phenomenon of perceptual epenthesis, it has been shown that the results do not

originate from such frequency statistics (Berent et al. 2007; Dupoux et al. 2001; Kabak and

Idsardi 2007). However, to test this frequency based explanation more explicitly, I conducted a

control experiment. I calculated the frequency of the sequences [#sәk, #sәp, #sәt] and [#sik, #sip,

#sit] in CELEX (Baayen et al. 1995). The log transformed frequencies are given in (5). The

frequencies with the two vowels have the following relative ordering for each of the three places

of articulation: (#sik > #sәk), but (#sәp > #sip) and (#sәt > #sit). Under the frequency based

account, the prediction is hence that for a stimulus like [skh…] a percept with an intrusive [i] will

be more likely than a percept with an intrusive [ә]. However, for [skh…] and [sth…], a percept

with an intrusive [ә] should be more likely. To test this, I ran a second experiment that was

identical to the one described just above with one exception: the recordings from which the

18
stimuli were created had the structure [CV.si.Chv́.CV] rather than [CV.sә.Chv́.CV]. In order to

make the two sets of stimuli as comparable as possible, tokens for the control experiment were

selected such that the durations of [i] and [ә] did not differ significantly (two-tailed t(22) = 1.52,

p = .14). Stimuli creation, and experimental design and procedure were exactly the same as for

Experiment 1. Thirteen native speakers of American English participated in the control

experiment. By comparing the response patterns between these experiments, the predictions of

the frequency based account can be tested. The expected differences in the response patterns

between the experiments under the frequency based account are given in 0.

The d'-scores for the crucial pairs in the two experiments are represented in (7). The d'-scores

were higher in the control experiment (when the vowel was [i]) than in Experiment 1 (when the

vowel was [ә]) for token pairs at all three places of articulation. This goes against the predictions

of the frequency based account given in 0. The fact that discriminability is lower when V = [ә],

shows that an intrusive vowel percept is most likely to be [ә], irrespective of the frequency of the

sequences involved. Given that epenthetic vowels in English production are also more [ә]-like

(Davidson 2006, 2007), this is what is expected if this intrusive vowel percept is the result of the

grammar of the listeners rather than an influence of frequency statistics calculated over the

lexicon.

(5) Log transformed CELEX frequencies

#sVk #sVp #sVt

V = [ә] 2.03 4.22 2.15

V = [i] 3.49 1.81 1.70

19
(6) Expected response patterns under the frequency based account

Expected epenthetic
Token pair Expected discriminability
vowel percept

sVkhámi skhámi [i] (#sik > #sәk) Better when V = [ә]

sVpháno spháno [ә] (#sәp > #sip) Better when V = [i]

sVtháki stháki [ә] (#sәt > #sit) Better when V = [i]

(7) Comparison between the results of Experiment 1 and the control experiment

d'-score

V = [ә] V = [i]
Token pair

sVkhámi skhámi 1.25 1.82

sVpháno spháno 0.61 1.54

sVtháki stháki 1.14 1.84

Having shown that a frequency based account is unlikely, there still remains an alternative

explanation for the effect that also assumes that the effect has its origins in the syllable structure

grammar of English. All that the discrimination experiment shows is that participants were less

successful at discriminating pairs like [sәthápi]~[sthápi]. However, based only on the results of

this experiment, we do not know whether this is because they perceive [sthápi] with an intrusive

schwa, or whether they perform perceptual deletion on [sәthápi] and perceive this stimulus

without the schwa. Experiment 2 was designed to differentiate between these two explanations.

20
In this experiment, participants were presented with tokens such as [sәthápi] and [sthápi] with the

task of counting the number of syllables in each token. If they perform perceptual epenthesis,

they should count three syllables in [sthápi]. On the other hand, if they perform perceptual

deletion, they should count two syllables in [sәthápi].

4. Experiment 2: Syllable count

Berent et al. (2007) use a syllable count task to investigate the phenomenon of perceptual

epenthesis. This task has two advantages over the discrimination task. First, as explained in the

previous paragraph, it is better at diagnosing what the perceptual system does—perceptual

epenthesis or perceptual deletion? Secondly, in a syllable count task, participants hear only one

token before they have to respond. The time that elapses between stimulus presentation and

response is hence considerably shorter, so that the syllable count task gives better insight into the

time course of the processing involved in the perception of these stimuli. However, the syllable

count task still taps into the later phonological interpretation stages of speech processing.

Participants are asked to count “syllables” which implies that prosodic structure must be present

by the time that they respond. The nature of the task also requires of participants to listen to the

whole stimulus before a response can begin, thereby also slowing down the response and

increasing the likelihood that the response will be given only after phonological interpretation

has taken place.

4.1. Methods

Participants. Twelve undergraduate students from the University of Michigan participated in this

experiment. All participants were native speakers of American English with no known speech or

21
hearing deficits. There was no overlap in participants between Experiment 1 and 2. Participants

were paid for their participation.

