This book is an anthology comprising fourteen original articles, which focus on various aspects of a rich repertoire of voicing phenomena in Japanese.Footnote *

Their approaches vary, ranging over descriptive, theoretical, engineering and experimental strands. Synchronic and diachronic perspectives are also incorporated. The phonology of voicing in Japanese has been one of the most fertile sources of empirical data in phonological research. However, the diverse phenomena discussed in this book indicate that a satisfactory understanding of this intensively explored area has not yet been reached. This book will be an invaluable resource to scholars of Japanese phonology in particular, as well as to those working on the phonology of voicing in general. The book consists of two parts: ‘Consonant voice’ and ‘Vowel voice’, preceded by a short overview by the editors. Some of the papers evolved from the research presented at the Linguistics and Phonetics 2002 conference held at Meikai University, Urayasu in September 2002.

Part I contains ten articles. It begins with Haruo Kubozono's foundational paper on rendaku (also known as sequential voicing). This is a morphophonological effect that voices the initial consonant of a second compound member (e.g. /mizu-tori/→[mizu-dori] ‘waterfowl’). He starts by reviewing earlier discussion that rendaku is partially productive but partially lexicalised. He subsequently lists linguistic conditions on its occurrence. Famous ones are Lyman's Law (no multiple voiced obstruents in a native morpheme) and the Right Branching Condition (rendaku is prevented in a second member of a right branching compound) (Otsu Reference Otsu, Otsu and Farmer1980, Itô & Mester 1986). Much less well known is a morpheme-specific length restriction. Comparing cases like [e-hoN] ‘picture book’ and [maNga-boN] ‘comic book’, /hoN/ ‘book’ undergoes rendaku only when the entire compound is larger than four moras (Ohno Reference Ohno, Nakayama, and 2000). Kubozono attempts to justify this view by describing several cases where quadrimoraic lengths function as a prosodic or morphological unit. This chapter will be useful for those unfamiliar with the rendaku phenomenon.

Keren Rice investigates a long-standing puzzle. Lyman's Law militates against more than one voiced obstruent in a morpheme, but /huN-ʃibaru/surfaces as [huN-zibaru] ‘immobilise’, with two apparently voiced obstruents. She solves this puzzle by proposing that a heteromorphemic NC cluster shares a single SV (sonorant/spontaneous voice), whereas a postnasal obstruent lexically bears LV (laryngeal voice) if the two consonants are tautomorphemic. She further proposes that Lyman's Law is sensitive to LV, but insensitive to SV. As a desirable consequence, tautomorphemic NC clusters properly prevent rendaku (e.g. [ʃirooto-kaNgae] ‘layman's idea’). Rice goes on to question postnasal voicing as a phonological cue for lexical stratification (Itô & Mester Reference Itô, Mester and Goldsmith1995a, Reference Itô, Mester, Beckman, Dickey and Urbanczykb, Reference Itô, Mester, Padgett, Ussishkin, Herrick, Kurisu and Sanders1999). Her analysis implies the absence of voicing alternations in tautomorphemic NC clusters. She argues that *N does not constitute a phonological criterion for distinguishing Yamato and Sino-Japanese classes (cf. Itô et al. Reference Itô, Mester and Tsujimura1999). One residual question is how Rice's analysis handles cases such as nominal [hʲootaN-kago] ‘gourd basket’. On her proposal that a heteromorphemic NC cluster has a single SV feature, incorrect [hʲootaN-gago] would respect Lyman's Law. Rice gives this example, but she does not explain why /k/ remains voiceless. To the best of my knowledge, [huN-zibaru] is the only verbal compound that disobeys Lyman's Law, but there are many nominal compounds like [hʲootaN-kago].

Kazutoshi Ohno summarises how the voicing contrast in obstruents came into existence in the orthography of contemporary Japanese. The voicing distinction was originally made in the eighth century. This distinction was lost with the advent of kana orthographies, but it was revived in Modern Japanese with the introduction of a diacritic to denote voicing. Ohno considers two possible connections between Japanese orthography and phonology: (i) the orthographic systems reflect phonological voicing (non-)distinctiveness of obstruents, or (ii) they are orthographic conventions proper. The second position can be further subdivided, depending on whether voicing has been phonologically distinctive or not, independently of orthographies. Ohno does not offer his own view, so the discussion remains inconclusive. He suggests that nasality should be scrutinised to uncover phonological voicing contrast in the history of Japanese.

