Old non downloadable version of nasalization

In this section I provide background information on the topics with which this paper is concerned. §2.1 discusses incomplete neutralisation and phonetic paradigm uniformity effects and shows how they are problematic for a modular architecture of grammar. §2.2 introduces the reader to the initial mutations of the Celtic languages and provides a detailed review of the existing literature on incomplete neutralisation in initial mutation. §2.3 discusses phonological and phonetic patterns of vowel nasalisation in Scottish Gaelic, and the manner in which vowel nasalisation interacts with a neutralising process triggered by an initial mutation.

Incomplete neutralisation and phonetic paradigm uniformity effects

Morphological paradigms often involve neutralisation, where a phonological contrast that is observable in one part of the paradigm is not realised in another. For instance, in German, the contrast between voiced and voiceless obstruents is neutralised word-finally as a result of final devoicing:

1. (1)

   Rad	/ʀaːd/	[ʀaːt]	‘wheel’	cf.	gen	Rades	/ʀaːd-əs/	[ʀaːdəs]
   Rat	/ʀaːt/	[ʀaːt]	‘advice’	gen	Rates	/ʀaːt-əs/	[ʀaːtəs]

However, many studies have found that the neutralisation brought about by final devoicing is incomplete. Acoustically, this incomplete neutralisation typically manifests itself in the partial retention by devoiced consonants of certain properties normally associated with voiced consonants, such as shorter closure duration or longer preceding vowel duration. As well as for German (Mitleb 1981; Fourakis and Iverson 1984; Charles-Luce 1985; Port and O’Dell 1985; Port and Crawford 1989; Piroth and Janker 2004), acoustic evidence of incomplete neutralisation in final devoicing has been reported for Catalan (Dinnsen and Charles-Luce 1984; Charles-Luce and Dinnsen 1987), Polish (Slowiaczek and Dinnsen 1985; Tieszen 1997), Russian (Barry 1988; Dmitrieva et al. 2010; Kulikov 2012; Kharlamov 2012, 2014; Shrager 2012), Afrikaans (van Rooy et al. 2003; Kaplan 2017) and Dutch (Warner et al. 2004). It has also been found that listeners perform at above-chance level when attempting to discriminate between minimal pairs that differ only in underlying voicing in German (Port et al. 1984; Port and O’Dell 1985; Port and Crawford 1989), Polish (Slowiaczek and Szymanska 1989), Afrikaans (van Rooy et al. 2003) and Russian (Matsui 2011; Kharlamov 2012, 2015). Besides final devoicing, incomplete neutralisation has also been reported in connection with various other phenomena, such as American English t/d-flapping (Fisher and Hirsh 1976; Fox and Terbeek 1977; Zue and Laferriere 1979; Patterson and Connine 2001; Herd et al. 2010; Braver 2011, 2013, 2014) and Russian unstressed vowel reduction (Padgett and Tabain 2005; Kaplan 2017).

Although statistically significant across large numbers of tokens, the differences between incompletely neutralised segments are small, gradient and not readily perceptible, and may generally be detected only through experimentation and statistical analysis. Under a traditional view of phonology as operating upon discrete symbolic representations, they therefore cannot be comfortably assigned to distinct phonological categories, which is problematic for a strictly modular feedforward architecture of grammar in which the phonetics is sensitive only to the categorical output of the phonology. Nevertheless, it has been suggested that incompletely neutralised segments do in fact have distinct phonological surface representations, rendering the phenomenon compatible with such a framework. Van Oostendorp (2008) argues that a devoiced segment continues to bear an unpronounced [voice] feature which may subtly influence its phonetic interpretation, while Iosad (2017:22–24) considers incomplete neutralisation to occur when fully distinct phonological representations are assigned ranges of phonetic realisations that happen to almost entirely overlap. However, if the phonetics is sufficiently powerful to bring about such near-complete neutralisation of a categorical phonological contrast, then this raises the question of whether a separate categorical phonology is necessary at all.