Token selection and stimuli creation. The stimuli used in this experiment were created from

the 12 non-words in (8). These tokens have the form [CV.sә.Chv́C] where [Ch] is one of the

voiceless aspirated stops of English, and [v́] is a stressed vowel. The same female speaker as in

Experiment 1 was recorded reading these non-words ten times. I again selected one repetition of

each token to use for stimuli creation. As in Experiment 1, the token was selected so that all of

the sounds in the token were clearly articulated, and these selected tokens were equalized for

intensity in Praat.

(8) Labial Coronal Dorsal

vɛsәphím masәphál kusәthíf lusәthám basәkhíp fisәkhán

lasәphíf visәpháf fasәthík misәthál masәkhíf pisәkháf

Each of these 12 non-words was processed in the same way to create stimuli. I discuss the

processing of [fisәkhán] here as an example. From [fisәkhán], I created two additional stimuli by

splicing out first only the schwa (i.e. [fiskhán]), and then both schwa and the aspiration

associated with [kh] (i.e. [fiskán]). One additional stimulus was created from each of these three

stimuli by also deleting the initial [fi-] (i.e. [sәkhán], [skhán], and [skán]). The other 11 non-

words were processed in the same way, creating six stimuli from each. These stimuli were

presented to participants in a syllable count task. Based on the hypotheses set out at the end of

section 2, the expected response pattern for this experiment is given in (9). The response is

expected to agree with the actual syllable count for all but one of the six token types. Well-

formed syllable parses exist for all of the stimuli except for [skhán]. In agreement with the

22
hypothesis that the perceptual system settles on a final percept that is identical to the acoustic

input if a well-formed prosodic parse of the acoustic input exists, the expectation is hence that

participants will accurately perceive the number of syllables for all but [skhán]. For [skhán], the

aspirated [kh] has to be parsed into syllable onset position, leaving the word-initial [s] stranded.

The perceptual system is then expected to perform perceptual epenthesis to supply [s] with a

syllabic nucleus.

Procedure. Stimuli presentation and response collection were performed exactly as for

Experiment 1. Six stimuli were created from each of the 12 non-words in (8), for a total of 72

stimuli. To these stimuli, 60 fillers were added. The fillers were created from non-words of the

form [CV.C1ә.C2v́C], where [C2] is a sonorant consonant, and [C1C2-] is a possible syllable onset

in English. From the non-word [lusәmát], for instance, were created the fillers [lusәmát],

[lusmát], [sәmát], and [smát]. All test stimuli were presented 8 times each, and fillers 4 times

each. Stimuli presentation was differently randomized for each participant. Participants were

allowed a self-paced break halfway through the experiment. Before a stimulus was presented, a

visual cue was presented in the middle of the computer screen. The stimulus was then presented,

and participants indicated their responses by pressing one of three buttons on a response box

marked as [1], [2] and [3], respectively. The response time was recorded from the onset of each

stimulus up to the point where the participant registered a response. Participants received no

feedback during the experiment. Before the experiment began, participants did 10 practice trials,

randomly selected from the fillers. Participants were instructed to respond as accurately and

quickly as possible.

23
(9) Expected responses in Experiment 2

Actual Expected
Example [s]-position Schwa? [+asp]?
syllable count response

fisәkhán Yes Yes 3 3

fiskhán Medial No Yes 2 2

fiskán No No 2 2

sәkhán Yes Yes 2 2

skhán Initial No Yes 1 2

skán No No 1 1

4.2. Results and discussion

Responses were coded as ‘correct’ or ‘error’ in terms of the actual syllable count of each token,

as given in (9). The average response pattern per participant is represented in Figure 3. The

percent correct responses on each stimulus type was subjected to a one sample t-test to determine

whether it was significantly above chance. The results of these t-tests are given in (10). Except

for the [skhán]-type stimuli, the responses were indeed significantly above chance both by

participants and by items. For the [skhán]-type stimuli, neither the participants nor the items

analysis returned a significant result. In fact, the average percent correct responses was actually

below chance for these stimuli.

24
 3 2 2
2

2
1

3
2 3 1
[fisәkhán] [fiskán] [fiskhán] [sәkhán] [skán] [skhán]

Figure 3 Mean syllable counts for Experiment 2 by participant for each token type.

Numbers written in the bars indicate the response represented by the bar. Each

category on the x-axis represents all stimuli of that type.

The most informative comparison here is between [fiskhán] and [skhán]. These stimuli are

acoustically identical except for the presence/absence of the initial [fi-]. Since participants were

very accurate in their responses to [fiskhán]-type stimuli, the extra syllable percept to [skhán]-

type stimuli cannot come from some unforeseen acoustic property of these stimuli. In the [s]-

initial condition, participants perceive an extra vowel between [s] and [kh], but in the [s]-medial

condition they do not perceive such a vowel between the exact same [s] and [kh]. This vowel

percept therefore originates in the perceptual system and not in the acoustic properties of the

stimulus. This result also answers the one question that was left open at the end of Experiment 1.

