3.1.1 Yaghnobi

PND was first proposed for Yaghnobi by Xromov (Reference Xromov1972). Yaghnobi is an Iranian language, spoken by approximately 13,500 speakers in five different areas of Tajikistan (Paul et al. Reference Paul, Abbess, Müller, Tiessen and Tiessen2010: 4). It is the only living descendant of Sogdian, an Eastern Iranian language that was spoken around the fourth century CE. Xromov (Reference Xromov1972) observes that NT sequences in Yaghnobi correspond to ND sequences in ancestral Sogdian; on the basis of this observation, he posits a sound change  in the development from Sogdian to Yaghnobi. Table 1 lists cognates from Yaghnobi and Sogdian that confirm this correspondence.Footnote ^[10]

Table 1

PND in Yaghnobi (from Xromov Reference Xromov1972: 128).

Outside of post-nasal position, original voiced stops surface as voiced fricatives [β, ð, ɣ] in Sogdian: this is the result of a sound change that turns voiced stops into voiced fricatives except post-nasally. In Yaghnobi, velar and labial fricatives in the elsewhere condition are preserved as fricatives [w/β] and [ɣ]. The Sogdian dental fricative [ð], on the other hand, surfaces as a stop: [dah] ‘ten’ for Sogdian [ðəsa] (data from Novák Reference Novák2010: 31; Novák Reference Novák2014). Voiceless stops remain unchanged in Sogdian, except post-nasally, where they become voiced (Yoshida Reference Yoshida2016) before being devoiced in Yaghnobi.

Regarding the Yaghnobi synchronic system, Novák (Reference Novák2010) reports voiceless stops to be aspirated except pre-consonantally and voiced stops to be phonetically voiced, although no instrumental studies are offered. Synchronic Yaghnobi phonology contrasts voiced and voiceless stops in all positions: initially, finally, intervocalically, in clusters, and post-nasally. The fact that [ $\pm$ voice] contrasts fully in NT and ND sequences in synchronic Yaghnobi means that PND is not an active alternation anymore. The post-nasal contrast in the feature [ $\pm$ voice] is, however, likely secondary, introduced late in the language’s development through borrowings from Tajik (cf. Xromov Reference Xromov1972). Some examples of borrowed ND sequences from Tajik are given in Xromov (Reference Xromov1972: 128): [angiʃt] ‘coal’, [ʧanɡ] ‘dust’, [baland] ‘high’, [lunda] ‘round’. It is even possible to find pairs of native and borrowed words with and without devoicing. For example, the inherited Yaghnobi [vant] ‘tie’ – with an unvoiced stop after a nasal – stands in contrast to the borrowed [band] ‘tie’ with a voiced variant.

3.1.2 Tswana, Shekgalagari, and Makhuwa

PND has been reported as a synchronic phonological process in the languages/dialects of the Sotho-Tswana group of Southern Bantu languages, especially in Tswana and Shekgalagari (Hyman Reference Hyman, Hume and Johnson2001, Solé, Hyman & Monaka Reference Solé, Hyman and Monaka2010), but also in Sotho (Janson Reference Janson1991/1992). The three languages are closely related and mutually intelligible (Makalela Reference Makalela2009). Because PND is least well described in Sotho, I will focus on Tswana and Shekgalagari. Tswana is spoken by approximately 4–5 million people in Botswana, Namibia, Zimbabwe, and South Africa (Coetzee & Pretorius Reference Coetzee and Pretorius2010), and Shekgalagari by approximately 272,000 people in Botswana (Solé et al. Reference Solé, Hyman and Monaka2010).

In Makhuwa, PND is reported as a sound change followed by nasal deletion (Janson Reference Janson1991/1992). Makhuwa is closely related to Sotho-Tswana too and is spoken by approximately 3.5 million speakers in Mozambique (Simons & Fennig Reference Simons and Fennig2018).

Table 2 shows that synchronic voiced stops (that surface as voiced stops in the elsewhere condition) in Shekgalagari become voiceless when preceded by a nasal. Voiceless stops remain unchanged both post-nasally and elsewhere.

Table 2

PND in Shekgalagari (table from Solé et al. Reference Solé, Hyman and Monaka2010). The N-prefix functions as the first person singular prefix, e.g. χʊ-pak-a ‘to praise’ vs. χʊ-m-pak-a ‘to praise me’ (Solé et al. Reference Solé, Hyman and Monaka2010).

Both Tswana and Shekgalagari have unaspirated voiced (D) and voiceless (T) and aspirated voiceless stops  in their inventories (for phonetic studies, see Coetzee & Pretorius Reference Coetzee and Pretorius2010 and Solé et al. Reference Solé, Hyman and Monaka2010): they surface initially, intervocalically, and post-nasally. The only permitted syllable structures are ‘V, CV, CGV, N’ (where G $=$ glide and N $=$ syllabic nasal; Coetzee & Pretorius Reference Coetzee and Pretorius2010: 406). Detailed instrumental acoustic studies of the Tswana and Shekgalagari stop system in Coetzee & Pretorius (Reference Coetzee and Pretorius2010) and in Solé et al. (Reference Solé, Hyman and Monaka2010) confirm that post-nasal devoicing is indeed realized as change in the feature [ $\pm$ voice], and not as a change in [ $\pm$ spread glottis].

Several peculiarities need to be noted with respect to Tswana. First, /ɡ/ never surfaces as a voiced stop: while it is devoiced to [k] post-nasally, it gets deleted elsewhere, e.g. [χʊ-at-a] for /χʊ-ɡat-a/ (cf. [χʊ-ŋ-kat-a], Solé et al. Reference Solé, Hyman and Monaka2010). Second, voiced alveolar stop [d] surfaces as an allophone of /l/ before the high vowels /i/ and /u/ (Coetzee & Pretorius Reference Coetzee and Pretorius2010). Third, voiced stops in nasal clusters of secondary origin (from NVD after vowel deletion) do not undergo devoicing, but undergo assimilation in Tswana (see Table 3). Finally, the nasal in NT sequences Tswana (and Shekgalagari) is retained when stressed and in the 1st person object prefix, but gets deleted elsewhere (see Dickens Reference Dickens1984). Hyman (Reference Hyman, Hume and Johnson2001) gives an example of preservation of N-: in the case of class 9 and 10 prefix, when roots are monosyllabic.

In Shekgalagari, secondary ND clusters (from NVD after vowel deletion) remain voiced (Table 3). /d/ can surface in the elsewhere condition (not only before /i/ and /u/), but it does not alternate with devoiced [nt]  $\leftarrow$  /nd/ (because /d/ corresponds to Tswana /tl/). Shekgalagari (unlike Tswana) also features the palatal series of stops that enters PND (Solé et al. Reference Solé, Hyman and Monaka2010). For a detailed discussion of differences between Tswana and Shekgalagari, see Dickens (Reference Dickens1984) and Solé et al. (Reference Solé, Hyman and Monaka2010).

Table 3

PND in Tswana and Shekgalagari (table from Solé et al. Reference Solé, Hyman and Monaka2010).

Makhuwa is also reported to undergo PND as a sound change. Sequences of a nasal and a voiced stop (ND) develop to voiceless unaspirated stops (T) (Janson Reference Janson1991/1992), but the process did not develop into a synchronic alternation in Makhuwa. To my knowledge, no elaborate accounts of the phonetic realization of voiced and voiceless stops in the elsewhere condition exist for Makhuwa.

Because the languages are closely related and there are other instances of common innovation between Sotho-Tswana and Makhuwa, Janson (Reference Janson1991/1992) and Hyman (Reference Hyman, Hume and Johnson2001) imply that PND in the three languages is likely a common innovation. For the same reason, I will treat these three languages together in this study. Since the description of PND in Makhuwa (and Sotho) is sparse and lacking in detailed phonetic descriptions, I will focus my discussion on Tswana and Shekgalagari; however in principle, the arguments for these two languages apply to Makhuwa as well.

3.1.3 Bube and Mpongwe

Unlike Makhuwa, Bube is not closely related to Tswana and Shekgalagari. Janssens (Reference Janssens1993) reports that Bube also features PND. As will be shown below, several aspects of Bube PND are highly reminiscent of the process reported for Tswana, Shekgalagari, and Makhuwa despite the languages not being closely related. Bube is a Northwest A Bantu language, spoken by approximately 51,000 speakers on Bioko island (Simons & Fennig Reference Simons and Fennig2018). Sequences of a nasal and a voiced stop in Pre-Bube develop to voiceless stops in Bube. Table 4 illustrates the development.