Because incompletely neutralised forms display subtle traces of the properties of paradigmatically related forms, incomplete neutralisation may be regarded as a type of paradigm uniformity effect. Paradigm uniformity effects at the level of categorical phonology are well-known. For instance, English post-nasal coda g-deletion overapplies in singer /sɪNg-ə/ [sɪŋə] (cf.finger /fɪNgə/ [fɪŋgə]) as a result of its paradigmatic relationship with sing /sɪNg/ [sɪŋ], in which deletion applies transparently. Under a modular architecture this can be handled either by means of cyclicity (Bermúdez-Otero 2011), whereby the deletion rule applies to the stem before suffixation takes place, or by output-output (OO-)correspondence (Kenstowicz 1996; Benua 1997), in which highly-ranked OO-correspondence constraints enforce phonological identity between corresponding morphemes in paradigmatically related forms. Regardless of which approach is correct, paradigm uniformity effects at the level of categorical phonology are unproblematic for a modular architecture. Note that the term incomplete neutralisation is therefore not appropriate for cases where the observed difference is sufficiently large, categorical and readily perceptible to constitute an ordinary phonological contrast—for example, in Friulian final devoicing, vowels before underlyingly voiced stops may be over twice as long as those before underlyingly voiceless ones (Baroni and Vanelli 2000). In this case, neutralisation clearly does not occur in the first place and the surface contrast can be considered straightforwardly phonological.

Unlike those at the level of categorical phonology, paradigm uniformity effects at the level of gradient phonetics can be used to argue for non-modular frameworks in which the phonetics has direct access to morphological and/or lexical information. Steriade (2000) and Braver (2013, 2019) argue that OO-correspondence constraints may make reference to phonetic properties, thus motivating identity between paradigmatically related forms at the level of fine-grained phonetic detail. Regarding incomplete neutralisation in final devoicing, Ernestus and Baayen (2006, 2007), Kleber et al. (2010), Winter and Röttger (2011) and Roettger et al. (2014) favour an exemplar-based approach in which both inflected and uninflected forms are stored together in a phonetically rich lexicon and the production of one form may be subtly influenced by partial co-activation of paradigmatically related forms. Both of these approaches are inherently non-modular, as they discard the traditional distinction between categorical phonology and gradient phonetics.

The overwhelming majority of work on incomplete neutralisation has focused on Indo-European languages and the straightforwardly concatenative morphology that is common throughout that family—notable exceptions include Peng (2000) and Yu (2007), who look at the neutralisation of tonal contrasts in Chinese languages, and Gouskova and Hall (2009) and Hall (2013, 2017), who investigate neutralising processes affecting vowels in a variety of Arabic. It is therefore likely that we do not yet have a complete picture of the types of processes that may bring about incomplete neutralisation. In particular, very little work exists on the Celtic languages, which—although they belong to Indo-European—display a typologically unusual type of morphophonological alternation known as initial mutation. While final devoicing is a highly transparent and phonetically natural process that serves to satisfy a phonotactic constraint, the alternations that occur in connection with initial mutation are different in nature, being directly triggered by the morphology itself. It is therefore far from clear that initial mutation should be subject to the same kinds of paradigmatic effects as have been observed in better-studied alternations. This study investigates an initial mutation in Scottish Gaelic in order to determine whether it brings about complete or incomplete neutralisation.

Initial mutation

All of the living Celtic languages are characterised by systems of morphophonological alternations in initial consonants known as initial mutations. Radical (unmutated) consonants are replaced by their mutated counterparts under a variety of conditions. For example, possessive particles in Irish and Welsh trigger a variety of mutations on a following noun:

1. (2)

   a.	Irish
   a cat	[ə ]	‘her cat’	(radical)
   a chat	[ə ]	‘his cat’	(lenition)
   a gcat	[ə ]	‘their cat’	(eclipsis)
   b.	Welsh
   eu cath	[i kaːθ]	‘their cat’	(radical)
   ei gath	[i gaːθ]	‘his cat’	(soft mutation)
   fy nghath	[və ̊ŋaːθ]	‘my cat’	(nasal mutation)
   ei chath	[i xaːθ]	‘her cat’	(aspirate mutation)

In Irish, radical initial /k/ undergoes frication to [x] under lenition and voicing to [g] under eclipsis. Meanwhile, in Welsh, radical initial /k/ undergoes voicing to [g] under the soft mutation, nasalisation to [̊ŋ] under the nasal mutation and frication to [x] under the aspirate mutation. An extensive literature exists on the analysis of the initial mutations and their exact place in the grammar (e.g. Hamp 1951; Oftedal 1962; Ellis 1965; Rogers 1972; Ó Dochartaigh 1978; Ewen 1982; Lieber 1983; Ball and Müller 1992; Kibre 1997; Pyatt 1997; Stewart 2004; Green 2006; Wolf 2007; Hannahs 2013; Iosad 2014). Importantly, initial mutation differs from the straightforwardly concatenative morphology that is common throughout Indo-European in that it appears to involve the substitution of one segment for another, although authors such as Lieber (1983, 1987), Wolf (2007) and Iosad (2014, 2017) offer accounts in which the relevant alternations are triggered by the affixation of featural autosegments such as [+continuant], [−spread glottis] or [+nasal]. While the examples in (2) contain an overt local trigger, initial mutations can also display purely morphological conditioning: for example, lenition in Irish marks past tense on verbs and feminine agreement on attributive adjectives. Throughout this paper, I will therefore assume that a set of forms such as Irish [] ∼ [] ∼ [] or Welsh [kaːθ] ∼ [gaːθ] ∼ [̊ŋaːθ] ∼ [xaːθ] make up a morphological paradigm and are thus comparable to sets like German [ʀaːt] ∼ [ʀaːdəs].