The results of Experiment 1 showed that participants cannot accurately discriminate stimuli such

as [sәthápi] and [sthápi]. What was not clear was whether this is because they incorrectly perceive

[sәthápi] as [sthápi], or whether they incorrectly perceive [sthápi] as [sәthápi]. The fact that the

25
participants in Experiment 2 perceived [skhán] with an intrusive vowel, indicates that it is [sthápi]

that was misperceived in Experiment 1.

(10) Results of t-tests for Experiment 2

Participant analysis Item analysis

Stimulus type t(11) p t(11) p

[fisәkhán] 12.3 < 0.001 36.0 < 0.001

[fiskán] 21.5 < 0.001 88.0 < 0.001

[fiskhán] 12.2 < 0.001 6.1 < 0.001

[sәkhán] 30.7 < 0.001 45.9 < 0.001

[skán] 2.0 < 0.04 9.2 < 0.001

[skhán] -0.3 0.78 -0.9 0.37

The response times for Experiment 2 were also statistically analyzed. For each participant, the

response times were first standardized. Response times for [s]-initial and [s]-medial stimuli were

separately standardized, since the [s]-medial stimuli were all slightly longer because of the extra

initial [CV-]. After the response times where thus transformed, all responses that were more than

two standard deviations from the mean were excluded. This led to the exclusion of just over 4%

of all responses.

The response times on the [fisәkhán]- and [fiskhán]-type stimuli, and the [sәkhán]- and

[skhán]-type stimuli were compared, using one-tailed paired-sample t-tests. [fisәkhán] contains an

extra [ә] that is absent from [fiskhán]. Since response time was measured from the onset of a

stimulus, the extra [ә] therefore contribute a little to the response time measured for [fisәkhán],

26
and similarly for the extra [ә] in [sәkhán]. If only the actual duration of a stimulus determines the

response time, then the response times to [fisәkhán] and [sәkhán] should on average be longer

than those to [fiskhán] and [skhán]. The analyses for [s]-medial tokens like [fisәkhán] and [fiskhán]

returned non-significant results both by items (t(11) < .01, p = .50) and by participants (t(11) =

0.29, p = .39). The response times to these two types of tokens therefore did not differ

significantly. However, the results for the [s]-initial tokens like [sәkhán] and [skhán] were quite

different. They were highly significant by items (t(11) = 8.3, p < .001) and tended toward

significance by participants (t(11) = 1.6, p = .07). These results are represented graphically in

Figure 4. Inspection of this figure shows that the response time for [fisәkhán] and [fiskhán] is

about equal, as is expected from the results of the t-tests reported just above. The response time

for [sәkhán] is also shorter than that for [fisәkhán] and [fiskhán], which is again expected since

[sәkhán] is physically shorter, lacking the initial [fi-]. However, the response time for [skhán]-

type stimuli is longer than that for [fisәkhán]- and [fiskhán]-type stimuli. This is true in spite of

the fact that the [skhán]-type stimuli are in reality the shortest of the four stimuli types compared

in this figure.
ms

fisәkhán fiskhán sәkhán skhán

Figure 4 Mean response times by item for [fisәkhán]-, [fiskhán]-, [sәkhán]-, and

[skhán]-type stimuli. Error bars show 95% confidence intervals.

27
 Most importantly, the response time for [skhán]-type stimuli is significantly longer than that

for [sәkhán]-type stimuli, in spite of the fact that the [sәkhán]-type stimuli are actually longer.

The longer response time for [skhán]-type stimuli therefore does not originate in the acoustic

properties of these stimuli, but rather in the processing time dedicated to these stimuli. In the

processing model represented above in Figure 1, stimuli like [skhán] are subject to an extra stage

of processing. The perceptual system first creates an acoustic representation of the stimulus, and

then imposes a prosodic parse on this representation. This prosodic parse is then measured

against the demands of the grammar, and if it is found to be well-formed a perceptual decision is

taken. However, no well-formed prosodic parse of [skhán] is possible. At this stage, the

perceptual system therefore has to change the mental representation of the stimulus (via

perceptual epenthesis) and re-parse the stimulus prosodically. This additional step takes up

additional processing time, resulting in slower response times for stimuli like these.

5. Experiment 3: Begins on …

The results of Experiments 1 and 2 show that listeners represent stimuli like [sthapi]~[sәthapi]

and [sәkhan]~[skhan] the same during late stages of perceptual processing, in agreement with the

processing model represented in Figure 1. However, these results do not give evidence for the

earlier stage of acoustic encoding proposed in Figure 1, where the members of these pairs are

hypothesized to have different mental representations. Although the results of Experiments 1 and

2 are hence consistent with the model represented in Figure 1, they are also consistent with other

processing models. Another possibility is that the percept [sәkhán] is activated for both [skhán]

and [sәkhán] stimuli right from the earliest stages of processing, but that activation of this percept

reaches the critical decision level sooner for an input stimulus that is actually identical to this

percept. Such an interpretation would be consistent with a perceptual processing model that does

28
not contain an initial stage of phonetic encoding that is informed only by the actual acoustic

input stimulus. If this were correct, then both [skhán] and [sәkhán] will be mentally represented

as [sәkhán] from the earliest stages of processing. There should then be no stage of processing

where the perceptual system can discriminate these stimuli consistently. On the other hand, if

there is an initial stage of phonetic encoding, as represented in Figure 1, where the mental

representations of [skhán] and [sәkhán] stimuli differ, then there must be a stage of processing

where the perceptual system can discriminate these two types of stimuli.