Table 4

PND in Bube (table from Janssens Reference Janssens1993).

Table 4 illustrates that Pre-Bube sequences of a nasal and voiced stop (ND) yield a single voiceless stop (T) with the nasal being lost. In the elsewhere condition, the labial voiced stop surfaces as such, the dental develops to [r/l], and the voiced velar stop gets lost (similar to the situation in Tswana and Shekgalagari) (Janssens Reference Janssens1993). Voiceless stops can either delete, develop to [h] (in the labial series), or continue to surface as voiceless stops. NT sequences develop to a plain voiceless stop (T; Janssens Reference Janssens1993). The Bube synchronic system has no alternation between voiceless and voiced stops in the post-nasal vs. elsewhere condition, but the development is intriguing from a diachronic perspective: it appears as if PND operated in Pre-Bube.

In the Mpongwe dialect of Myene (Bantu B language, spoken in Gabon, Simons & Fennig Reference Simons and Fennig2018), PND is reported marginally for root initial [ɡ] after some prefixes that historically ended in a nasal, e.g. *gàmb-  ${>}$  [i-kamba] (Mouguiama-Daouda Reference Mouguiama-Daouda1990). However, because PND is marginal in Mpongwe – it is morphologically limited and does not apply categorically – I will for the most part leave it out of the ensuing discussion.Footnote ^[11]

3.1.4 Konyagi

Recently, PND as a sound change has been discovered in Konyagi (also known as Wamey, among others; Merrill Reference Merrill2014, Reference Merrill2016a, Reference Merrillb). Konyagi, a member of the Atlantic subfamily of the Niger-Congo group, is spoken by approximately 21,000 speakers in Senegal (Simons & Fennig Reference Simons and Fennig2018). Note that Konyagi is not part of the Bantu family, which means that it is only very distantly related to the Bantu languages above with PND. Merrill (Reference Merrill2014, Reference Merrill2016a, Reference Merrillb, p.c.) reconstructs a detailed picture of Konyagi’s pre-history. Notable in this reconstruction is a series of voiceless stops in post-nasal position that correspond to voiced stops in the neighboring languages Bedik and Basari of the Tenda group (data in Table 5 is from Merrill Reference Merrill2014, Reference Merrill2016a, Reference Merrillb, based on Ferry Reference Ferry1991 and Santos Reference Santos1996). It thus appears as if Konyagi underwent PND. Voiceless NT sequences develop to a plain voiceless stop (T) in Konyagi. Both voiceless and voiced stops fricativize in the elsewhere condition. Table 5 illustrates PND in Konyagi.

Table 5

PND in Konyagi (from Merrill Reference Merrill2014, Reference Merrill2016a, Reference Merrillb).

Synchronically, Konyagi features voiceless and voiced stops as well as voiceless and voiced fricatives. Non-nasal clusters are not permitted. Post-nasally, only voiceless stops are allowed. Elsewhere, voiceless and voiced stops can contrast with voiceless and voiced fricatives, except initially where only fricatives are allowed (Merrill Reference Merrill2014, Reference Merrill2016a, Reference Merrillb). PND is part of a synchronic mutation process in Konyagi: devoiced stops after prefixes that historically ended in a nasal alternate with voiced fricatives after prefixes that historically ended in a vowel and voiced stops after prefixes that ended in another consonant (see Table 27 and Merrill Reference Merrill2014, Reference Merrill2016a, Reference Merrillb). No instrumental phonetic data is available for Konyagi.

3.1.5 South Italian dialects

Sicilian and Calabrian are dialects of Italian spoken in the corresponding regions of Italy by approximately 4.7 million speakers (Simons & Fennig Reference Simons and Fennig2018). PND has been reported for these dialects in Rohlfs (Reference Rohlfs1949: 424–425). The peculiarity of South Italian PND is that the sound change targets only the voiced affricate *ʤ, which is devoiced to [ʧ] after the nasal  Elsewhere, *ʤ develops to [j] in these dialects (Rohlfs Reference Rohlfs1949: 265, 358–359; Kümmel Reference Kümmel2007: 376). Voiced stops are not reported to be devoiced in the post-nasal position: the feature [ $\pm$ voice] is contrastive for stops. The voiceless affricate either remains unchanged or develops to [ʃ] (Rohlfs Reference Rohlfs1949). Table 6 illustrates PND in Sicilian and Calabrian.

Table 6

PND in South Italian (from Rohlfs Reference Rohlfs1949: 424–425).

3.1.6 Buginese and Murik

PND has been reported in three Austronesian languages: Buginese, Murik, and the Bengoh dialect of Land Dayak (Blust Reference Blust2005, Reference Blust2013). PND in the latter is simply mentioned without accompanying data (Rensch et al. Reference Rensch, Rensch, Noeb and Ridu2006: 69; Blust Reference Blust2013); I therefore leave Land Dayak out of the discussion that follows. Buginese is spoken by approximately 5 million people in Sulawesi (Indonesia); Murik is spoken in Sarawak (Malaysia and Brunei) by approximately 1,000 speakers (Simons & Fennig Reference Simons and Fennig2018). These three Austronesian languages are not closely related, so we cannot attribute PND to developments in a common ancestor; it is likely that PND developed independently in all three branches (Blust Reference Blust2005, Reference Blust2013).

Apparent PND in Buginese is represented in Table 7, showing the development of Proto-Malayo-Polynesian (PMP) voiced stops (data from Blust Reference Blust2013). Velar stops are devoiced post-nasally. Labial stops appear devoiced after nasals, but surface as [w] initially and word-internally (with a sporadic reflex [b] in initial position). The dental stop *d is not implicated in PND; Pre-Buginese *d develops to /r/ in all positions, which does not undergo devoicing, e.g. *dindiŋ  ${>}$  [renriŋ]. Word-initially, however, *d is sporadically preserved as a voiced stop [d] or develops to [l]. The voiced fricative *z is occluded to a voiced palatal stop initially, and develops to a sonorant [r] intervocalically. Post-nasally, *z is devoiced to [c]. Word-finally, all stops develop to [ʔ]. Voiceless NT sequences develop to a geminate voiceless stop (TT) in Buginese (Mills Reference Mills1975, Blust Reference Blust2005). Voiceless stops remain unchanged in the elsewhere position (Mills Reference Mills1975).

Table 7

Summary of PND in Buginese.

The data cited as evidence of PND as a sound change in Buginese are given in Table 8.

Table 8

PND in Buginese with cognates from reconstructed Proto-South-Sulawesi (Proto-SS; from Blust Reference Blust2005, Reference Blust2013).

In Murik, labials, alveolars, and velars undergo devoicing in post-nasal position. In the elsewhere condition, the developments vary according to the place of articulation. Bilabial voiced stops surface as such in the elsewhere condition. Voiced alveolars appear as [l] word-initially and [r] word-internally. Velars are devoiced not only post-nasally, but sporadically also in the elsewhere condition. PMP *z develops to the voiced palatal affricate [ɟʝ] initially, to [s] word-internally, and to the voiceless palatal affricate [cc¸] post-nasally. Table 9 illustrates the development of voiced stops from Proto-Kayan-Murik. Voiceless NT sequences develop to a plain voiceless stop (T) in Murik, while plain voiceless stops remain unchanged (Blust Reference Blust1974, Reference Blust2005).

Table 9

Summary of PND in Murik (as reconstructed in Blust Reference Blust2005).

PND is confirmed for Murik by the examples in Table 10.

Table 10

PND in Murik (from Blust Reference Blust2005: 259–262; Blust Reference Blust2013: 668).

Synchronically, PND in Buginese is reported to be an active phonological alternation (Sirk Reference Sirk1983) as well as an active sandhi process (Noorduyn Reference Noorduyn and Macknight2012/1955; see Section 3.5 and Table 26). Buginese contrasts voiceless and voiced stops (Noorduyn Reference Noorduyn and Macknight2012/1955), but voiced labial, palatal, and velar stops /b, d, ɡ/ and a glide /w/ are reported to devoice (and change to a stop in the case of /w/) post-nasally in synchronic Buginese (see also Table 26 below). In Murik, PND is reported as a synchronic alternation only for the palatal stop series which alternates between [ɟ] in the elsewhere condition and [c] post-nasally. For other places of articulation, Blust (Reference Blust2005: 260) reports PND to be synchronically inactive due to ‘phonemic restructuring’.