Kuniya Nasukawa studies how voicing features should be represented. Japanese has a two-way contrast of voiced vs. voiceless. Aspiration does not exist. Framing his discussion in Element Theory (Kaye et al. Reference Kaye, Lowenstamm and Vergnaud1990), he refutes the basic dichotomy of [L] (voiced) vs. [H] (voiceless), where [L] and [H] are associated with low and high F0 values of voiced and voiceless segments respectively. Based on the observation that voiced obstruents and nasals partially pattern alike in Japanese phonology (i.e. nasals should be specified for [voice] for the purpose of postnasal voicing), he labels both voiced obstruents and nasals as [N]. They are partially different too (i.e. voiced obstruents block rendaku, but nasals do not). Nasukawa proposes that voiced obstruents and nasals are specified as headed [N] (voiced obstruents) and bare [N] (nasals) respectively. Voiceless obstruents are not marked at all, since features are monovalent in Element Theory. Some cross-linguistic insights are cited in support of this conclusion, but the repercussions of this element-based representation on Japanese phonology overall are not fully discussed.

Timothy J. Vance presents a statistical study of rendaku in compounds consisting of inflected items (i.e. verbs and adjectives). On the one hand, verbal compounds are more likely to exhibit rendaku when a whole compound is a noun than otherwise. This contrast can be seen from pairs like deverbal nominal [toori-gakari] ‘passing by’ and verbal [toori-kakaru] ‘pass by’. This tendency is consistent with Vance's inference that rendaku originated from reduction of a genitive particle (i.e. /-no/ in contemporary Japanese) involved in noun phrases. If this inference is on the right track, it is natural that rendaku is most frequently attested in nominal compounds. On the other hand, compounds involving an adjective tend to undergo rendaku, regardless of the grammatical category of resulting compounds. This observation is obviously incompatible with the inferred historical development of rendaku, because there is no sensible reason to suppose that compounds involving an adjective contained the genitive particle or equivalent sometime in the past. Vance leaves this mystery open for future research.

Haruka Fukazawa & Mafuyu Kitahara provide an OT analysis of rendaku and postnasal voicing involved in loanwords. Rendaku is blocked by Lyman's Law if a second member of a compound contains a lexical voiced obstruent. The domain of this OCP effect is a morpheme. Following their own work (Fukazawa & Kitahara Reference Fukazawa, Kitahara, Weijer and Nishihara2001), they propose that Uniformity should be relativised with regard to morphemes, contra Ito & Mester (2003), who specify the domain in a markedness constraint. The novel aspect of their analysis is the proposed extension to loanwords involving the English plural suffix -s. Loanwords normally retain voicing of original words in postnasal contexts (i.e. Ident[voi]≫*N). However, examples with the plural suffix exhibit postnasal voicing (e.g. [saNzu] ‘Suns’, but [kabusu] ‘Cubs’), suggesting the opposite ranking. Fukazawa & Kitahara claim that this ranking paradox can be resolved by further extending Uniformity and Ident[voi] with regard to morphological categories such as stems and affixes. They argue that their analysis is desirable because it makes use of faithfulness relativisation that is necessary in OT anyway, obliterating relativised markedness. But Kurisu (Reference Kurisu, Miyamoto and Ochi2007) argues against their analysis. One problem is that morpheme-specific Uniformity does not penalise the faithful parsing of two voiced obstruents that share one and the same [voice] feature in the underlying representation. This incorrectly implies that Lyman's Law does not have to be respected in native Yamato-Japanese. Kurisu draws the conclusion that relativised markedness constraints are necessary in OT grammar.

Noriko Yamane-Tanaka analyses the gradual disappearance of intervocalic prenasalised stops in Japanese. Starting with [^mb ⁿd ŋ m n] in Old Japanese (OJ), [^mb] was lost in Middle Japanese (MJ), [ⁿd] disappeared in Modern Japanese (ModJ) and [ŋ] was the last to be eliminated, in present-day Japanese (PJ). This historical change is harmonically complete. She posits the ranking *[^ŋg]≫*[^mb]≫*[ⁿd]≫*[ŋ]. The successive disappearance of prenasalised stops and [ŋ] is analysed via demotion of Max[nas]. In OJ, Max[nas] was placed between *[^ŋg] and *[^mb], and this faithfulness constraint has been demoted by one step in each subsequent historical stage. Interestingly, this diachronic change corresponds to synchronic variation in Japanese dialects. Yamane-Tanaka conjectures that the voicing contrast in stops has been gradually established via disappearance of prenasalised stops. In OJ, voiced stops were absent. Rather, voiceless stops and prenasalised stops existed. Prenasalised stops have been substituted by plain voiced stops, leading to the contrast between voiced and voiceless stops in PJ.