The experimental tasks used in Experiments 1 and 2 by their very nature tap into the later

stages of perceptual processing and therefore cannot give evidence about the earlier stages of

processing. Experiment 3 is designed to address this issue. In this experiment, participants are

presented with the same stimuli as in Experiment 2. However, rather than counting syllables,

participants are instructed to decide whether a stimulus begins on, for instance, [sk…] or [sәk…].

This task does not require prosodic parsing, nor does it require of participants to listen to the full

token before responding. Participants can thus respond faster so that their responses may reflect

earlier stages of processing.

5.1. Methods

Participants. Eleven undergraduate students from the University of Michigan participated in this

experiment. All participants were native speakers of American English with no known speech or

hearing deficits. There was no overlap between participants in this and earlier experiments.

Participants were paid for their participation.

Token selection and stimuli. The exact same stimuli as in Experiment 2 were used in

Experiment 3.

29
 Procedure. As with Experiments 1 and 2, stimulus presentation and response collection were

controlled with Superlab. Presentation of each token was preceded by a visual cue on the

computer monitor of two options represented in English orthography. The two options always

corresponded to a percept with and without a schwa, with schwa represented with the letter e.

For [s]-initial tokens, the options therefore consisted of s or se followed by one of {p, t, k}. For

example, a token like [skhán] would be preceded by the visual cue “sk… OR sek…”. For [s]-

medial tokens, the visual cue consisted of an orthographic representation of the initial syllable,

followed by s or se and one of {p, t, k}. For example, [fisәkhán] would be preceded by “feesek…

OR feesk…”. In the first syllable of these [s]-medial tokens, [i] was always represented by the

letters ee, [u] by oo, [ɛ] by e, and [a] by a. The order between the cue with the schwa and without

the schwa was balanced by token type. Participants were instructed to read the visual cue before

the auditory token was presented, and to decide on which sound sequence the auditory stimulus

begins by pushing either the leftmost or the rightmost button on the response box. Response

times were recorded from the onset of the auditory stimulus to the moment that a participant

pushed a button on the response box. In order to increase the likelihood that participants’

responses will be based on earlier stages of processing, they were encouraged to respond as

quickly as possible, even before the full token has played if that was possible. Each token was

included once per experimental block, for 72 total tokens. The block was presented 4 times, with

participants allowed a short self paced break between blocks. Tokens were differently

randomized for each participant and each block. Before the experimental trials began,

participants received 20 practice trials, randomly selected from the experimental trials. A

different selection was made for each participant. Participants received no feedback.

30
5.2. Results and discussion

In order to determine whether the different design succeeded in eliciting faster responses, and

hence responses based on earlier stages of processing, the response times of Experiment 3 were

analyzed. The response times were first standardized, exactly as in Experiment 2, and responses

more than 2 standard deviations from the mean for each participant were excluded, resulting in

the exclusion of just under 4% of all responses. As a rough measure of success, the response

times to different tokens in Experiment 2 (µ = 999 ms) and Experiment 3 (µ = 930 ms) were

compared using a one-tailed, paired sample t-tests, returning a highly significant difference (t(71)

= 3.76, p < .001). On average, response times in Experiment 3 were therefore shorter than in

Experiment 2. However, the more important comparison is the response time to [skhán] type

tokens between the two experiments. The goal of Experiment 3 is to elicit responses for these

tokens at an earlier stage of processing, and we therefore need the response time to these tokens

to be faster in Experiment 3 than Experiment 2. The response times on these tokens for the two

experiments are represented in Figure 5. These response were also compared using one-tailed,

two sample t-tests. These tests returned significant results both by items (t(22) = 5.8, p < 0.001)

and by participants (t(21) = 2.0, p < 0.03). I conclude that the difference in task in Experiment 3

has therefore resulted in speeding up the responses of the participants.

31
 ms

Experiment 2 Experiment 3
Figure 5 Average response time by item to [skhán] type stimuli in Experiments 2 and 3.

Error bars show 95% confidence intervals.

Responses were coded as ‘correct’ or ‘error’ in terms of the actual acoustic propertiess of each

token. The average response pattern per participant is represented in Figure 6. The percent

correct responses on each stimulus type was subjected to a one sample t-test to determine

whether it was significantly above chance. The results of these t-tests are given in (11)

32
 [fisәkhán] [fiskán] [fiskhán] [sәkhán] [skán] [skhán]
Figure 6 Mean percent correct for Experiment 3 by participant for each token type.