3.1.7 Nasioi

The most recent report of PND is that in Brown (Reference Brown2017) for the South Bougainville language Nasioi, spoken in Papua New Guinea by approximately 20,000 speakers (Simons & Fennig Reference Simons and Fennig2018). Brown (Reference Brown2017), based on Hurd & Hurd (Reference Hurd and Hurd1970), claims that PND in Nasioi is a synchronic alternation and supports this claim by showing that while bilabial and alveolar voiced stops /b/ and /d/ contrast in voicing with unaspirated voiceless /p/ and /t/ word-initially and after a glottal stop, only voiceless variants appear in post-nasal position. Elsewhere, stops fully contrast in the feature [ $\pm$ voice], but are not permitted other than word-initially and after a glottal stop: intervocalically, labial and alveolar stops surface as a voiced fricative [β] and a flap [ɾ], respectively. The velar voiced stop /ɡ/ does not occur in the system. This distribution points to a clear case of PND. PND in Nasioi is not only a phonotactic restriction, but also seems to be an active synchronic alternation, which is illustrated by the example in (4) from Brown (Reference Brown2017). The voiceless version of the personal pronoun -p/b- surfaces post-nasally; the voiced one surfaces elsewhere.

Brown (Reference Brown2017) specifically argues that PND is an innovation in Nasioi and that the fact that a related language, Nagovisi, shows traces of post-nasal voicing (with a speculation that this might be indicative of the proto-language) suggests that Nasioi PND operated as a single sound change. In other words, this suggests that PND is not ‘derived from the confluence of multiple independent changes’ (according to Brown Reference Brown2017: 275).

The data presented in this subsection seem to suggest, at first glance, that PND operated as a single sound change in the development of all thirteen of these languages. However, I will demonstrate below that a thorough investigation reveals a common pattern of complementary distribution in all thirteen cases along with other pieces of evidence, strongly suggesting that a combination of natural sound changes operated in place of a single PND sound change.

3.2.1 Hypercorrection

Xromov (Reference Xromov1972) appears to invoke hypercorrection to explain PND in Yaghnobi. He postulates that in Sogdian and Yaghnobi, all stops voice post-nasally, but that this sound change is in progress, which results in variation between [NT] ${\sim}$ [ND] for the underlying /NT/. He further claims that the underlying /ND/ joined this pattern (probably through hypercorrection), which means that at some stage variation between [NT] ${\sim}$ [ND] existed for /ND/ as well. Once the voicing process was finished, the voiceless variant became more frequent for both inputs. No motivation is given for why the voiceless variant becomes prevalent, except for an assumption that devoicing might have been influenced by morphological analogy in dentals and then spread to other places of articulation. Xromov’s explanation might be more convincing if we assumed hypercorrection arose from an interaction between two hypothetical dialects, where one voices all stops, and the other preserves the contrast; speakers of the neutralizing dialect could then hypercorrect their voiced stops to voiceless ones in post-nasal position. Not only do we lack evidence for the stage with the two dialects, but this assumption also fails to explain the connection between PND and the development of voiced stops to voiced fricatives in the elsewhere condition.

Blust (Reference Blust2005, Reference Blust2013) offers three possible explanations for the emergence of PND in Buginese, Murik, and Land Dayak. First, he notes that, much as PNV can be understood as an assimilation of stops to a voiced environment, PND can be explained as dissimilation. This assumption, however, lacks explanatory power: it simply restates that PND is the opposite process from PNV and does not specify whether such dissimilation would be driven by perceptual or other factors. Blust (Reference Blust2013: 668) himself notes that ‘this does little to explain why a change of this type would occur’.

Blust’s second explanation for Austronesian PND postulates that the three languages in question first underwent PNV: voiceless stops became voiced in post-nasal position, thus eliminating NT sequences. According to Blust (Reference Blust2013: 669), after PNV, ‘voice was free to vary’ post-nasally, and the ‘voiceless variant of post-nasal obstruents prevailed over time’. In other words, Blust (Reference Blust2005) seems to suggest that the lack of contrast in [ $\pm$ voice] enables devoicing in this position. There are three major issues with this approach. First, it is not parsimonious to assume the independent occurrence of PNV three times without any comparative evidence. Second, it is difficult to explain why a voiceless variant would prevail in an environment in which voicing is preferred compared to other environments (e.g. word-initially), but would fully contrast in those other environments, where voicing is dispreferred. While it is true that the functional load/frequency of phonemes can influence the probability of a merger (Wedel Reference Wedel2012, Wedel, Kaplan & Jackson Reference Wedel, Kaplan and Jackson2013, Hay et al. Reference Hay, Pierrehumbert, Walker and LaShell2015), it is unclear why a merger would first happen to voiced stops and then to voiceless stops.Footnote ^[12] Finally, this line of reasoning, like many others proposed thus far, fails to explain some common patterns, such as fricativization in the elsewhere condition or the presence of unconditioned devoicing in related languages/dialects (e.g. in Tswana, Section 3.1).

Somewhat related to Blust’s (Reference Blust2005) second proposal is the idea that the loss of contrasts could influence the operation of sound change (cf. Keyser & Stevens Reference Keyser, Stevens and Kenstowicz2001, Hyman Reference Hyman2013). PND is, however, attested both in languages in which [ $\pm$ voice] likely does not contrast post-nasally (e.g. in Yaghnobi) as well as in languages in which [ $\pm$ voice] fully contrasts post-nasally at the time of devoicing (e.g. in Konyagi or Tswana; see Table 2 and Section 3.3.2). Similarly, an analysis of devoicing as a loss of contrast in [ $\pm$ voice] in the elsewhere condition is not appealing. Devoicing is attested in languages in which voiced stops fricativize elsewhere (which means that the contrast is that of T  ${\sim}$  Z, e.g. in the Austronesian languages) and there are no voiceless fricatives (S) in the systems, meaning [ $\pm$ voice] could be redundant (and [ $\pm$ continuant] contrastive), as well as in languages in which the contrast in the elsewhere position is between voiceless and voiced fricatives (S  ${\sim}$  Z, e.g. in Konyagi or Tswana) and stops surface only post-nasally (and contrast in voice), meaning [ $\pm$ voice] cannot be redundant. Finally, the fact that complete devoicing occurs in some subdialects of Tswana without any contrast being lost beforehand and that in Yaghnobi devoicing targets all stops, not only post-nasal ones, speaks against this line of reasoning (see also Sections 3.3.1 and 3.3.2 for further discussion).

A third explanation offered in Blust (Reference Blust2005) invokes dissimilation by hypercorrection. Blust notes that NT sequences in Buginese and Murik develop either to T or TT. This means that, at a certain point, NT sequences were absent from the language and only voiced stops surfaced after nasals (ND). At this point, according to Blust (Reference Blust2005: 261), speakers ‘may have assumed that pre-nasalized obstruents had acquired voicing by assimilation’ and then ‘undid’ that assumed voicing. This account faces three major difficulties. As already pointed out by Blust (Reference Blust2005: 261), it is unclear what would ‘prompt speakers to assume that voicing assimilation had taken place in earlier clusters’ of ND. Second, even if they had made this assumption, the speakers would still have to apply dissimilation, i.e. devoicing in a context that strongly promotes voicing compared to other positions. Finally, this particular approach with hypercorrection lacks broader explanatory power, since it cannot be extended to cases of apparent PND in other languages, where the sound change NT  ${>}$  TT, T is not attested (e.g. in Tswana).

