• §2.1 Introduction To Chapter 2
• §2.2 High Vowel Devoicing
  • §2.2.1 Factors Affecting the Application of HVD
    • §2.2.1.1 Speech Rate
    • §2.2.1.2 Pitch Accent Placement
    • §2.2.1.3 Sociological Factors
    • §2.2.1.4 Moraic Position of the Vowel
    • §2.2.1.5 Consonantal Environment
  • §2.2.2 HVD as a Feature-Changing Rule
  • §2.2.3 An Alternative Account: Gestural Overlap
  • §2.2.4 HVD as Both Phonetic and Phonological
  • §2.2.5 Summary of Arguments For Both Processes
• Chapter 2 notes (§§2.1 ~ 2.2)

2.1 Introduction to Chapter 2 top

This chapter presents an overview of the various characterizations of loss of high vowel voicing in Japanese.

The organization of this chapter is as follows. §2.2 presents an overview of previous work done on vowel devoicing in Japanese. It begins with the traditional phonological feature-changing formulation of the rule and notes problems with that formulation. It continues with the alternative gestural overlap account of devoicing presented in Jun & Beckman (1993) that is based on Browman & Goldstein (1986, 1990, 1992). The analysis of Tsuchida (1994, 1997) is then reviewed, which contends that the apparent discrepancies between the phonological and gestural accounts of devoicing are best resolved by positing both processes at work in Japanese--a gestural overlap account that causes loss of voicing of non-high vowels and high vowels between fricatives, and a phonological process involving the spread of the feature [+spread glottis] from the preceding voiceless obstruent that causes devoicing of high vowels in other devoicing environments. Evidence from the current data set is also presented supporting and refining this analysis.

§2.3 continues with an aspect of Japanese vowel devoicing that apparently has not yet received proper attention in the literature, the fricativization accompanying devoiced vowels that is prevalent in the current data set and routinely observable in media and public speech. Vowels in the devoicing environment are often fricativized following coronal and velar obstruents, consistent with the report of the ‘obstruentized’ vowels of Uyghur reported on in Hahn (1991) and other fricativized vowels in Standard Chinese (Ladefoged & Maddieson 1990, 1996). In the current data set this fricativization occurs both with voiced and devoiced vowels, implying independent manipulation of the oral closure of the voicing gestures.

§2.4 attempts to clarify the terminology associated with the devoicing and fricativization processes, and §2.5 gives geometric representations of the various changes that vowels in Japanese can undergo. Finally, §2.6 summarizes this chapter.

2.2 High Vowel Devoicing top

The devoicing of high vowels is one of the hallmarks of Japanese in the Kanto region surrounding Tokyo. Its description has a long history in the literature (e.g. Sakuma 1929; Bloch 1950; Martin 1952). According to Maekawa (1989), citing Miyajima (1961), the earliest reference to devoicing is apparently a Latin grammar of Japanese by D. Collado in 1632. This section will discuss the application of the rule of High Vowel Devoicing (HVD), beginning with factors said to affect its application.

2.2.1 Factors affecting the application of HVD top

There are many factors affecting the application of HVD, including vowel height, consonantal environment, position in the word, phrase and utterance, the quality of the vowel in the following mora, the number of consecutive devoicable vowels, pitch accent location, formality of speech, rate of speech, dialectal and ideolectal variation, generational variation, and grammatical structure of the word (Kondo 1997; see §4.5 of that work for an excellent overview). This section will review the factors investigated in this study and the literature that has reported them.

2.2.1.1 Speech Rate top

HVD was characterized as a fast speech rule, as opposed to a casual speech rule, by Hasegawa (1979). That is to say, it was classified as a rule dependent on Speech Rate (SR) and not variations in formality. Previous research has provided evidence for this SR effect (Kuriyagawa & Sawashima 1989, Maekawa 1990, Kondo 1993, Jun & Beckman 1993). In fact, it has been shown that non-high vowels are also devoiced by some speakers at faster speech rates (Maekawa 1990; Kondo 1993).

However, as noted by both Kondo (1997) and Tsuchida (1997), in actuality it is very difficult to categorize HVD as a fast speech rule--Sakuma (1929) maintains that if a word contains only one devoicable vowel, it must be devoiced; Bloch (1950) and Kawakami (1977) both maintain that high vowels in the devoicing environment are voiced only in unnaturally careful speech. In addition, Vance (1987:55) suggests and Beckman (1994:57) states outright that high vowels are devoiced even in the most carefully read speech. Kondo (1994) notes that one speaker in that study did not voice any of the 9 vowels in non-consecutive devoicing sites at the determined normal SR, and only 1 out of 9 vowels at the slow SR. Han (1994: 76) reports that devoicing of [i] and [u] were quite variable within that group of participants as well, with some participants devoicing many vowels at all SRs. The participants in Varden & Sato (1996) and Kondo (1997) also showed loss of high vowel voicing at slow SRs.

While a SR effect was still observed in Varden & Sato (1996) and the data in the current study, these observations and studies suggest that the ‘fast speech’ requirement of HVD is being lost for at least younger speakers of the Tokyo dialect. It was this loss of SR that initially prompted the study being reported on in this work. As will be seen in subsequent chapters, many vowels were devoiced at even the slowest SR utilized in this study.