(11) Results of t-tests for Experiment 3

Participant analysis Item analysis

Stimulus type t(10) p t(11) p

[fisәkhán] 76.5 < 0.001 94.9 < 0.001

[fiskán] 33.9 < 0.001 54.1 < 0.001

[fiskhán] 8.3 < 0.001 5.9 < 0.001

[sәkhán] 45.3 < 0.001 52.2 < 0.001

[skán] 21.0 < 0.001 25.8 < 0.001

[skhán] 5.4 < 0.001 7.2 < 0.001

The percent correct response was significantly above chance for all token types. Most

importantly, the response pattern to [skhán] token types in Experiment 3 is different from those in

33
Experiment 2. In Experiment 2, participants were about equally as likely to perceive these tokens

as [sәkhán] (with two syllables) or as [skhán] (with only one syllable). However, in Experiment 3,

participants overwhelmingly perceive these stimuli as starting on [sk…], and therefore

distinguish these stimuli from [sәkhán] type stimuli. Participants differentiate [skhan] and [sәkhan]

at the faster response rates of Experiment 3 but they confuse [skhan] with [sәkhan] at slower

response rates in Experiment 2. I interpret this as evidence that though [skhan] and [sәkhan]

receive the same mental representation at later stages of processing, there is, in accordance with

the model in Figure 1, an earlier stage of phonetic encoding where these types of tokens have

separate representations.

6. General discussion

6.1. Mismatch between stimulus and percept as evidence about grammar

The human perceptual system does not function like a photocopier. It does not merely translate

the input signal into a faithful mental representation. It rather takes the input signal and

transforms it into a percept that bears resemblance to the input signal, but that also differs from it

in lawful ways. This is true of all aspects of human perception—visual (Gordon 2004), olfactory

(Wilson and Stevenson 2006), tactile (Schiff and Foulke 1982), auditory (Warren 2008), etc. One

of the central goals of cognitive psychology is to discover the laws that dictate the ways in which

the percept can deviate from the actual input. Part of the deviation is just the result of the

physiological properties of the different organs involved in perception. But physiology does not

explain all of perception. There are also operational or processing laws that influence the way in

which the stimulus is processed once it has been translated by the primary perceptual organs into

some kind of mental representation. One of the processing related factors that influence the

auditory perception of speech is the grammar of the listener. Over the past several decades, a

34
large body of evidence has been amassed showing that the exact same acoustic speech stimulus

can be perceived very differently by speakers of different languages. Under the assumption that

there are no major differences in the physiology of the auditory perception organs between

speakers of different languages, such differences must have their origin somewhere in the

processing that happens in the mind after the auditory speech stimulus has been translated into a

mental representation. In order to understand the human perceptual system and the cognitive

processes that support perception, it is therefore necessary to understand which aspects of

grammar can influence speech perception, and how these aspects influence speech perception.

In the same way that the study of speech perception is important for our general

understanding of human cognition, it is also important for our understanding of the grammatical

competence of the language user. Studying the mismatches between acoustic speech stimuli and

the percepts that listeners form, can give us insight into the structure of grammar. In this paper,

we have seen how such a mismatch can provide evidence that speech is parsed into higher order

prosodic structure. The percept at which listeners arrive for the exact same acoustic stimulus

depends partially on the position that the stimulus occupies in higher order prosodic structure.

Specifically, an acoustic stimulus that corresponds to the phone sequence [sth] is perceived as

such when it appears in a context where a syllable boundary can be lawfully inserted between the

[s] and [th]—i.e. in a context such as [lu__ápi]. But this same acoustic stimulus is perceived as

[sәth] when it appears in a context where this syllable parse is not possible—i.e. in a context such

as [__ápi].

In English, aspirated stops are only tolerated in absolute syllable initial position. This

presents a possible solution for the mismatch between the acoustic stimulus [sth] and the percept

[sәth] in the [__ápi] context. The listener accurately perceives the aspiration in the acoustic

35
stimulus. The aspiration is taken as the cue for the start of a new syllable, and the first three

phones of the [sthápi] stimulus are hence prosodically parsed as [s.thá…]. This leaves the word-

initial [s] stranded without a syllabic nucleus with which it can affiliate, and causes the

perceptual system to perceive an illusory schwa between the [s] and [th]. In the [lu__ápi] context,

this is not necessary. As before, the aspiration is taken as a cue to start a new syllable, resulting

in a prosodic parse [lus.tha…]. Since the [s] is now preceded by a vowel with which it can

affiliate, this sequence is well-formed, and there is no need for an illusory vowel. Somewhere

during the processing of the speech stimulus, higher order prosodic structure is therefore

imposed on the stimulus. Even though it is impossible to detect a syllable boundary between [s]

and [th] in the acoustics of the stimulus, the way in which this stimulus is processed gives

evidence for the reality of such a prosodic boundary.