3.2.2 Phonetic motivation

Some analyses have attempted to account for PND by motivating the process phonetically. Solé et al. (Reference Solé, Hyman and Monaka2010: 612) specifically identify PND as a ‘historical process’, suggesting that they assume PND operated as a single instance of sound change. Moreover, these authors claim that PND is not necessarily an unnatural process and may in fact have a phonetic explanation. The main evidence for this claim comes from Shekgalagari, which is assumed to feature ‘early velic rising’ (Solé et al. Reference Solé, Hyman and Monaka2010: 612) in NT sequences. This process is supposed to follow from the fact that (i) speakers do not show any passive voicing in the NT sequences in Shekgalagari, and (ii) underlying nasal-fricative sequences /nz/ yield a nasal affricate [nts]. This process of early velic rising, which is argued to account for both of these observations, would also have caused a ‘long stop closure’ in ND sequences. Because voicing is difficult to maintain, especially during longer closure, the result would be devoicing of the stop (Solé et al. Reference Solé, Hyman and Monaka2010: 612).

This explanation has three major drawbacks. First, secondary ND sequences surface as NN in Tswana and ND in Shekgalagari. The two examples in Table 11 (repeated from Table 3) illustrate this distribution.

Table 11

Secondary ND sequences in Tswana and Shekgalagari (table from Solé et al. Reference Solé, Hyman and Monaka2010).

If early velic rising in Shekgalagari were indeed a phonetic process, we should expect to see devoicing in secondary ND sequences as well. The fact that the stops in secondary ND sequences surface as voiced speaks against the proposal in Solé et al. (Reference Solé, Hyman and Monaka2010) and Solé (Reference Solé, Solé and Recasens2012). Of course, one could assume that early velic rising operated prior to the period during which secondary ND sequences arose. However, there is a flaw in this assumption: Solé et al. (Reference Solé, Hyman and Monaka2010) provide evidence for early velic rising from synchronic phonetic data. If we postulate that early velic rising is responsible for PND as a synchronic phonetic process, we should expect secondary sequences to undergo devoicing as well. Conversely, if we posit that early velic rising should have been completed by the time secondary ND sequences were introduced, we should not expect to find continuing evidence for this process in the current phonetic data. The only reasoning that could explain why secondary ND sequences do not undergo devoicing under the early velic rising approach would be to assume that devoicing is blocked in order to prevent the merger of two grammatical morphemes, /m-/ and /mʊ-/. Even if this is indeed the case, it still means that early velic rising is not completely automatic in synchronic Shekgalagari, but can be overridden for functional reasons. Also, Solé et al. (Reference Solé, Hyman and Monaka2010) note that Shekgalagari has ND sequences in approximately ten items that do not arise from the /mʊ-/ morpheme. While ‘most’ of the ten items are assumed to be borrowings, it is unclear if all items indeed are borrowings (at least one of them is a grammatical marker [ǹde-]). This reinforces the view that early velic rising is not a completely automatic phonetic tendency in Shekgalagari, but can be blocked by at least functional and loanword factors. Second, realization of /nz/ as [nts] is typologically common and proceeds from general phonetic tendencies cross-linguistically (cf. Ali, Daniloff & Hammarberg Reference Ali, Daniloff and Hammarberg1979, Steriade Reference Steriade, Huffman and Krakow1993, Ohala Reference Ohala1997, Busà Reference Busà, Solé and Speeter2007, Kümmel Reference Kümmel2007; see also Section 3.4). The fact that Shekgalagari /nz/ surfaces as [nts] hardly tells us anything about early velic rising. Third, an explanation along these lines fails to account for other cross-linguistic cases of apparent PND where no traces of early velic rising can be found. If PND were indeed motivated by early velic rising, we should expect to find evidence that early velic rising is similarly responsible for reported cases of PND outside Tswana and Shekgalagari (to my knowledge, there exist no studies of this phenomenon). Finally, the explanation proffered in Solé et al. (Reference Solé, Hyman and Monaka2010) fails to explain the connection between PND and the recurrent pattern of fricativization in the elsewhere condition or the pattern of unconditioned devoicing of voiced stop found in subdialects of Tswana (see Section 3.3.2 and Table 16).

3.2.3 Perceptual enhancement

This paper also speaks to a long-standing discussion on the role of production vs. perception in sound change and phonology in general. One of the approaches that understands PND as a natural process is the perceptual account outlined in Kirchner (Reference Kirchner2000) and Stanton (Reference Stanton2016b). Post-nasal devoicing can be analyzed as a contrast enhancement with the following reasoning: sequences of a nasal and a voiced stop (ND) are perceptually very close to plain nasal stops (N). In order to enhance the perceptual contrast between N and ND, speakers can devoice stops in the latter sequence to NT. There is no doubt that perceptual (or auditory in Garrett and Johnson’s Reference Garrett and Johnson2013 terms) enhancement plays a role in sound change and phonology (for an overview, see Garrett & Johnson Reference Garrett and Johnson2013, Garrett Reference Garrett, Bowern and Evans2015, Fruehwald Reference Fruehwald2016). For example, lip protrusion in [ʃ] has no articulatory grounding, but a clear perceptual one: lip protrusion results in lowering of the frequency peak, which in turn results in perceptual enhancement of the contrast [ʃ]  ${\sim}$  [s]. The majority of cases of perceptual enhancement discussed in Garrett & Johnson (Reference Garrett and Johnson2013) or Garrett (Reference Garrett, Bowern and Evans2015), however, involve phonetic properties of phonemes in all positions. Additionally, I am not aware of cases of perceptual enhancement that would operate in the unnatural direction, as would be the case for PND. Perceptual factors usually play a role in enhancing a contrast that already has some underlying articulatory motivation and consequently underlying phonetic motivation (e.g. vowel lengthening before voiced stops in English; de Jong Reference de Jong1991, Reference de Jong2004; Solé Reference Solé, Solé and Beddor2007; Beguš Reference Beguš2017).Footnote ^[13] Below, I argue that the unnatural PND does not result from perceptual enhancement, but from a combination of three natural sound changes. To be sure, this does not suggest that perceptual enhancement is not a possible mechanism for sound change, but only that perceptual enhancement is likely not the underlying cause of PND.

Two studies of apparent PND have recently argued that the process is perceptually motivated. Stanton (Reference Stanton2016b: 1106) identifies a number of languages that devoice sequences of ND to NT word-finally, but not word-medially: Neverver, Kobon, Naman, Avava, Páez, and Tape. This distribution can be analyzed as contrast enhancement word-finally, where the contrast is additionally reduced. Lapierre (Reference Lapierre2018) similarly argues that devoicing of ND to NT in Panará can be motivated as perceptual enhancement of the contrast between the following nasal and oral vowels.

These proposals, however, face problems. First, it needs to be stressed that all languages with alleged PND in Stanton (Reference Stanton2016b) and Lapierre (Reference Lapierre2018) lack plain voiced stops in their inventories altogether: all voiced stops are pre-nasalized (ND) and there is no series of plain voiced stops (D). The reported PND is thus better analyzed as unconditioned (or word-final) devoicing of voiced stops and as such does not qualify as a true case of PND: it is not the case that voiced stops devoice post-nasally and not elsewhere. For example, the languages with reported PND in word-final position in Stanton (Reference Stanton2016b: 1106) (Neverver, Kobon, Naman, Avava, Páez, and Tape) lack plain voiced stops completely. Final   can thus simply be analyzed as devoicing of voiced stops in word-final position, which is a phonetically well-motivated development and does not lend support to the perceptual enhancement explanation.

The Panará example is even more relevant to our discussion (Lapierre Reference Lapierre2018). Panará and Timbira lack phonemic voiced stops altogether: the inventory only features voiceless stops and nasals. Nasals (N) and pre-nasalized stops (ND) are in complementary distribution in Proto-Jê: before an oral vowel, the nasal stop /N/ is realized as [ND]. Proto-Jê ND surfaces as devoiced NT in Panará and Timbira (Lapierre Reference Lapierre2018). The Panará example rules out devoicing as an enhancement between N and ND: in Pre-Panará, N and ND do not contrast, but are in allophonic distribution (one before nasal vowels and the other before oral vowels), which means there is no contrast between N and ND to be enhanced. It is true that devoicing can still be analyzed as enhancement of the contrast between the following nasal and oral vowel (as proposed in Lapierre Reference Lapierre2018). Herbert (Reference Herbert1986) (via Stanton Reference Stanton2018b), Stanton (Reference Stanton2018b), and others propose that shielding (e.g. NV  $\rightarrow$  NDV) is contrast enhancement – but oral closure in ND is sufficient that the nasalization would not spread into the following vowel. The perceptual analysis should show that devoicing (or longer oral release along the lines of Stanton’s Reference Stanton2018a explanation) increases the contrast between the following V and Ṽ. The mechanisms for this enhancement are not immediately obvious and experimental evidence is, to my knowledge, lacking. The alternative proposal where devoicing of ND in Panará and Timbira is analyzed as unconditioned devoicing of voiced stops is strongly supported by comparative evidence. All voiced stops (both plain and pre-nasalized) in other closely related languages devoice unconditionally (either simultaneously or D devoices before ND) (Nikulin Reference Nikulin2017). In other words, unconditioned devoicing of voiced stops operated as a sound change in the pre-history of Panará and Timbira. Devoicing of the pre-nasalized voiced series (ND) is thus just an extension of such a sound change. In Section 3.3, I will propose that such unconditioned devoicing of stops is precisely what underlies all thirteen cases of PND in which voiced stops seemingly devoice only post-nasally and are preserved elsewhere.