2.2.1.2 Pitch accent placement top

Perhaps the factor most discussed in the literature is the effect of lexical pitch accent placement on the application of HVD. Traditionally it has been held that accented vowels tend not to be devoiced (McCawley 1977; Haraguchi 1977); accented vowels are either not voiced, or the accent is shifted to another vowel. However, it has been noted that this accent shift appears to be disappearing, especially for younger speakers (Sugito & Hirose 1988; Kondo 1997; Tsuchida 1997: 23-31; Kitahara 1997, 1998; Varden 1997). In particular, in work unfortunately not well known in the west, Sugito (1966, 1969, 1971) notes that the loss of voicing on an accented vowel results in the lost pitch being determined by the level of pitch on the following vowel. (note 2-1) Additionally, perhaps the most authoritative accent dictionary of Standard Japanese, NHK (1985), lists both voiced and devoiced variants for many accented vowels. Further, Tsuchida (1997: 26) claims that the only interaction still seen between accent and loss of voicing for younger speakers is in borrowed vocabulary. Finally, Kitahara (1997, 1998) provides evidence from phrasal pitch contours that when accented high vowels were devoiced, the pitch accent location had indeed been lost and had not been shifted to another mora.

It will be noted that in the present study it also appears that accent and loss of voicing did still interact to some degree--of the 10 participants examined in this study, all but 2 showed at least some variability in pitch accent placement accompanied by variable devoicing of vowels (i.e. voicing the 1st vowel of a token and devoicing the 2nd in one repetition, and then devoicing the 1st vowel and voicing the 2nd in another repetition of the same token). Further research is necessary to determine how much of an influence the experimental setup had on this variability.

Overall the devoicing of accented vowels is in general supported by the current data set, and will be discussed more fully in Ch. 3. The devoicing of accented vowels, together with the alternative strategies available for avoiding devoicing, indicate that there will be no significant effect of pitch accent placement on vowel devoicing. Effects of pitch accent placement on the devoicing of vowels in a later chapter will therefore not be checked.

2.2.1.3 Sociological factors top

It would appear that sociological factors are now exerting the greater influence on the application of the rule than either SR or accent placement. Beckman (1994: 66) notes both the spreading of devoicing in the ‘prestige’ dialect of Tokyo, and that purposeful voicing of sentence-final vowels (i.e. suppression of devoicing) in the Osaka region seems to be used to exclude speakers from the Tokyo region. Imaizumi et al. (1995) notes a decrease in the number of vowels devoiced when teachers are interacting with hearing-impaired children then when they are interacting with normal-hearing children. Perhaps surprisingly, Yuen & Hubbard (1997) found that sociological factors (gender, speaking style, etc.) played a greater role in the number of vowels devoiced by the participants in that study than either segmental environment or pitch accent placement. And it is the impression of the author that sentence-finally the [u] of the copula desu and -masu (a non-past verb ending) are voiced only in situations when an impression of formality or respect is desired (e.g. during newscasts, or when talking to a those of higher social rank or strangers). It would appear that sentence-final devoicing of these vowels, at least for many speakers, should now be considered ubiquitous, with its suppression to indicate formality or respect being the marked case.

2.2.1.4 Moraic position of the vowel top

Another factor affecting the application of HVD that has been noted in the literature is that vowels in the 1st mora of two-mora words tend to devoice more than the vowels in the 2nd mora (Kuriyagawa & Sawashima 1989). This finding was supported by the current data set, although the effect was not consistent for all tokens or at all 3 SRs utilized in this study. §6.4.1 will discuss this further.

2.2.1.5 Consonantal environment top

The consonantal environment that the devoicable vowel is found in can also have an effect on its application. Han (1962b) suggests that a preceding fricative has a greater devoicing influence than an affricate, which has a greater influence than a stop. According to Takeda & Kuwabara (1987), high vowels are more likely to devoice after fricatives and affricates than after plosives. Simada et al. (1991) found that some speakers tend to devoice vowels after plosives, while others tend to delete them. Consonantal effects observed in this study will be noted in following discussion.

Other factors affecting the application of HVD such as the number of devoicable high vowels in a row and focus accent will not be discussed in this work. However, see Kondo (1994, 1997) for extensive discussion of devoicing in consecutive devoicing environments, see also Vance (1987) Ch. 6 for a more complete overview of factors affecting devoicing.

The various formulations of HVD that have been proposed in the literature will now be reviewed.

2.2.2 HVD as a feature-changing rule top

HVD has been traditionally characterized as a phonological (i.e. feature-changing) rule (e.g. McCawley 1968: 127) which could be formulated as follows.

Figure 2.1 A traditional formulation of the rule of HVD.

That is to say, high vowels devoice between voiceless obstruents, or after a voiceless obstruent and before a pause.