It is important to note what the repair is that is imposed on the stimulus. When faced with a

prosodically ill-formed representation such as [s.tha…], there are several ways in which it could

be changed to become well-formed. The aspiration can be perceptually deleted, so that the

percept can be parsed syllabically as [stá…]. Alternatively, the initial [s] can be perceptually

deleted, giving the well-formed parse [thá…]. Another conceivable perceptual repair is to

perceptually insert a vowel before [s], as in [әs.thá…]. All of these would result in prosodically

well-formed percepts. However, the English listeners in the experiments reported above

preferred a different repair—perceptually inserting a vowel between the [s] and [th]. A question

that this paper does not address is whether the specific perceptual repair used to remedy such

prosodically ill-formed stimuli is determined by the grammar of the listeners (i.e. it is language

specific), or whether the same repair will be observed irrespective of the language of the listeners.

Further research would be required to answer this question—though see Fleischhacker (2001) for

36
an exploration of the idea that the perceptual repair is selected that is acoustically most similar to

the actual acoustic stimulus.

6.2. The time course of perceptual epenthesis

Behavioral data, such as those reported in this paper, are rather course-grained in their time

resolution, and therefore not optimal for probing into the time course of mental processing where

very small time differences can be at play. Even so, behavioral data can reveal much about

processing. The model represented in Figure 1 makes two very specific predictions with regard

to the time course of processing, and the experiments presented above provide evidence for both

of these predictions.

First, the model predicts that there will be an additional processing step involved in tokens

that are prosodically ill-formed, and hence that such tokens will take longer to process. The

results of Experiment 2 are in agreement with this prediction. Although [stháp]-type stimuli are

physically of shorter duration than [sәtháp]-type stimuli, participants are slower at responding to

[stháp]-type stimuli. During acoustic encoding, faithful acoustic representations of both of these

token are made. In the next step of processing, prosodic parses are imposed on these acoustic

representations, and these parses are compared against the grammar for well-formedness. Since a

well-formed prosodic parse of [sәtháp] is possible, namely [sә.tháp], no additional processing is

required for this token. On the other hand, no well-formed parse of [stháp] exists—both [stháp]

and [s.tháp] are not in agreement with English phonological grammar. An additional processing

step is hence required to bring this token type into agreement with the phonology of English,

resulting in slower processing.

However, although the longer processing time associated with [stháp]-type stimuli are

consistent with the extra processing step of the model represented in Figure 1, it does not

37
actually confirm the existence of such an extra processing step. It is also possible that multiple

perceptual hypotheses are entertained from the earliest stages of perceptual processing, so that

the percept [sәtháp] is activated for both [stháp] and [sәtháp] stimuli even during early auditory

processing, but that activation of this percept reaches the critical decision level sooner for an

input stimulus that is actually identical to this percept. Such an interpretation would be consistent

with a perceptual processing model that does not contain an initial stage of phonetic encoding

that is informed only by the actual acoustic input stimulus. However, the results of Experiment 3

give evidence for the existence of such an early acoustic encoding stage in speech processing. In

this experiment, participants responded faster to and clearly differentiated [stháp] and [sәtháp]

type stimuli.

Even the task in Experiment 3 is still behavioral in nature, and therefore not ideally suited to

tap into the earliest stages of auditory processing. Non-behavioral methods that directly measure

real time brain response to stimulus presentation, such as ERP (event related potentials) studies,

are better suited to studying early stages and the fine-grained time course of processing

(Praamstra et al. 1994; Praamstra & Stegeman 1993). In ERP studies, two stimuli are presented

with a short inter stimulus interval, and brain activity is measured during the presentation of the

second stimulus. Some of the stimuli pairs presented are identical, while others differ from each

other. By comparing brain activity between the “same” and “different” conditions, it can be

determined whether these two conditions are processed differently by the brain.

Wagner and Schafer (2008) conducted such an ERP study with stimuli very similar to those

used in the experiments discussed in this paper. The stimuli in their “same” condition contained

pairs like [ptaki]~[ptaki]/[pәtaki]~[pәtaki], and the stimuli in their “different” conditions pairs

like [ptaki]~[pәtaki]/[pәtaki]~[ptaki]. The participants in their study performed a same/different

38
task similar to that used in Experiment 1 above. They found that English listeners could not

consistently identify [ptaki]~[pәtaki] pairs as “different”, giving more evidence of perceptual

epenthesis. The crucial difference between their experiment and Experiment 1 is that they

collected ERP data while their participants were performing the same/different task. They found

evidence for early differences in brain activity between same and different pairs, around 500 ms

after the onset of the second pair member. But this difference disappears at later processing

stages, somewhere around the 1000 ms mark. They interpret this as evidence for a separate early

stage of processing, consistent with the phonetic encoding stage proposed in Figure 1 above. The

behavioral same/different response happens later than the 1000 ms time point, and is therefore

based on the later processing stages where the [ptaki] and [pәtaki] are no longer distinguished.