The fact that in some varieties of Setswana, PND involves not only devoicing but also ejection likewise fails to support the enhancement explanation. Gouskova et al. (Reference Gouskova, Zsiga and Boyer2011) observe that PND in Setswana occasionally results in ejective stops. This distribution would be in line with the perceptual enhancement approach: to maximize the contrast, the speaker can use ejection as an additional cue for the original N  ${\sim}$  ND contrast. While it is true that some stops in Setswana are realized with ‘weak’ ejection, this does not necessarily point to perceptual enhancement either. Setswana has three series of stops: voiced (D), voiceless aspirated  and voiceless unaspirated that are in variation with ejective articulation (T/T’). While at first glance, the alternation between voiced stops elsewhere and voiceless unaspirated/ejective (T/T’) in post-nasal position appears to be fortition, it can also be analyzed simply as devoicing and is actually expected under the historical approach presented below. From the data described in Gouskova et al. (Reference Gouskova, Zsiga and Boyer2011), it appears that the ratio of voiceless unaspirated vs. ejective articulation in the T/T’-series of stops is not much higher in post-nasal position than in the elsewhere condition. For example, Gouskova et al. (Reference Gouskova, Zsiga and Boyer2011) count the number of recorded tokens that have ambiguous acoustics (can be interpreted as either ejective or voiceless aspirated) and the number of unambiguous ejectives. In post-nasal position, the ratio is 41 (ambiguous) to 60 (ejective), and in intervocalic position 43 (ambiguous) to 54 (ejective). This means that ejection in post-nasal position can be analyzed as part of the common trend in Setswana to realize unaspirated voiceless stops (and in our case, devoiced stops) with ejection. In fact, I will argue below that PND in Tswana results from a combination of sound changes (following Dickens Reference Dickens1984, Hyman Reference Hyman, Hume and Johnson2001), one of which is the devoicing of all voiced stops. Because voiced stops have significantly less aspiration than voiceless aspirated stops, we expect a devoiced variant of an unaspirated voiced stop to be phonetically closest to the voiceless unaspirated series (T). This series of stops can then get realized as ejective (either when devoicing occurs or after), which would explain the Setswana distribution without resorting to post-nasal fortition or contrast enhancement.

There exists some additional evidence against the perceptual enhancement explanation of PND. First, there are currently no experimental studies supporting the claim that NT is perceptually more salient than N or ND in intervocalic position. Kaplan (Reference Kaplan2008) is the only study known to me that tests this contrast perceptually and does not limit it to final position (Katzir Cozier Reference Katzir Cozier2008). While NT is more salient than N and ND word-finally,Footnote ^[14] no such significant effect has been found word-medially: all three stimuli, N, NT, and ND, were perceived with equal rates (Kaplan Reference Kaplan2008). While this result could also be due to a ceiling effect, it nevertheless shows that the contrast between ND/N and NT is substantially more salient in intervocalic position than word-finally. This means that, according to the perceptual enhancement approach, we would expect many more cases of word-final PND than of unrestricted PND. As already mentioned, all cases in Stanton (Reference Stanton2016b) are better analyzed as word-final stop devoicing and, to my knowledge, no cases of true word-final PND are attested (where voiced stops remain voiced word-finally except if preceded by a nasal). Second, PND is not attested as a repair strategy even in cases where we should expect it. For example, many languages disallow NC $_{1}$ VNC $_{2}$ sequences. A recent study in Stanton (Reference Stanton, Hammerly and Prickett2016a, Reference Stanton2018a) suggests that avoidance of these sequences constitutes a strategy to avoid a difficult contrast with NVNC $_{2}$ . The vowel in NC $_{1}$ VNC $_{2}$ is universally phonetically nasalized, a process which reduces cues for the contrast between NC $_{1}$ and N. One way to repair this contrast would be to devoice the first consonant C $_{1}$ . However, the survey in Stanton (p.c.) shows that NDVNC  ${>}$  NTVNC is not attested. It is unclear to me how this gap is explained within the perceptual enhancement approach. Finally, while perceptual enhancement might explain PND, it cannot, to my knowledge, explain other unnatural processes discussed elsewhere (especially intervocalic devoicing; Beguš Reference Beguš2018).

As already mentioned, this paper does not argue against the existence of perceptual enhancement as a sound change in general. A number of arguments presented here, however, favor the three-sound-changes approach over the perceptual enhancement approach when dealing with PND. It seems suspect to suggest that a contrast will only be enhanced when a set of three sound changes happen to operate in the pre-history of a system. Future research involving processes other than PND should show whether other unnatural processes are better explained by perceptual enhancement or the three-sound-change approach called the Blurring Process proposed here.

While I am claiming that PND did not arise as contrast enhancement, I do, however, acknowledge the possibility that perceptual factors play a role in the phonologization of PND and the preservation of the alternation once it arises. It is thus possible that perceptual enhancement helps PND survive as a productive process longer than would be expected otherwise.

3.2.4 Multiple sound changes

Finally, Dickens (Reference Dickens1984) and Hyman (Reference Hyman, Hume and Johnson2001) propose an explanation for PND in Tswana that assumes a set of three non-PND sound changes that conspire to produce apparent PND: fricativization, devoicing of stops, and occlusion of fricatives. I will argue in the remainder of this paper that this is in fact the correct explanation not only for Tswana and Shekgalagari, but that an essentially similar historical scenario played out in all thirteen cases of reported PND described above. Unfortunately, at the time of Dickens’ and Hyman’s work, no historical parallels existed in the literature that would support their explanation, which led other authors to propose alternative accounts of the data (see the three types of explanations in Sections 3.2.1, 3.2.2, and 3.2.3 above). Admittedly, in the absence of typological parallels, one might judge an explanation that operates with a single sound change (motivated by perceptual enhancement, articulatory factors, or hypercorrection) more parsimonious and justified than an explanation that requires three separate sound changes. By bringing numerous cases of PND from disparate language families together and arguing that diachronic developments very similar to that proposed for Tswana in Dickens (Reference Dickens1984) and Hyman (Reference Hyman, Hume and Johnson2001) operated in all thirteen languages, the next section dispels this concern and validates the three-sound-change analysis on typological grounds. In addition, I provide crucial new evidence in favor of the three-sound-change approach and argue that the three sound changes are directly historically attested in Avestan, Sogdian, and Yaghnobi, three languages in an ancestral relationship.

3.3.1 Yaghnobi

Yaghnobi is a descendant of Sogdian, a Middle Iranian language spoken in the first millennium CE and preserved in documents from that period.Footnote ^[15] In Sogdian, all voiced stops surface as voiced fricatives except post-nasally. This complementary distribution is directly attested in Sogdian and confirmed by the writing system (cf. Sims-Williams Reference Sims-Williams and Schmitt1987: 178). It is equally clear that the Sogdian pattern developed through a sound change  in an earlier stage of the language. In Avestan, which is closely related to Sogdian (but more archaic) and can be used to represent the parent language of Sogdian, voiced stops correspond to Sogdian voiced fricatives except post-nasally, e.g. Avestan [dasa] vs. Sogdian [ðasa] ‘ten’, Avestan [gari-] vs. Sogdian [ɣarí] ‘mountain’ (Kümmel Reference Kümmel2006), Avestan [asənga-] vs. Sogdian [sang] ‘stone’ (Bartholomae Reference Bartholomae1961: 210).Footnote ^[16]

Given the complementary distribution attested in Sogdian, I propose that, on the way to Yaghnobi, an additional sound change operated: unconditioned devoicing of voiced stops (D  ${>}$  T; for phonetic motivation and attestations of this change, see Section 3.4). Evidence for this sound change in Yaghnobi is attested as directly as a sound change can be attested diachronically: Stage 2 (Sogdian) which is attested in written sources features voiced stops which surface as voiceless at Stage 3 (Yaghnobi). Because voiced stops surface after nasals, this combination of sound changes results in apparent PND. The development is summarized in Table 12.