Devoicing before a pause was included in this rule primarily to account for the devoicing of the sentence-final [u] of desu (the copula) and -masu (a formal verb ending) that was discussed above. This sentence-final devoicing will not be discussed in this thesis due to its almost universal application at all SRs and levels of formality, except, as mentioned before, when the speaker wishes to present a formal impression by consciously suppressing devoicing. (note 2-2)

It has long been recognized that HVD can occur at slower SRs and in more formal speech (e.g. Sakuma 1929; Bloch 1950), indicating that it is a phonological process: if a change in pronunciation occurs at a SR slow enough for there to have been sufficient time for any existing articulatory gestures to be realized, (note 2-3) then the change must have happened before the articulatory instructions were issued--there must have been a phonological change. Examples occurring at the slowest SRs will be presented below that show exactly this situation.

Looking at the rule above, and as discussed in other work (e.g. Jun & Beckman 1993), the rule as formulated here predicts that only two types of vowels will be produced: 1) vowels specified for [+voice] and hence produced with full voicing; and 2) vowels specified for [-voice] and hence produced without any voicing (i.e. whispered vowels). The rule as formulated above implies that no other aspect of the vowel production will change other than the normal adjustments in temporal duration due to speech rate (SR) or prosodic changes.

However, the characterization of devoiced vowels as simply lacking vocal cord vibration suggested by Figure 2.1 above is not accurate in at least two respects.

First, as noted in Jun & Beckman (1993), the rule as stated above does not account for the fact that the amount of voicing evidenced during vowel production is gradient. The traditional formulation of the rule predicts either fully voiced vowels or fully devoiced vowels. However, the voicing of a vowel does not begin at full intensity or end with a sudden decrease to zero; the intensity of the glottal vibrations increases over a period time, typically maintains some plateau, and then decreases over a period of time. This gradiency increases at faster SRs for high vowels between voiceless obstruents (Beckman 1994). Globally, the duration of voicing that can be found in multiple repetitions of a given vowel ranges from the full duration of the vowel to the slightest amount of voicing. As little as one cyclic glottal pulse at a frequency appropriate for the voicing of a speaker has been observed in the data in this study and elsewhere (Beckman & Shoji 1988; Kondo 1993). The range of voicing durations that can be found in the production of vowels at a given SR would seem to be too great to be due to adjustment of segment, mora, and word durations because of SR and stylistic variation (for discussion of temporal adjustments, see Kawasaki 1983; Port et al. 1987; Campbell 1992; Han 1994; Pirello et al. 1997; Farnetani 1997).

Second, the rule as stated above does not reflect the large-scale articulatory changes of the devoiced vowels routinely observable in public speech and found in this study. The devoiced vowels found in this study are not whispered vowels; their mode of production is quite distinct from vowels produced during normal whispering. Vowels produced during whispering exhibit no appreciable closure of the oral tract. The air rushing through the spread glottis provides the energy for the resonance of the oral tract that is associated with the production of the vowel (Catford 1977: 96; Ladefoged 1996: 107-108), with the clarity of the whispering often being strengthened apparently by the glottal tension utilized for ‘breathy voice’. In contrast, the closure associated with the production of vowels undergoing HVD is that of a fricative, a fricative closure with enough oral tract vowel configuration maintained to provide the perception of a vowel: frication caused by the air rushing out of this newly-formed closure provides the energy for the resonance of the oral cavity that results in formant structure formation.

In addition to the articulatory changes taking place in the oral cavity, Tsuchida (1994, 1997) notes that the glottal aperture of these vowels matches that of a fricative, not that of a vowel. There is a significantly larger glottal opening associated with both true fricatives and devoiced vowels attributed to the assignment of the feature [+spread glottis] to both; in the case of fricatives, underlyingly; in the case of devoiced vowels, by phonological rule. Indeed, it seems that Tsuchida (1997) comes quite close to calling the devoiced vowels seen in that study fricatives.

For the two reasons given above, the characterization of HVD as a spread of the feature [voice] is therefore not consistent with the observed data. The fricativization mentioned above will be discussed further in §2.3. An answer to the first of these criticisms, that a feature-changing rule cannot account for the observed gradiency of voicing, is found in Jun & Beckman (1993).

2.2.3 An Alternative account: Gestural overlap top

As noted in Jun & Beckman (1993), the Gestural Score framework proposed in Browman & Goldstein (1990, 1992) provides an explanation for the gradiency of voicing duration observed both locally in one vowel production and globally across repetitions. In the Gestural Score account, all vowels are phonologically specified as [+voice]; devoicing of vowels is due to the instructions for voicing being overlapped by the preceding or following instructions to produce a voiceless consonant. This overlapping of articulatory instructions is shown in Figure 2.2 below for the word kiki ‘crisis’. (The spreading symbols represent instructions to the glottis to inhibit or stop voicing, and the closing symbols represent instructions to the glottis to initiate voicing. They are not meant to represent actual vocal cord movement or degree of glottal spread or closure.)

Figure 2.2 Gradient vowel devoicing in the Gestural Score framework.

At slower SRs, represented to the left of the figure, there is adequate time for the voicing gestures associated with a vowel to be fully initiated and sustained for the full duration of the vowel. At medial SRs, represented in the middle of the figure, although there is sufficient time to initiate voicing, the previous and following devoicing gestures for the unvoiced stops (i.e. either active glottal abduction or suppression of adduction) cause the vowel to be devoiced for only a portion of its duration. At progressively faster and faster SRs, there is less and less time for voicing to be initiated and sustained, until finally at some sufficiently fast enough SR there is no time to initiate voicing. At this point the devoicing gestures completely overlap the voicing gesture in the temporal domain, and the vowel becomes completely devoiced.