Dehaene-Lambertz et al. (2000) perform a similar study with Japanese listeners, investigating

the results of Dupoux et al. (1999, 2001) further. Dupoux et al. showed that Japanese listeners

cannot discriminate tokens like [ebzo] and [ebuzo] consistently because of perceptual epenthesis

in [ebzo]. The ERP results of Dehaene-Lambertz et al., however, are in conflict with those of

Wagner and Schafer described just above. They did not find evidence for consistent differences

in the brain activity of Japanese listeners between the “same” ([ebzo]~[ebzo], [ebuzo]~[ebuzo])

and “different” ([ebzo]~[ebuzo], [ebuzo]~[ebzo]) conditions, not even during the earliest stages

of processing. They conclude that their results “suggest that the impact of phonotactics takes

place early in speech processing and support models of speech perception, which postulate that

the input signal is directly parsed into native language phonological format” (Dehaene-Lambertz

et al. 2000, 635). Their results thus speak against the initial stage of phonetic encoding proposed

in Figure 1.

39
 Given these conflicting results, it cannot be decided definitively at current whether there is

indeed a separate level of phonetic encoding—however see Kingston (2005) and Poeppel et al.

(2008) for more arguments for such a level. More research, and specifically more ERP type

investigations of similar phenomena, is necessary.

References
Baayen, R. Harald, Richard Piepenbrock and Leon Gulikers L. Gulikers. 1995. The CELEX
Lexical Database (Release 2) [CD-ROM]. Philadelphia, PA: Linguistic Data Consortium,
University of Pennsylvania.
Beckman, Mary E. and Jan Edwards. 1990. Lengthenings and shortenings and the nature of
prosodic constituency. In John Kingston and Mary E. Beckman (eds.) Papers in
Laboratory Phonology I: Between the Grammar and the Physics of Speech. Cambridge:
Cambridge University Press, 152-178.
Berent, Iris and Tracy Lennertz. 2007. What we know about what we have never heard: Beyond
phonetics. Reply to Peperkamp. Cognition. 104:638-643.
Berent, Iris, Donca Steriade, Tracy Lennertz and Vered Vaknin. 2007. What we know about
what we have never heard: Evidence from perceptual illusions. Cognition. 104:591-630.
Boersma, Paul and David Weenink. 2008. Praat: Doing Phonetics by Computer (Version 5.0.24).
[Computer program available at http://www.praat.org/.].
Byrd, Dani and Elliot Saltzman. 2003. The elastic phrase: modeling the dynamics of boundary-
adjacent lengthening. Journal of Phonetics. 31:149-180.
Cholin, Joana, Niels O. Schiller and Willem J.M. Levelt. 2004. The preparation of syllables in
speech production. Journal of Memory and Language. 50:47-61.
Clifton, Charles, Katy Carlson and Lyn Frazier. 2006. Tracking the what and why of speakers’
choices: Prosodic boundaries and the length of constituents. Psychonomic Bulletin &
Review. 13:854-861.
Coetzee, Andries W. and Daan Wissing. 2007. Global and local durational properties in three
varieties of South African English. The Linguistic Review. 24:263-289.

40
Connine, Cynthia M., Debra Titone and Jian Wang. 1993. Auditory word recognition: extrinsic
and intrinsic effects of word frequency. Journal of Experimental Psychology: Learning,
Memory, and Cognition. 19:81-94.
Davidson, Lisa. 2006. Phonology, phonetics, or frequency: Influences on the production of non-
native sequences. Journal of Phonetics. 34: 104-137.
––––––– 2007. The relationship between the perception of non-native phonotactics and loanword
adaptation. Phonology. 24:261-286.
Dehaene-Lambertz, Ghislaine, Emmanuel Dupoux and Ariel Gout. 2000. Electrophysiological
correlates of phonological processing: a cross-linguistic study. Journal of Cognitive
Neuroscience. 12:635-647.
Dupoux, Emmanuel, Kazuhiko Kakehi, Yuki Hirose, Christophe Pallier and Jacques Mehler.
1999. Epenthetic vowels in Japanese: A perceptual illusion? Journal of Experimental
Psychology: Human Perception and Performance 25, 1568–1578.
Dupoux, Emmanuel, Christophe Pallier, Kazuhiko Kakehi & Jacques Mehler (2001). New
evidence for prelexical phonological processing in word recognition. Language and
Cognitive Processes. 5:491–505.
Fleischhacker, Heidi. 2001. Cluster-dependent epenthesis asymmetries. In Adam Albright and
Taehong Cho (eds.) UCLA Working Papers in Linguistics 7. Papers in Phonology 5. Los
Angeles: UCLA Linguistics Department, 71-116.
Fromkin, Victoria. 1971. The non-anomalous nature of anomalous utterances. Language. 47:27-
52.
Gordon, Ian E. 2004. Theories of Visual Perception. New York, NY: Psychology Press.
Kabak, Bariş and William J. Idsardi. 2007. Perceptual distortions in the adaptation of English
consonant clusters: Syllable structure or consonantal contact constraints? . Language and
Speech. 50:23-52.
Kingston, John. 2005. Ears to categories: New arguments for autonomy. In Sonia Frota, Marina
Vigario and Maria João Freitas (eds.) Prosodies: With Special Reference to Iberian
Languages. New York: Mouton de Gruyter, 177-222.
Lennertz, Tracy and Iris Berent. 2007. Markedness constraints on the perception of s/z-initial
onset clusters. Paper presented at the Workshop on Variation, Gradience and Frequency