Table 12

Development of PND from Avestan to Yaghnobi.

Sogdian thus provides direct historical evidence showing that the apparent case of PND in Yaghnobi is a side effect of two natural and well-attested sound changes: (i) fricativization of voiced stops except in post-nasal position  and (ii) unconditioned devoicing of voiced stops (D  ${>}$  T).

To get PND, we need a third sound change: occlusion of voiced fricatives to stops (Z  ${>}$  D). Yaghnobi provides additional evidence for this process too: the original voiced labial and velar stops (Avestan [b] and [g] that develop to Sogdian [β] and [ɣ] in the elsewhere position) still surface as voiced fricatives [v] and [ɣ] in the elsewhere position in modern Yaghnobi (e.g. Yagh. [vant] ‘tie’ from Sogd. [βand]; Yagh. [ɣar] ‘mountain’ from Sogd. [ɣarí]). Nevertheless, the voiced dental fricative (Sogdian [ð]) in the elsewhere position gets occluded in Yaghnobi and surfaces as a voiced stop [d] (Xromov Reference Xromov1972: 123), thus blurring the original complementary distribution. Apparent PND in Yaghnobi thus fully holds only for the dental series of stops, because only this series of stops underwent a sound change that turned the original voiced fricatives ‘back’ to stops (Z  ${>}$  D). Table 13 illustrates the three sound changes that operated on the dental series of stops to produce PND.

Table 13

Development of coronals from Avestan to Yaghnobi.

Yaghnobi and Sogdian contain yet more crucial evidence in favor of the three-sound-change approach outlined above. One of the most serious objections to the three-sound-change approach in Yaghnobi and elsewhere is that devoicing should not be analyzed as unconditioned, but as post-nasal – precisely because it operates only in post-nasal position. This being the case, devoicing should indeed be considered unnatural according to the definition in (2). However, Sogdian and Yaghnobi provide strong and unique evidence to suggest that devoicing of voiced stops is in fact unconditioned. I show below that devoicing targets all surface voiced stops in Yaghnobi, not just those in post-nasal position.

It is true that voiced stops surface primarily in post-nasal position in Sogdian, but they also surface in one additional, more marginal context: after voiced fricatives. The origin of such a distribution is straightforward: original voiced stops did not fricativize (i) after nasals (as already mentioned above) and (ii) after another fricative in order to avoid clusters of two fricatives. This property of Sogdian and Yaghnobi, which has up to now been a less discussed fact (Novák Reference Novák2013), essentially means that voiced stops surface as fricatives in Sogdian in all positions except post-nasally (ND) and occasionally after a voiced fricative (ZD). We already saw that ND sequences devoice to NT in Yaghnobi. Crucially, Sogdian sequences of a voiced fricative and a voiced stop (ZD), too, devoice in the development from Sogdian to Yaghnobi – to ST (in central and western dialects),Footnote ^[17] e.g. Sogd. [pəzda], Yagh. [past] or [pajst] ‘smoke’; potentially also Sogd.  Yagh. [duxtar]; Sogd. [suɣd-], Yagh. [suxta]; Sogd. [əχʃɨβdi], Yagh. [xiʃift]Footnote ^[18] (from Novák Reference Novák2013, Reference Novák2014 and Xromov Reference Xromov1987). This devoicing is automatically explained by my proposal that voiced stops devoice unconditionally in Yaghnobi. ZD is thus predicted to first develop to ZT and then to ST when the voiced fricative assimilates in voicing to the following voiceless stop. That this scenario indeed took place is suggested by evidence from Western Yaghnobi, in which forms like [avd] and [avt] for Sogd. [əβd´a] are attested (data from Novák Reference Novák2014). While the first form [avd] dialectally lacks devoicing, the second form [avt] shows that only voiced stops devoice: the fricative is not yet assimilated to the following voiceless stop (compare the assimilated form from Eastern Yaghnobi [aft]).

Conversely, my model proposes voiced fricatives and consequently clusters of voiced fricatives do not devoice. This proposal is borne out by the data: Sogdian clusters of two voiced fricatives (that arise from a cluster of a voiced fricative and an approximant [dw]), [ðβ], do not devoice in Yaghnobi:Footnote ^[19] Sogd. [ðβəri] ‘door’ and [əzβa:k] ‘tongue, language’ yield Yagh.  and  (Novák Reference Novák2013). Table 14 illustrates the development of voiced fricative–stop and fricative–fricative sequences.

Table 14

Correspondences of ZD and ZZ sequences.

Sequences of a voiced fricative and a voiced stop ZD that devoice in Yaghnobi to ST strongly support the claim that voiced stops undergo unconditioned (not only post-nasal) devoicing in Yaghnobi: all voiced stops that surface in the Sogdian and pre-Yaghnobi system get devoiced in the development of Yaghnobi. The development of ZD clusters in Yaghnobi is summarized in Table 15.

Table 15

Development of ZD sequences from Sogdian to Yaghnobi.

In the other twelve languages with PND, non-nasal clusters are generally not allowed (or they became simplified before the emergence of PND), so we do not see devoicing anywhere other than in post-nasal position. Devoicing in clusters in Yaghnobi thus offers a crucial piece of evidence in favor of the proposal that the devoicing of voiced stops that occurs in the development of PND is unconditioned; if it were not, we would not be able to unify our account of the devoicing of ND and ZD.

3.3.2 Tswana, Shekgalagari, and Makhuwa

If Yaghnobi provides crucial evidence for the three-sound-change approach on historical grounds, Tswana offers crucial dialectal evidence for this hypothesis. There are at least three different systems of stops in the microdialects of Tswana. Among one set of speakers, voiced stops get devoiced in all environments: no voiced stops are allowed in the system. Speakers of this system have been labeled ‘devoicers’ (Coetzee, Lin & Pretorius Reference Coetzee, Lin, Pretorius, Trouvain and Barry2007: 861). Another set of speakers changes voiced stops into fricatives in all positions but post-nasally (these speakers are called ‘leniters’; Coetzee et al. Reference Coetzee, Lin, Pretorius, Trouvain and Barry2007: 861). A third set of speakers use the so-called PND system: for these speakers, voiced stops surface as voiceless only post-nasally (Zsiga, Gouskova & Tlale Reference Zsiga, Gouskova, Tlale, Davis and Rose2006, Coetzee et al. Reference Coetzee, Lin, Pretorius, Trouvain and Barry2007, Coetzee & Pretorius Reference Coetzee and Pretorius2010). The three systems of Tswana are represented in Table 16.

Table 16

Microdialects of Tswana (from Zsiga et al. Reference Zsiga, Gouskova, Tlale, Davis and Rose2006, Coetzee et al. Reference Coetzee, Lin, Pretorius, Trouvain and Barry2007).

As Hyman (Reference Hyman, Hume and Johnson2001) argues, the PND system arises precisely through the combination of two other (devoicing and leniting) systems: leniters take on fricativization except after nasals  while devoicers undergo unconditioned devoicing (D  ${>}$  T). Following Dickens (Reference Dickens1984), Hyman (Reference Hyman, Hume and Johnson2001) argues that post-nasal devoicers undergo both sound changes. In other words, Dickens (Reference Dickens1984) and Hyman (Reference Hyman, Hume and Johnson2001) argue that PND in Tswana results from a combination of three sound changes, the first of which is fricativization of voiced stops except post-nasally (the sound change that operates in the leniters’ system). Unconditioned devoicing happens next (the sound change that operates in the devoicers’ system), and because voiced stops surface only after nasals, the result is apparent PND. This interstage is confirmed by Kutswe, a variety of Sotho in which the alternation is between [p] post-nasally and [β] elsewhere (as mentioned in Dickens Reference Dickens1984). This pattern is obscured in Tswana and Shekgalagari, however, by an additional change in the dialect with PND: unconditioned occlusion of fricatives (Z  ${>}$  D). After this change, voiced stops surface as voiceless after nasals and, crucially, the feature [ $\pm$ voice] is fully contrastive in the elsewhere condition (Hyman Reference Hyman, Hume and Johnson2001). Recall that this final sound change also occurred in Yaghnobi, but only for the dental series of stops.