This characterization of devoicing as a gestural overlap process supports other studies (e.g. Maekawa 1990) which have shown devoicing of both high and non-high vowels to be gradient in the sense that widely varying durations of voicing activity associated with vowels have been observed. If devoicing were strictly phonological in nature, one would not expect to find such a wide range of voicing duration values for a given SR--vowels should either be voiced, with voicing lasting nearly all the duration of the formant activity of the vowel, or they should be devoiced, with no trace of voicing observed since they are surrounded by voiceless consonants. In actuality, a wide range of voicing duration values for vowels in similar consonantal contexts at the same SR have been observed. Examples of this range of voicing duration values will be provided from the current data set as well.

Imaizumi et al. (1995: 775~776), in their discussion of the gestural overlap account of devoicing presented in Jun & Beckman (1993), delimit three ways that overlapping of the vowel voicing gesture can be achieved: 1) the devoicing gestures of the surrounding voiceless obstruents can be temporally shifted toward the vowel due to an increase in SR (note 2-4); 2) the surrounding devoicing gestures can be lengthened (note 2-5); or 3) a combination of the two. The 3 strategies are represented graphically below (after Imaizumi et al. 1995: 775, Figure 4).

Figure 2.3 Three strategies for achieving overlap of vowel voicing gestures by surrounding devoicing gestures

While Imaizumi et al. (1995) does not provide evidence for which strategy was employed by the participants in their study, evidence will be presented below that both overlapping and strengthening was utilized by the participants in this study.

For much of the devoicing observed in this study a temporal shift in the location of the frication associated with the preceding obstruent can be seen; it appears that a temporal shift in the centering of the frication--and hence the glottal spreading responsible for the frication--was responsible for the devoicing of the vowel. Further, it will be seen that only the preceding obstruent’s glottal spreading gesture appeared to be shifted; the frication associated with the following obstruent remained fixed with respect to the succeeding vowel within token productions of similar duration. Therefore the devoicing of the vowel will be attributed to the glottal spread of the preceding voiceless obstruent, with the glottal spread associated with the following voiceless obstruent not being actively involved in the devoicing.

At least one participant in this study appears to have utilized the second strategy mentioned above; in her productions spectrographic evidence points to a lengthening of the glottal spreading gesture. Indeed, in some productions by some participants the frication associated with the token-initial obstruent is seen to continue through to the 2nd mora, with no clear medial stop closure or release being made. These cases will be discussed further below, with representative productions being given.

However, characterizing devoicing as a phonetic overlap of glottal spreading gestures also fails to account for the most basic of observations regarding HVD: vowels often devoice even when the SR must be slow enough to allow full achievement of articulatory targets.

The fundamental problem with extending the gestural overlap account of devoicing depicted in Figure 2.2 above to all instances of HVD is that it predicts that high vowels will devoice only when the SR is fast enough for the devoicing gestures to completely overlap the voicing gestures. As noted in Tsuchida (1997), this characterization of HVD is not tenable. HVD occurs even at slow SRs where there must be more than sufficient time for the voicing gesture to be realized. Devoicing at slow SRs has been noted in the literature (Kondo 1993, 1994, 1997; Varden & Sato 1996) and can be seen in the current data set as well. The fact that a vowel devoices even at slow SRs indicates that the voicing gesture must not have been present at the phonetic level--that the instructions to voice the vowel were changed before the vowel was pronounced.

There would seem to be one possible analysis involving gestural overlap that would allow devoicing to occur at even slow SRs. This would be an analysis where the glottal spreading gestures of the surrounding voiceless obstruents could be freely realigned within the temporal domain of the segments production. If the glottal spread associated with a voiceless obstruent were allowed to be centered at any point of time in the production of the obstruent, it could also be centered toward the end of a obstruent’s production. This would result in the glottal spread continuing into the following vowel. An overlap of the vowel’s voicing gesture could occur, and the vowel could be (at least partially) devoiced. In another repetition of the obstruent and vowel, the obstruent’s glottal spreading gesture could be centered more toward the beginning of the obstruent’s production. This could result in the glottal spreading gesture being completed when the production of the following vowel began. No overlap of the vowel’s voicing gesture would occur; the vowel would not be devoiced.

However, based on other studies (Kim 1970; Kingston 1990; Kingston & Diehl 1994) and the data that will be presented below, attributing devoicing to the free alignment of the glottal spreading gesture does not seem tenable. The data in this study shows that the alignment of the glottal spreading gesture associated with a voiceless fricative or affricate, as indicated by the location of the frication associated with that obstruent, is not found in a range of temporal locations as would be expected if alignment of the glottal spreading gesture were free. The location of the frication associated with a consonant in the current data set is fixed. In addition, no reports of free alignment of the frication associated with a segment’s glottal spreading gesture appear to have been made in any of the literature involving glottal spreading. When a voiceless fricative or affricate precedes a voiced vowel, the frication is centered on the fricative component of the consonant. When a voiceless fricative or affricate precedes a voiceless vowel, the frication is centered midway between the consonant and the vowel site. This is consistent with both the data and analysis of Kingston (1990), Iverson & Salmons (1995), and Tsuchida (1997). The temporal shift of the glottal spreading gesture is therefore thought to be due to a phonological process that will be discussed below, not free alignment of the spreading gesture.