Hideki Zamma examines how rendaku and accentuation are correlated in 347 frequent surnames. In her limited research on Japanese surnames ending in /ta/ ‘field’, Sugito (Reference Sugito1965) claims that rendaku and accent appear in complementary distribution in Japanese surnames (e.g. [jóko-ta] (accented without rendaku) vs. [jama-da] (unaccented but with rendaku)). But Zamma's finding reveals that such complementarity is indeed not regular across the board. He argues that sensitivity to rendaku and accentuation is a morpheme-specific characteristic. Morphemes are grouped into the following five categories: (i) both kinds of information are specified, (ii) sensitivity only to rendaku is specified, (iii) sensitivity only to accentuation is specified, (iv) neither is specified and (v) factors other than rendaku and accentuation must be taken into account. This work is quite interesting at the descriptive level, but no theoretical analysis is given. Zamma remarks that we need some theoretical machinery that is capable of incorporating quite a few patterns of variation.

Tomoaki Takayama considers grammatical and sociolinguistic conditions on the occurrence of rendaku in loanwords. Rendaku is often described as a characteristic of native items, but some recent loanwords are subject to this voicing process. He argues that Japanese speakers tend to construe loanwords as native items if they conform to native phonotactic patterns and if they are culturally not associated with the countries where the source languages are spoken. However, phonotactics and cultural association do not explain rendaku in Sino-Japanese. For example, [keNka] ‘quarrel’ contains an N cluster prohibited in native vocabulary, but it still undergoes rendaku (e.g. [ojako-geNka] ‘parent-child quarrel’). Sino-Japanese items often sound formal, but not all do. Takayama maintains that vulgarised informal Sino-Japanese items are likely to undergo rendaku. He leaves open the question as to why recent loanwords and Sino-Japanese items behave differently. But he hints at the importance of studying historical background of loanword borrowings.

Keiichiro Suzuki reports on a development in speech engineering, Large Vocabulary Continuous Speech Recognition (LVCSR). Japanese has a large number of numeral classifiers. They display various context-dependent phonological changes and free variation, including voicing alternations. After LVCSR acquires raw data and its lexicon is fixed such that all variants of numeral classifiers exist and the frequency is adjusted with n-gram extrapolation, the goal is to test to what extent LVCSR succeeds in recognising sentences with 65 major numeral classifiers, of which 12 begin with /h/. /h/-initial classifiers were separated from the others because they exhibit the richest phonological alternations. The test was conducted against four versions of LVCSR that differ in the degree of pre-experimental fixes. The overall results showed that the performance of LVCSR improved if the pre-experimental fixes were more thorough. Suzuki concludes that fixes are effective strategies in engineering, and suggests that phonologists and engineers in automatic speech recognition can learn from each other.

Part II comprises four articles. Kikuo Maekawa & Hideaki Kikuchi recast and confirm an earlier finding that devoicing of high vowels is largely affected by the manner of consonants in ₁V₂ contexts. Examining a corpus of spontaneous speech, they made two findings. First, interconsonantal high vowels are most and least likely to be devoiced if C₂ is a stop and a fricative respectively. They connect this difference with perception of mora boundaries. If a vowel is devoiced when C₂ is a fricative, the VC sequence is realised as successive frication. This blurs a mora demarcation, so devoicing is dispreferred. By contrast, stops involve a burst of oral closure. This enhances recognition of a mora boundary even if the preceding vowel is devoiced. The other finding is that consecutive devoicing tends to be avoided in VV contexts. Only 26·5% of samples showed devoicing of both vowels. Most of the first vowels were devoiced, but second vowels underwent devoicing much less frequently. Maekawa & Kikuchi argue that devoicing of a second vowel is contingent on whether the first vowel undergoes devoicing or not, but not vice versa. They also suggest that speech rate and laughter can have an influence on vowel devoicing. They conclude that more comprehensive linguistic and extralinguistic research must be carried out for a full understanding of vowel devoicing. It may be worthwhile to investigate the dispreference for consecutive devoicing from phonological perspectives (e.g. foot structure).