The attestation of the subdialect of Tswana that has unconditioned devoicing of voiced stops also speaks against analyses that assume PND results from the reduced functional load of the contrast in the feature [ $\pm$ voice] in post-nasal stops, when voiced stops develop to voiced fricatives elsewhere (see Section 3.2). In the ‘devoicer’ subdialect (Coetzee et al. Reference Coetzee, Lin, Pretorius, Trouvain and Barry2007: 861), unconditioned devoicing occurs without the development of voiced stops to fricatives (i.e. when the feature [ $\pm$ voice] is fully contrastive).

The pattern of development of Tswana voiceless stops provides a structural parallel that speaks in favor of the proposed explanation. As Hyman (Reference Hyman, Hume and Johnson2001) points out, voiceless stops underwent fricativization (except post-nasally) along with voiced stops. Table 17 shows this development.

Table 17

Development of voiceless stops in Tswana with examples (table from Hyman Reference Hyman, Hume and Johnson2001).

The data in Table 17 provide yet another piece of evidence that complementary distribution occurred first in the development of PND in Tswana and related dialects, in both the voiced and voiceless series of stops (Hyman Reference Hyman, Hume and Johnson2001). The voiceless and voiced series underwent lenition except in post-nasal position (as in the leniters dialect), and then voiced stops underwent further changes (unconditioned devoicing, as in the devoicers dialect) to produce PND, whereas voiceless stops retained the complementary distribution (Table 17).Footnote ^[20]

3.3.3 Bube and Mpongwe

Despite the fact that Bube and Mpongwe are not closely related to Tswana, Shekgalagari, and Makhuwa, we find striking similarities between these languages that uniformly point to the conclusion that complementary distribution (with lenition of voiced stops except post-nasally) operated in the pre-history of Bube and Mpongwe, too. The data for the two languages, however, is sparse and detailed phonetic descriptions are lacking.

Several indicators in Bube data clearly point to a pre-stage with complementary distribution. Janssens (Reference Janssens1993: 37) reports that in the labial series, the voiced stop and voiced fricative [b] and [β] are in free variation in medial position. Also, exactly parallel to Tswana, the voiced velar stop is lost in Bube, pointing indirectly to an interstage with [ɣ] (also reconstructed independently in Janssens Reference Janssens1993: 27). Finally, the dental series of stops undergoes lenition as expected: in the elsewhere condition, [d] develops to [l] or [r] (likely through an interstage *[ð]), depending on the vowel quality, e.g.  (Janssens Reference Janssens1993: 23). In fact, Janssens independently reconstructs a proto-stage of Bube with exactly the complementary distribution we observe in other languages with PND: voiced stops surface as fricatives except post-nasally.

In Mpongwe, PND that targets only the velar series also arises through a combination of three sound changes. Mouguiama-Daouda (Reference Mouguiama-Daouda1990) independently reconstructs a stage in which [g] surfaced as a voiced velar fricative [ɣ] except post-nasally, citing as evidence the fact that the voiced velar stop is still realized as [ɣ] by some older speakers (Mouguiama-Daouda Reference Mouguiama-Daouda1990). PND in Bube and Mpongwe thus follows the usual trajectory: complementary distribution and unconditioned devoicing of stops that surface only post-nasally.

3.3.4 Konyagi

Konyagi and related languages, too, provide strong evidence in favor of a stage with complementary distribution, just like in Yaghnobi or Tswana. Proto-Tenda, an ancestor of Konyagi, is reconstructed on the basis of Konyagi and its neighboring languages Basari and Bedik (Merrill Reference Merrill2014, Reference Merrill2016a, Reference Merrillb). The first stage of Proto-Tenda features a phonemic inventory with voiced and voiceless stops and fricatives. In Stage II, all stops fricativized everywhere but in post-nasal position. This fricativization in Proto-Tenda was quite radical, targeting both voiced and voiceless stops, as well as nasal stops (Merrill Reference Merrill2014, Reference Merrill2016a, Reference Merrillb) and, as I reconstruct, geminates/clusters.Footnote ^[21] The fricativization is directly confirmed by the Basari dialect, which preserves the Proto-Tenda II stage with minor deviations. Table 18 shows the development of consonants in non-post-nasal position from Proto-Tenda I and II to Basari and Konyagi. Just like in Basari, original voiced stops surface as fricatives in Konyagi (further developing to sonorants in some places of articulation), except after nasals.

Table 18

Development of non-post-nasal consonants from Proto-Tenda to Konyagi (table from Merrill Reference Merrill2014, Reference Merrill2016a, Reference Merrillb; Santos Reference Santos1996).

As already mentioned, Proto-Tenda voiced stops remain stops in post-nasal position. Table 19 summarizes the development of consonants in post-nasal position in the descendants of Proto-Tenda (from Merrill Reference Merrill2014, Reference Merrill2016a, Reference Merrillb, p.c.).Footnote ^[22] Post-nasally, no fricatives occur in either Konyagi or Basari.

Table 19

Development of post-nasal consonants from Proto-Tenda to Konyagi (table from Merrill Reference Merrill2014, Reference Merrill2016a, Reference Merrillb; Santos Reference Santos1996).

While in Bedik and Basari, original post-nasal voiced stops remain voiced, they devoice in Konyagi. Here too, however – like in all other cases of PND – the voiced stops that devoice post-nasally were in fact the only voiced stops in the language (the stage confirmed in today’s Basari). Thus, again, the sound change operating in Konyagi is in fact unconditioned devoicing of voiced stops. The apparent PND is once more the result of a combination of sound changes.

The third sound change, occlusion back to stops, is lacking in Konyagi. As a result, it is not voiceless and voiced stops that alternate at some stage of development in Konyagi, but rather voiced fricatives and voiceless stops. Konyagi, however, does not lack voiced stops completely: original fricative geminates were later occluded to stops and simplified, resulting in voiced stops being reintroduced into the synchronic inventory.Footnote ^[23] All these changes result in a synchronic mutation system that involves voiced and voiceless stops and voiced fricatives or approximants (see Section 3.5).

So far, I have argued that all cases of apparent PND can be accounted for through a combination of two or three well-motivated sound changes. I now turn to a case of PND from the South Italian dialects to illustrate that such sound change combinations involving complementary distribution are not limited to stops, but can apply to other segments as well.

3.3.5 South Italian

On the surface, the data in South Italian suggest that *ʤ devoices to [ʧ] only in post-nasal position (Section 3.1.5). However, if we look at the development of *ʤ elsewhere, we observe that it gets de-occluded and further develops to [j] except after nasals (e.g. *faʤina  ${>}$ Calabrian [fajina], *leʤere  ${>}$ Calabrian [lejere], Sicilian [lejiri]; Rohlfs Reference Rohlfs1949: 358). Given the cases of PND discussed so far, de-occlusion of the affricate and the development to the glide provides evidence for a stage with complementary distribution. At the point when *ʤ appears post-nasally and [j] (or probably *ʒ) elsewhere, an unconditioned devoicing of voiced affricates occurs. This, too, is a well-attested and motivated sound change: voiced affricates are highly dispreferred, which is why devoicing is a natural tendency in these segments. Voice is difficult to maintain, especially in affricates, which combine two articulations that are highly antagonistic to voicing: closure and frication (Ohala Reference Ohala and MacNeilage1983, Reference Ohala and Brown2006). Table 20 illustrates the reconstructed development: Stage 2 shows a period of complementary distribution, and Stage 3 represents the development after the unconditioned devoicing of voiced affricates.

Table 20

Devoicing of post-nasal affricates in South Italian.

Note that this set of sound changes is in principle the same as in the previous cases, but here complementary distribution targets affricates instead of stops. Also, Stage 4 is absent from South Italian: the change that would occlude [j] to [ʤ] is absent.