There is one other point which argues against free temporal alignment of the glottal spreading gestures. If there were great enough freedom in aligning the spreading gesture with the rest of the segmental articulation so that shifting the spreading gesture’s point of alignment caused the devoicing at slower SRs, the strong prediction is made that all vowels--high, mid and low--would be subject to this type of devoicing at slower SRs. There seems to be no phonetic reason to allow the temporal alignment of the glottal spreading gesture to be freer for high vowels than for low vowels. (note 2-6) The fact that non-high vowels do not devoice at slower SRs as high vowels do argues against free temporal realignment of the devoicing gesture.

Analyzing devoicing as a phonological process does provide a ready answer as to why generally only high vowels devoice, however--historical reassessment of a phonetic phenomenon as a phonological one. The close oral closure accompanying high vowels makes it much more likely for air flow to drop below the minimum requirement for voicing to be sustained (Locke 1979); in Japanese and several other languages, this apparently has led to the grammaticalization of fricativization/devoicing process for high vowels.

Disallowing a phonetic account of HVD at slow SRs leaves only a phonological account; namely, that some sort of feature adjustment is responsible for the devoicing observed at slow SRs. However, the fact that non-high vowels devoice more at faster SRs (Maekawa 1990) and that gradiency of voicing is seen both within individual productions and across many productions indicates that a phonetic overlap of the voicing gesture is also present.

The apparent discrepancy in the data is reconciled in Tsuchida (1997), who proposed that both phonetic and phonological devoicing are present in Japanese.

2.2.4 HVD as both phonetic and phonological top

Tsuchida (1997) provides particularly strong experimental evidence for the phonological devoicing of high vowels between plosives. In that study, electromyographical (EMG) measurements of muscle activity and fiberscopic observation of the accompanying glottal spreading were made during the production of both voiced and devoiced vowels. Tsuchida found that when high vowels devoiced between plosives, there was only one associated large spike of activity of the laryngeal muscle involved in glottal spread, the posterior crycoarytnoid (PCA) muscle, and only one large accompanying glottal spreading gesture. These findings supported Sawashima (1971), which also found one PCA activation accompanied by one large glottal spread when high vowels devoiced between plosives.

In contrast, the participant in the Tsuchida (1997) study generally did not devoice high vowels between fricatives. When high vowels were preceded and followed by voiceless fricatives, two clear activations of the PCA and two clear glottal spreading gestures were observed, one for each consonant preceding and following the vowel. (The same two activations of the PCA and two accompanying glottal spreading gestures were also found for non-high vowels in all devoicing environments.)

Tsuchida (1997) therefore argued for both phonological and phonetic devoicing in Japanese: in the case of high vowels between voiceless plosives, a phonological spread of the feature [+spread glottis] from the preceding voiceless obstruent; in the case of high vowels between voiceless fricatives and non-high vowels in all devoicing environments, a phonetic overlap that causes loss of voicing and devoicing at faster SRs. Indeed, this proposal was anticipated as far back as Bloch (1950: 136), who stated that while "[voiceless [i] and [u] are] paralleled, especially in slow or careful speech, by an otherwise identical synonymous phrase containing [voiced [i] or [u]] instead", voiceless [a] and [o] occur "in rapid speech only". And as discussed in Kondo (1997), the results of Maekawa’s (1990) studies into devoicing of /u/ vs. devoicing of /a/ and intervocalic voicing of /h/ also implied two processes at work: one influenced by SR and the other independent of it.

As for the mechanism responsible for devoicing, Beckman (1994: 58-59) states that devoicing can be attributed to closeness of the oral constriction. When there is a close oral constriction maintained at the release of the preceding consonant, as there is for fricatives and affricates, the air pressure that has built up in the oral tract due to the closure cannot be vented as quickly when the segment is released as if the oral tract were relatively open. This maintenance of the air pressure in the oral tract above the larynx makes it more difficult for voicing to be initiated, since there must be a minimal pressure difference across the larynx for voicing to be able to take place (Lieberman & Blumstein 1988: 101-103). When the vowel following a fricative or affricate is a high vowel, itself made with greater closure than mid or low vowels, the effect of closure blocking the venting of air pressure will compound, and the likelihood that the realization of the voicing gesture will be completely blocked increases. This leads to the prediction that high vowels will devoice more readily after fricatives and affricates, a prediction born out by Takeda & Kuwabara (1987) and Kondo (1993) (as cited by Beckman 1994). Indeed, Kondo (1994, 1997) has found good evidence of vowels in environments that are adjacent to other vowel devoicing environments (i.e. consecutive devoicing environments) undergoing reduction in intensity even when they are not devoiced. This suggests that devoicing or deletion of vowels in these consecutive devoicing contexts can be seen as an end result of an overall vowel reduction process (see Kondo 1997 Ch. 8 for discussion).