41
 in Phonology. Stanford University. [Downloaded on June 23, 2008 from
www.stanford.edu/dept/linguistics/linginst/nsf-workshop/Lennertz&Berent_Poster.pdf.].
Levelt, Willem J.M. 2001. Spoken word production: A theory of lexical access. Proceedings of
the National Academy of Sciences of the United States of America. 98:13464-13471.
MacMillan, Neil A. and C. Douglas Creelman. 2005. Detection Theory: A User’s Guide.
Mahwah, NJ: Lawrence Erlbaum.
Mattys, Sven L. and James F. Melhorn. 2005. How do syllables contribute to the perception of
spoken English? Insight from the migration paradigm. Language and Speech. 48:223-253.
McQueen, James M. 1998. Segmentation of continuous speech using phonotactics. Journal of
Memory and Language. 39:21-46.
Nespor, Marina and Irene Vogel. 1986. Prosodic Phonology. Dordrecht: Foris.
Newman, Rochelle S., James R. Sawusch and Paul A. Luce. 1997. Lexical neighborhood effects
in phonetic processing. Journal of Experimental Psychology: Human Perception and
Performance. 23:873-889.
Poeppel, David, William J. Idsardi and Virginie van Wassenhove. 2008. Speech perception at the
interface of neurobiology and linguistics Philosophical Transactions of the Royal Society.
B 363:1071-1086.
Port, Robert F. and Adam P. Leary. 2005. Against formal phonology. Language. 81:927-964.
Praamstra, Peter, Antje S. Meyer and Willem J.M. Levelt. 1994. Neurophysiological
manifestations of phonological processing: Latency variation of a negative ERP
component timelocked to phonological mismatch. Journal of Cognitive Neuroscience.
6:204-219.
Praamstra, Peter and Dick F. Stegeman. 1993. Phonological effects on the auditory N400 event-
related brain potential. Cognitive Brain Research. 1:73-86.
Schafer, Amy, Katy Carlson, Charles Clifton and Lyn Frazier. 2000. Focus and the interpretation
of pitch accent: Disambiguating embedded questions. Language and Speech. 43:75-105.
Schiff, William and Emerson Foulke (eds.) 1982. Tactual Perception: A Sourcebook. Cambridge:
Cambridge University Press.
Selkirk, Elizabeth O. 1978. On prosodic structure and its relation to syntactic structure. In
Thorstein Fretheim (ed.) Nordic Prosody II. Trondheim: Tapir, 111-140.

42
––––––– 1981. On the nature of phonological representation. In Terry Myers, John Laver and
John Mathieson Anderson (eds.) The Cognitive Representation of Speech. Amsterdam:
North Holland, 379-388.
––––––– 1986. On derived domains in sentence phonology. Phonology Yearbook. 3:371-405.
––––––– 2001. On the phonologically driven non-realization of function words. In Charles Chang,
Michael J. Houser, Yuni Kim, David Mortensen, Mischa Park-Doob & Maziar
Toosarvandani (eds.) BLS 27: Proceedings of the Twenty-Seventh Annual Meeting of the
Berkeley Linguistics Society. Berkeley: Berkeley Linguistics Society, 257-270.
Sugahara, Mariko. 2005. Post-FOCUS prosodic boundaries in Tokyo Japanese: Asymmetric
behavior of an F0 cue and domain-final lengthening. Studia Linguistica. 59:144-173.
Turk, Alice. E. and Stefanie Shattuck-Hufnagel. 2000. Word-boundary-related duration patterns
in English. Journal of Phonetics. 28:397-440.
Wagner, Monica Palmieri and Valerie Shafer. 2008. Phonotactic influences on the perception of
consonant clusters by English and Polish listeners. [Handout and audiorecording of
presentation downloaded on June 26, 2008 from
http://www.cunyphonologyforum.net/SYLLPAPERS/.].
Warren, Richard M. 2008. Auditory Perception: An Analysis and Synthesis. Cambridge:
Cambridge University Press.
Wightman, Colin W., Stefanie Shattuck-Hufnagel, Mari Ostendorf and Patti J. Price. 1992.
Segmental durations in the vicinity of prosodic phrase boundaries. Journal of the
Acoustical Society of America. 91:1707-1717.
Wilson, Donald A. and Richard J. Stevenson. 2006. Learning to Smell: Olfactory Perception
from Neurobiology to Behavior. Baltimore, MD: Johns Hopkins University Press.

43