Initial mutation very often brings about neutralisation. In some cases, a particular segment may occur both as radical initial and also as the mutated grade of a different radical initial: for example, initial [b^ɣ g] in Irish may represent either radical /b^ɣ g/ or the eclipsis grade of radical /p^ɣ k/. In other cases, a particular segment may occur as the mutated grade of more than one radical initial: for example, initial [v] in Welsh represents the soft mutation grade of both radical /b/ and radical /m/. A small number of recent studies have investigated whether neutralisations such as these are incomplete, with no convincing positive results. Archangeli et al. (2014) report preliminary results, from three speakers, of an ultrasound study of the articulation of lenited consonants in Scottish Gaelic. The authors compare the position of the tongue body during the articulation of various neutralised consonants, such as [h] from lenition of // (e.g. thachd [haxk] ‘choke’, radical tachd []) vs. [h] from lenition of // (e.g. shad [] ‘toss’, radical sad []) vs. non-alternating [h] (e.g. tha [haː] ‘be.prs’). Although they do report some incomplete neutralisation, the items that they compare are poorly matched for following environment. All of the target consonants are either dorsal or glottal, meaning that the position of the tongue body during their articulation will be highly sensitive to anticipatory co-articulation with the following vowel. In a pair such as thachd [haxk] ‘choke’ (radical tachd []) vs. shad [] ‘toss’ (radical sad []), for instance, it is likely that the differing place of articulation of the following consonant results in a slight difference in the articulation of the vowel, which in turn will be reflected in the position of the tongue body during [h]. Some pairs involve a short vowel vs. a long vowel, e.g. dhiubh [ju] ‘of them’ (non-alternating) vs. ghiùlain [] ‘behave’ (radical giùlain []), or a short vowel vs. the first element of a diphthong, e.g. ghabh [ɣav] ‘take’ (radical gabh [kav]) vs. dhall [] ‘blind’ (radical dall []), where an exact match in quality is not necessarily to be expected. Additionally, one of the three speakers is from Lewis, where it is well documented that [u(ː)] displays a highly distinct retracted allophone next to velarised consonants such as [] (Borgstrøm 1940:32–33; Oftedal 1956:75; Ladefoged et al. 1998; Nance 2011), and gabh ∼ ghabh is normally pronounced [kɔ ∼ ɣɔ], potentially rendering these pairs very poorly matched for this speaker in particular. I therefore do not consider these results to be convincing evidence for incomplete neutralisation.

As part of a broader psycholinguistic study, Ussishkin et al. (2017) investigate whether speakers of Scottish Gaelic are able to discriminate between [f] from lenition of /p/ (e.g. phioc [f̯ioxk] ‘pick’, radical pioc [p^h̯ioxk]) and [f] representing radical /f/ (e.g. feòrag [f̯ioːɾak] ‘squirrel’). No statistically significant effect is found.

Welby et al. (2017; see also2011, 2014) investigate whether Irish [b^ɣ g] from eclipsis of /p^ɣ k/ (e.g. gcasúr [] ‘hammer’, radical casúr []) differ acoustically from [b^ɣ g] representing radical /b^ɣ g/ (e.g. gadaí [] ‘robber’). No statistically significant effect is found for VOT, consonant duration or closure duration, but the average intensity of the stop burst is found to be significantly greater for [b^ɣ g] from eclipsis of /p^ɣ k/ than for [b^ɣ g] representing radical /b^ɣ g/. However, as acknowledged by the authors themselves, this study encounters a serious problem. All of the target items are nouns immediately preceded by the definite article an [], which may be realised with or without []. In the eclipsis items, eclipsis is triggered by the presence of a preposition (e.g. ar an gcasúr [εɾ^j) ] ‘on the hammer’), while no preposition is present in the radical items (e.g. an gadaí [] ‘the robber’). Unexpectedly, it is found that the []-less variant of the article occurs far more frequently after prepositions than when no preposition is present. Because the authors were unable to control for this, any difference in average burst intensity between eclipsis voiced stops in e.g. ar an gcasúr and radical voiced stops in e.g. an gadaí is likely to be due to the fact that the stop is preceded by [] far more frequently in the radical context than in the eclipsis context.