Taken together with the glottal spreading overlap analysis of Jun & Beckman (1993), the prediction is that adjustment of two parameters can be responsible for devoicing of vowels: 1) adjustments to the glottal spreading associated with an adjacent voiceless obstruent; and 2) adjustments to the oral closure of the vowel due to an adjacent fricative or affricate, whether they are phonemically or allophonically derived. Further work is needed to completely delimit the two processes.

Another analyses involving both phonological and phonetic derivation of the same acoustic end in a given language is the proposal made for English by Hayes (1992) in a comment on Nolan (1992). The same argument is made for English /s/ to /sh/ assimilation in Holst & Nolan (1995).

Nolan (1992), using palatalographic measurement of productions of the English string late calls, noted that many speakers produce a complex coronal/velar articulation for the medial [t k] cluster. However, the coronal articulation of the complex segment is weakened to varying degree, while the velar articulation is not; the velar component is ‘robust’. Hayes (1992) accounted for this by positing both a binary phonological process and a gradient phonetic one: the phonological process spreads the velar place of articulation forward to the coronal to produce the complex segment, while a phonetic process weakens the coronal portion of the complex segment to varying degree--the gradiency of the weakening is the result of the process being due to overlap of gestures. (note 2-7)

Holst & Nolan (1995) provided further evidence for both processes in the production of /s/ and /sh/ clusters both within and across a syntactic boundary. Their data indicated that within syntactic boundaries, /s/ fully assimilated to a following /sh/, indicating phonological assimilation due to feature spreading, while across syntactic boundaries the assimilation was gradient, indicating phonetic overlap of gestures.

In regard to Japanese, Jun & Beckman (1993) and Varden & Sato (1996) have reported gradient amounts of voicing associated with high vowels that are not between voiceless fricatives. This gradiency of voicing durations at a given SR is observed in the current data set as well.

Returning now to the phonological devoicing posited by Tsuchida (1997), recall that the temporal shift of the glottal spreading gesture of a consonant preceding a devoiced vowel is due not to overlap or free alignment of glottal spreading gestures, but instead is due to a phonological spread of [+spread glottis] from the obstruent to the vowel. This double-linking of the feature then leads to its temporal realignment to the midpoint between the obstruent and the vowel. This realignment of the frication associated with fricatives and affricates can be seen in the current data set. A representative pair of productions is given in the figure below.

Figure 2.4 Waveforms and spectrograms for voiced and devoiced 1st vowel of tsuki ‘moon’, participant HK. HK 1s tsuki

The 1st vowel of the second production is devoiced, even though the 556 ms token duration is toward the long end of all productions and indicates a fairly slow SR (this participant’s productions averaged 554 ms for the slow repetitions, 383 ms for the normal repetitions, and 309 ms for the fast). In contrast, the first production of the same token with a token duration of 558 ms shows 78 ms of voicing (as measured from the filtered waveform) associated with the 1st vowel.

Notice that the various components of the two signals are temporally fairly closely aligned: the weak release of the initial affricate [ ts ] as evidenced by the vertical striation to the left of the spectrograms, (note 2-8) the onset of the closure of the medial stop [k], its release, the amount of aspiration of the second stop (i.e. the VOT), the onset of the second vowel’s voicing, and the onset of the closure of the following clitic’s initial stop [t] are all fairly well temporally aligned in the two productions.

The developed formant structure on the frication of the first vowel in the first production and the weak formant structure on the first vowel in the second production shows that there was a tongue-shaping gesture at the vowel site of both productions. This indicates that the tongue-shaping gesture is independent of voicing of the vowel. The equivalent duration of the 1st mora in both productions also speaks to the retention of the vowel’s timing slot, eliminating an analysis where the vowel is deleted after a coloring of the frication of the obstruent for this and similar productions.

The VOT associated with the 2nd mora vowel is approximately the same in both productions, indicating that the glottal spreading gesture associated with the release of the medial stop [k] is of the same duration and has the same temporal alignment in both. As noted earlier, this is taken to indicate that the preceding consonant (i.e. the 1st mora consonant) is responsible for devoicing, and that the glottal spreading gesture associated with a following consonant is not involved in devoicing of a vowel that precedes it.

Turning now to the frication associated with the 1st mora fricative, it can be seen that the duration of the frication is approximately the same in both productions--approximately 120 ms and 110 ms, respectively. This indicates that the glottal spreading gestures were also of about the same duration, arguing against an analysis involving a lengthening of the glottal spreading gesture in this case, whether or not it is strengthened.

Instead, note that the peak of the frication of the affricate in the devoiced vowel production is temporally shifted when compared to the location of the peak frication in the voiced vowel production. This is consistent with what Kingston (1990: 427) noted in discussion of s-voiceless stop clusters in languages like English (e.g. the initial cluster in stop). It was observed that glottal spreading in these clusters reached its maximal width at a point that "…is close to the boundary between the two oral articulations, a temporal compromise between the early peak [glottal spread] of the fricative and the late peak [glottal spread] of the stop." This is in contrast with the location of the maximal spread for voiceless fricatives and plosives when they are not found in clusters--in the case of voiceless fricatives, the maximal spread is usually aligned with the middle of the segment; for voiceless aspirate stops, it is aligned with the release of the closure (Browman & Goldstein 1986; Kingston 1990, Goldstein 1990; Tsuchida 1997).