Mariko Kondo offers a syllable-based analysis of the fact that not all high vowels are actually devoiced in _ and _# contexts in consecutive devoicing environments. Under the assumption that devoiced vowels are unable to host a syllable node, she argues that devoicing occurs only when consonants in the same syllable as the devoiced vowel can be resyllabified appropriately. She assumes that Japanese sanctions CCVVNC maximal syllables, contrary to the standard idea that onsets and codas must not be complex. According to her, this assumption is motivated by the presence of examples like [ʧa] ‘tea’ and [a.ba.diiNk.ko] ‘Aberdonian’. In /dookutu/ ‘cave’, for example, only the penultimate /u/ is devoiced (i.e. [dooktu]→[doo.ktu]). Kondo claims that /k/ is resyllabified as an onset with /t/. The final /u/ is not devoiced, to prevent an illicit [kt] coda cluster. The basic assumption is questionable, however. Palatalised consonants are normally interpreted as single segments. In contemporary Japanese, palatalised coronals are realised as (alveo)palatal consonants, and the palatalisation feature is superposed on primary place as a secondary articulation for labials and dorsals. Also, [kt] does not appear as an onset elsewhere in Japanese, and NC coda clusters are very restricted. Kondo's analysis does not predict these distributional restrictions.

Miyoko Sugito explores the acoustic and physiological properties of devoiced accented vowels in the Osaka dialect of Japanese, especially with reference to speech rate. Accent often shifts in many dialects when accented vowels are devoiced, but devoiced vowels retain accent in this dialect. Acoustic experiments revealed that accent tends to remain on devoiced vowels in fast speech, compared with normal speech (for more on the phonetic realisation of accent, see the summary of Tanaka's article below). In very fast speech, devoicing was accompanied by the change of HL to HH, where H and L stand for high and low pitch respectively. Sugito conjectures that this is due to the lack of enough transitional time from a devoiced vowel to the following vowel. One more finding, obtained with physiological experiments using EGM, is that the quality of vowel after a devoiced vowel has much to do with the accent change. Comparing /kúʃi/ ‘comb’ and /kúsa/ ‘grass’, F0 falls more abruptly in the latter after the accented vowel. [a] is inherently longer than [i]. Sugito claims that the vowel after a devoiced vowel must have sufficient duration for substantial pitch fall to occur.

Shin-ichi Tanaka articulates a prominence hierarchy and applies it to interactions of accent placement and vowel devoicing. He puts forward the following scale of phonological salience: accent>tone/length>sonority>voicing. This hierarchy is based on the degree of correlated articulatory effort. Voicing requires only vibration of vocal cords, whereas accent assignment entails factors such as aperture, pitch, duration and intensity, as well as vocal cord vibration, given that accent is assigned to vowels. Among other things, accent implies voicing in the hierarchy. Tanaka observes that accent often falls on a devoiced vowel when the locus of accent vacillates (e.g. [k-sok] or [k-sók] ‘rule’). These two forms sound equally natural to native speakers of Japanese. The implicational relation is disobeyed in [k-sok]. Derivationally, [k-sok] arises from a counterfeeding order of Accent Shift (relocation of accent on devoiced vowels) and Devoicing. Devoicing feeds Accent Shift, but [k-sok] surfaces if the two rules apply in the opposite order. As a result, such data present two challenges: disobedience to the prominence hierarchy and opacity. Tanaka presents a solution which appeals to Sympathy Theory (McCarthy Reference McCarthy1999). The idea is that [kí-soku] without devoicing is chosen as the designated candidate, and Max-❀O[accent] compels the ultimate output to agree with [kí-soku] in accent placement. Then opaque [k-sok] wins the competition since it successfully respects both Devoicing and Max-❀O[accent].

To summarise, the volume as a whole covers rendaku, postnasal voicing, vowel devoicing and issues surrounding voicing contrasts and correlation of (de)voicing with accentuation. One of the voicing phenomena not covered here is mimetic voicing (e.g. /pata-pata/→[bata-bata] ‘rattling’). Some recent studies reveal that this voicing effect exhibits quite regular behaviour. It also has numerous theoretical consequences with regard to relativisability of OT constraints and the phonological organisation of the Japanese lexicon (see Kurisu Reference Kurisu, Vance and Jones2006, Reference Kurisu, Miyamoto and Ochi2007 and works cited therein for relevant studies).

The volume as a whole suffers from a lack of stylistic coherence. American and British English spelling conventions are mixed, data transcriptions are not consistent, and there are a number of typographical and descriptive errors. As a result, those unfamiliar with Japanese may be confused or misled. Also, some of the papers seem preliminary. Nevertheless, this book will be informative to researchers of Japanese phonology and to those interested in voicing phonology in general. Some of the articles discuss closely related issues, and the authors refer to each other's work, so the reader will be easily aware of the connections among them. Further rigorous, comprehensive and systematic study of the issues taken up in this book will bring us fruitful insights and more interesting questions. I appreciate each contribution in the volume as a positive step toward our deeper understanding of the phonology of voicing.