Venturing outwith Celtic, some Austronesian languages display a morphophonological process known as nasal substitution, which resembles the initial mutations of the Celtic languages. In an ultrasound study of Javanese and Sasak, Archangeli et al. (2017) compare the articulation of nasal [ɲ] from radical /s/ to that of nasal [ɲ] from radical //, while Seyfarth et al. (2019) investigate whether Javanese [m ŋ] from radical /p k/ differ acoustically from non-alternating [m ŋ]. No statistically significant effect is found in either case.

Welby et al. (2017:131–132) also cite two cases from the existing descriptive literature of what they refer to as “incomplete neutralisation” in initial mutation. First of all, Falc’hun (1951) reports for the Bas-Léon dialect of Breton that [b₁ d₁ g₁] representing radical /b d g/ (e.g. dour [d₁uːr] ‘water’) are longer and have stronger release bursts than [b₂ d₂ g₂] from lenition of /p t k/ (e.g. dour [d₂uːr] ‘tower’, radical tour [tuːr]). However, what Falc’hun is describing here is part of a phonological contrast between fortis and lenis consonants that pervades almost the entire consonant system of the dialect in question, whereby all radical initial consonants are fortis and are thus considerably longer and more strongly articulated than the lenis consonants that result from lenition (see also Falc’hun 1943:43–44; Kervella 1947; Hamp 1951; Carlyle 1988). The distinction is readily perceptible to native speakers and it appears unlikely that phonological neutralisation is involved in the first place. In any case, the difference is in the opposite direction to what would be expected under incomplete neutralisation.^{Footnote 1} Secondly, Borgstrøm (1940) and Oftedal (1956) report for the Lewis dialect of Scottish Gaelic that [m₁] representing radical /m ɲ/ (e.g. mara [m₁aɾə] ‘sea.gen’) are fully nasal while [m₂ ɲ₂] from nasalisation of /p t^j/ (e.g. am bara [ə m₂aɾə] ‘the wheelbarrow’, radical bara [paɾə] ‘wheelbarrow’) are post-stopped. However, given that the distinction is so great as to be readily perceptible by fieldworkers, it is again unlikely that phonological neutralisation is involved. It is also somewhat doubtful that nasalisation of initial stops in Scottish Gaelic should be regarded as part of the initial mutation system, rather than as a purely phonological process of sandhi between a nasal consonant and a following word-initial stop. There is therefore no convincing evidence so far for incomplete neutralisation in initial mutation.

Vowel nasalisation in Scottish Gaelic

The primary focus of this study is the interaction of vowel nasalisation with an initial mutation in Scottish Gaelic. During the production of a nasalised vowel, the velum is lowered as for a nasal consonant and a proportion of the pulmonic egressive airstream is allowed to escape through the nasal cavity rather than through the mouth. Nasalisation may be measured experimentally in a number of ways, some more invasive or impractical than others (see Krakow and Huffman 1993 for a detailed discussion). Krakow (1989, 1993) measures the degree of opening of the velopharyngeal port using the Velotrace, a mechanical device that is inserted through the nose and directly records the position of the velum (Horiguchi and Bell-Berti 1987), while Solé (1992, 1995) employs a nasograph, a photoelectric device inserted through the nose and into the oesophagus (Ohala 1971). More recently, Byrd et al. (2009), Proctor et al. (2013), Carignan et al. (2015) and Barlaz et al. (2018) use real-time magnetic resonance imaging to directly observe velic and other articulatory movements during the production of nasalised vowels. It is also possible to measure nasalisation acoustically by spectral analysis (e.g. Chen 1996, 1997; Beddor 2007; Carignan et al. 2011; Shosted et al. 2012; Li 2014; Zellou and Tamminga 2014; Scarborough et al. 2015; Cho et al. 2017; Seyfarth et al. 2019), although the effect of nasalisation on formant values is often complex. An uninvasive, highly practical technique involves measuring the rate of nasal airflow at the nostrils in order to obtain an indirect measure of velic opening (e.g. Cohn 1990, 1993; Huffman 1990; Jun 1993; Basset et al. 2001; Shosted 2006