This midpoint location of the shared spreading gesture also led Iverson & Salmons (1995) to posit the sharing of the feature [+spread glottis] in voiceless fricative/stop clusters in English, similar to the spread of the same feature posited in Tsuchida (1997). Iverson & Salmons (1995) posit an underlyingly marked specification of [+spread glottis] for voiceless stops in English, rather than the redundant specification of [-voice] that is usually assumed for this class of segments. (note 2-9) Since both voiceless fricatives and voiceless stops are then specified for this feature, the OCP (McCarthy 1987) forces the dual specification of this feature to coalesce in this cluster-initial environment. The temporal alignment of the shared feature is thought to be triggered by the feature sharing and determined by the two segments’ midpoint.

As mentioned in §1.4.3, in this analysis the degree of observed aspiration in English is attributed to the independent prosodic strengthening of segments caused by English stress assignment (Nespor & Vogel 1986; Halle & Vergnaud 1987; Kingston & Diehl 1994). Japanese aspiration generally patterns that of English (Homma 1980; 1981) and can be plainly seen in the current data set; the differences between aspiration in English and Japanese can be attributed to different prosodic strengthening strategies in the two languages.

Independently using observation of glottal spreading and activation of the muscles causing it, Tsuchida (1997) posited devoicing of high vowels between voiceless fricatives as a spread of the feature [+spread glottis] from the voiceless fricative to the vowel. The sharing of the feature would then trigger a realignment of the spreading gesture to midway between the two segments. This realignment was evidenced in that study in both the plots of muscular activity and in the plots of the resulting glottal spread, and was held to be responsible for the devoicing of high vowels in this environment.

In light of the discussion of Iverson & Salmons (1995) above, a slight revision of the laryngeal system given in Tsuchida (1997: 53, Table 2.7) is required; here voiceless stops have been merged with voiceless fricatives and affricates, and /h/.

Table 2.1 Modified laryngeal classification system for Japanese.

Let it be noted that an attempt to account for the observed glottal spreading in many studies (e.g. Tsuchida 1997) has not yet been attempted.

Returning to the spectrograms in Figure 2.4 above, it can also be seen that the frication associated with the mora containing the voiced vowel is centered between the release of the consonant and the onset of the vowel; i.e. the frication is centered on the latter half of the initial affricate, the [s] portion of the affricate. This same alignment is seen in all [tsu] morae containing a voiced vowel in this study. In contrast, the peak of the frication associated with the devoiced vowels can be seen to be aligned within the vowel site, consistent with the phonological coalescence of the initial consonant’s and vowel’s specification for [+spread glottis] posited by both Tsuchida (1997) and Iverson & Salmons (1995).

In addition to the devoicing at slower SRs, alternate voicing and devoicing of vowels at the same SRs can be seen in the following two repetitions of chiki ‘friend(s)’ by participant ANa. In the first production the first vowel has been devoiced, whereas in the second, the second vowel has been devoiced. (The 1st panel gives the waveform; the 2nd, the pitch trace; and the 3rd, the spectrogram.)

Figure 2.5 Alternate devoicing of vowels at fixed SR, 1st vowel devoiced (participant ANa).

Figure 2.6 Alternate devoicing of vowels at fixed SR, 2nd vowel devoiced (participant ANa).

Note again the close alignment of the formant structure of the first vowel and the stop closure and releases in both productions. Again, the large difference in the two productions, aside from fo activity, is in the duration and centering of the frication on the first vowel, and the duration of frication on the second vowel. The approximately equal durations of both morae argue against an overlapping gesture analysis based on SR; there seems to be no plausible explanation why voicing of the vowel in each mora would have time to be initiated and sustained in one production but not the other. A phonological analysis again provides a ready answer, however--in the first production, the participant simply chose to accent the second vowel and devoice the first; in the second production, she chose to accent the first vowel and devoice the second. (note 2-10)

The characteristics of the frication in both devoiced vowels are also quite consistent with a devoicing analysis utilizing the spread of the feature [+spread glottis]. In the first production, again the duration and temporal location of the frication are consistent with the sharing of the feature spread from the voiceless fricative. The VOT observed in the second production, as with the tsuki example above, provides the same analysis for the devoicing of the second mora as well. The longer frication observed is entirely consistent with the spread of the feature [+spread glottis] from the aspirate stop to the vowel, again resulting in a shared specification for the feature that is temporally aligned midway between the two segments; i.e. substantially after the release of the stop.

However, it must be noted that temporal shift of the glottal spreading gesture is not observed for all productions for all participants. In particular, participant TO showed much less frication associated with her productions, and it appears that the period of frication lengthened rather than was shifted.

The following figures contrast two productions of the token tsuki ‘moon’ by participant TO.

Figure 2.7 Waveform and spectrogram for voiced 1st vowel of tsuki ‘moon’, participant TO.

Figure 2.8 Waveform and spectrogram for devoiced 1st vowel of tsuki ‘moon’, participant TO.

Again, the first vowel of the second production is devoiced, even though the 499 ms token duration is toward the long end of all productions and indicates a fairly low SR (this participant’s productions averaged 485 ms for the slow repetitions, 434 ms for the normal, and 319 ms for the fast). In contrast, the first production of the same token with a token duration of 501 ms shows 77 ms of voicing (as measured from the filtered waveform) associated with the first vowel.

However, notice that the frication associated with the initial obstruent of the 1st production is not shifted in the 2nd production. Instead, it is lengthened so that it continues from shortly after the release of the affricate on through the vowel site. This indicates the glottal spreading gesture associated with the obstruent is also lengthened in duration. It is thought that the cause of this lengthening is phonological in nature; this will be pursued in the next section.

It would appear from this and other similar productions from this participant that the individual strategies used to align the glottal spreading gesture show variation both across speakers and, as will be seen below in the next section, for different obstruents. For some speakers, the glottal spreading gesture associated with the preceding obstruent is temporally shifted, realigning at the midpoint between the obstruent and the vowel. For others, the glottal spreading gesture of associated with the preceding obstruent is lengthened, lasting the duration of the obstruent’s and the vowel’s production. And as noted above in the discussion of Imaizumi et al. (1995) in §2.2.3 of this work, some productions by some participants displayed frication lasting throughout the entire token, with no clear medial stop being produced. Representative productions of this will also be given in the next section.

This chapter will therefore adopt Tsuchida (1997), that vowel devoicing is caused by the spreading of the feature [glottal spread] from a preceding obstruent to a following vowel and subsequent realignment of the spreading gesture to the midpoint between the obstruent and the vowel. However, it has been noted that realignment of the gesture does not always occur; some speakers evidently utilize a lengthening strategy that achieves the same end.

Further, it was noted that phonetic overlap is active during the production of high vowels in all environments. While the overlap can not be responsible for devoicing at slower SRs, it is responsible for the gradient amounts of voicing durations observed, and may be responsible for devoicing at faster SRs. This would be consistent with the increased number of devoiced vowels seen at the higher SRs, as noted in the previous chapter.

In addition to these production arguments, statistical evidence from the current data set will be presented in Ch. 6 that shows quite clearly the data being separated into two distinct groups. The immediate conclusion is that the two groups of data were generated by two separate processes; i.e. phonetic overlap of voicing instructions and a phonological deletion of the voicing instructions.

2.2.5 Summary of arguments for both processes top

The following four considerations were discussed above:

Based on these considerations, it is concluded that both processes affecting voicing are required to account for the data observed in this and other studies. In summary, this work follows Tsuchida (1997) in claiming two loss of voicing processes at work in Japanese:

Including a phonetic reduction of voicing durations in the analysis of the data for high vowels as well as non-high vowels leads to the prediction that there will be high vowels that will lose all their voicing due to overlapping of glottal gestures at high SRs. Indeed, this may account for the statistically significant increase in the number of devoiced vowels seen in the previous chapter in the discussion of the effect of SR on whether or not a vowel is devoiced.

Chapter 2 notes (§§2.1 ~ 2.2) top

1. Thanks to Tsutomu Sato for bringing Professor Sugito’s earlier papers to my attention.

2. Reflective of the wide-spread application of sentence-final devoicing of [u] is the fact that 2 of the 3 Taiwanese students learning Japanese reported on in Varden & Sato (1996) devoiced all sentence-final [u] in that study; the remaining participant retained voicing on only a portion of them.

3. The term ‘sufficient time’ refers to the minimal time required for the articulatory trajectories of a given articulation to be realized; see Löfqvist (1997 §4.3) for discussion.

4. Imaizumi et al. 1995 attributes the temporal shift of the devoicing gesture to an increase in SR. This implies that the devoicing gesture is more robust than the voicing gesture of the vowel since it is the voicing gesture that is overlapped and not vice versa. This point will be accepted here as well.

5. Imaizumi et al. (1995) use the term ‘strengthened’, but the net effect of strengthening a devoicing gesture is an increase in duration of the gesture.

6. Kate Davis (p.c.) has pointed out to me that the inherent lower pitch of mid and low vowels is manifest by varying degrees of vocal fold tension, and this varying tension may allow greater freedom in alignment. I am unaware of any studies indicating this has been observed or argued against.

7. But see Browman (1995) for reinterpretation of their data as only phonetic overlap.

8. For the 2nd repetition, the 2nd of the token-initial closure releases was used in taking the token duration measurement.

9. Aspirate voiceless stops are then the default case in Iverson’s proposal, with aspiration being enhanced according to prosodic position. This is quite reasonable since it is the VOT, largely concommitant with the strength of aspiration, that is used to distinguish voiced and unvoiced stops in English (Lisker & Abramson 1964).

10. It appears that, for this participant, pitch accent placement for the token chiki is not completely fixed, possibly due to areal influences from adjacent dialects and/or the infrequent use of this word--the commonly-used word for ‘friend(s)’ is tomodachi. In 2 of the 18 repetitions she accented the 1st mora; the 1st mora vowel was voiced and the 2nd mora vowel was devoiced. In 15 of the 18 repetitions she accented the 2nd mora; this vowel was voiced, and the 1st vowel was devoiced. In 1 repetition both vowel were voiced. There is no difference in the meaning of the productions due to different placement of the accent other than the participant did not devoice the accented vowels.