3.3.6 Buginese and Murik

The emergent pattern that we have seen in all the apparent cases of PND so far can be generalized as follows: (1) a set of segments enters complementary distribution; (2) a sound change occurs that operates on the unchanged subset of those segments; (3) optionally, another sound change occurs that blurs the original complementary distribution.

Let us now turn to the three Austronesian languages. On the surface, the data from Buginese and Murik seem to point to PND operating as a single sound change. Moreover, there is no direct historical or dialectal evidence to suggest otherwise, as is the case for Yaghnobi, Tswana, and Konyagi. If the only attested instances of PND were those found in Austronesian languages, we would likely be forced to assume the operation of a single sound change – PND. However, these languages do, at least, show clear traces of a stage with complementary distribution. Below, I argue that the three-sound-change explanation again better captures the data, despite the lack of direct historical or dialectal evidence.

The main evidence against PND as a single sound change in Austronesian comes from the voiced labial stop in Buginese. Already in Proto-South-Sulawesi (PSS; from which Buginese developed), *b had developed to [w] except word-initially and post-nasally (Mills Reference Mills1975: 547). Later, the change *b  ${>}$  [w] also targeted the word-initial position, as is clear from initial stop in cases like *bumbun  ${>}$ [wumpun]. Thus, at one stage in the language’s development, voiced labial stops surfaced only post-nasally: again, we have clear evidence for complementary distribution. From there, the development followed the trajectory described above: unconditioned devoicing of voiced stops occurred, but produced apparent PND because voiced stops surfaced only post-nasally. The development is illustrated in Table 21.

Table 21

Development of PND in bilabials in Buginese.

In Buginese, /w/ continues to surface as a non-plosive (since only two changes operated in the labial series). Based on the development of the labial series, we can reconstruct that the velar series likewise undergoes complementary distribution and unconditioned devoicing. Unlike the labial series, however, the voiced velar fricative *ɣ undergoes occlusion to [ɡ] (all three sound changes operated in the velar series), thus obscuring evidence for an interstage with complementary distribution. Note that this is precisely the same scenario attested in Yaghnobi, with the only difference being that, in Yaghnobi, it was the dental series of fricatives that underwent occlusion, whereas in Buginese, it was the velar series. The development leading to apparent PND in velars (with all three sound changes operating) is illustrated in Table 22.

Table 22

Development of PND in velars in Buginese.

The dental series of stops in Buginese escapes PND because *d developed to [r] in all positions (*dindiŋ  ${>}$ [renriŋ]) and, as such, became ineligible for the devoicing of voiced stops. The development of *z also conforms to the proposal above: intervocalically, it undergoes rhotacism to [r] (possibly through an interstage with *ʝ); post-nasally, I reconstruct it occludes to *ɟ and devoices to [c] (according to the unconditioned devoicing of voiced stops which predictably targets the palatals); initially, it remains a fricative *ʝ and later occludes together with *ɣ to [ɟ] (in other words, there was likely a stage of complementary distribution: *ɟ post-nasally, *ʝ/r elsewhere).

The initial stage with complementary distribution in Buginese stops is additionally confirmed by a recent description of Buginese by Valls (Reference Valls2014) which reports that voiced stops [b, d, g] have voiced fricatives [β, ð/ɾ, ɣ] as allophones in apparent free variation in non-post-nasal position. This suggests that the last sound change that turns voiced fricatives to voiced stops is likely in progress in Buginese.

Just like in Buginese, evidence exists that complementary distribution underlies PND in Murik as well. The main evidence comes from the development of voiced alveolar stops. On the surface, Proto-Kayan-Murik *d surfaces as [l] initially, [r] intervocalically, and [t] post-nasally (see Table 23). These data point to a stage with complementary distribution (PKM in Table 23). I propose a historical development with the following trajectory: Proto-Kayan-Murik *d lenited to [l] or [r] (probably through an interstage with *ð) initially and intervocalically. After a nasal, however, *d remained a stop.

Table 23

Reconstructed development of alveolars in Murik.

Next, I argue that unconditioned devoicing of voiced stops occurred. Because devoicing operated during a period when voiced stops surfaced only after nasals, apparent PND is the result. Based on the development of the alveolar series, I reconstruct that fricativization targeted labials and velars too: original voiced stops fricativized to *β and *ɣ (except after a nasal), at which point voiced stops (surfacing only post-nasally) got devoiced (e.g. *b  ${>}$  [p], *g  ${>}$  [k]). Unlike the alveolars, *β and *ɣ then underwent occlusion ‘back’ to stops, resulting in apparent PND. Table 24 traces the proposed trajectory from (Pre-)Proto-Kayan-Murik to Murik for the labial series of stops.

Table 24

Reconstructed development of labials in Murik.

A peculiarity of the development in Murik is that it combines the PND of stops that we saw, for example, in Tswana, with the PND of affricates that we saw in Sicilian and Calabrian – i.e., whereas the previously discussed languages devoice either stops or affricates, Murik devoices both. The development of stops in this language is straightforward – it follows the usual trajectory of PND: complementary distribution, unconditioned devoicing, and then optional occlusion to stops (as we saw above). The development of affricates is more complicated, but nevertheless revealing. PMP *z develops to *s intervocalically already in Proto-Kayan-Murik (Blust Reference Blust2005); this development cannot be considered part of PND, because it happens at an earlier stage. Elsewhere, *z is preserved as voiced and palatalizes to *ʝ. Post-nasally, occlusion to an affricate *ɟʝ takes place.Footnote ^[24] The affricate then gets devoiced, exactly as in Sicilian and Calabrian, together with devoicing of other voiced stops. Because they surface only post-nasally at the time of unconditioned devoicing, the result is apparent PND. I further reconstruct that the initial fricative *ʝ then gets occluded to a voiced stop, together with other voiced fricatives in Murik.Footnote ^[25]

In sum, even though Buginese and Murik offer neither dialectal nor historical evidence for complementary distribution, there is enough language-internal evidence to posit that, at one stage, voiced stops (and affricates) surfaced only after nasals: the original voiced labial stops surface as fricatives even today in Buginese, whereas in Murik voiced alveolar stops surface as lenited [r] or [l], in accordance with the reconstructed complementary distribution.

3.3.7 Nasioi

That PND in Nasioi results from a combination of three natural sound changes is strongly suggested by Nasioi internal evidence and by diachronic development of its related languages. Evidence for a stage with complementary distribution comes from Nasioi itself: voiced stops /b/ and /d/ surface as a fricative [β] and a flap [ɾ] intervocalically. Lenition of voiced stops in related languages strongly suggests that a complementary distribution consisting of voiced stops developing to voiced fricatives except post-nasally, operated already in pre-Nasioi. In Buin and Nagovisi, for example, [ɾ] and [d] are in complementary distribution whereby [d] surfaces post-nasally and [ɾ] elsewhere (Brown Reference Brown2017). In Motuna, the complementary distribution is not limited to alveolars, but targets labials as well. The stops [b] and [d] are in complementary distribution with [w] and [r] such that the stops surface only post-nasally and the approximants surface elsewhere, including in word-initial position (Onishi 2012). Table 25 summarizes the development of Motuna voiced labial and alveolar stops.

Table 25

Summary of the development of Motuna labial and velar stops (Onishi Reference Onishi2011).

It is reasonable to assume that Motuna represents the proto-stage of Nasioi. In the development of Nasioi, voiced stops devoice, but because they surface only post-nasally, the change looks like post-nasal devoicing. Finally, Nasioi undergoes occlusion of fricatives to stops (like Tswana and Shekgalagari, Yaghnobi, and Buginese and Murik), but this occlusion is contextually limited to the initial and post-obstruent positions (see Section 3.1.7). Intervocalically, voiced fricatives still surface as fricatives. Occlusion of fricatives that is limited to initial position or to clusters is a well-motivated and common sound change (see Kümmel Reference Kümmel2007 and the discussion in Section 3.4 below).

4.2.1 Potential exceptions

There are two potential exceptions to the MSCR. First, unnatural alternations might arise from less than three sound changes when the change from $+\unicode[STIX]{x1D719}_{1}$ to $-\unicode[STIX]{x1D719}_{1}$ (Stage 2 to 3 in Table 28) automatically entails the change from $\neg \unicode[STIX]{x1D6FC}$ to $\unicode[STIX]{x1D6FC}$ in $\unicode[STIX]{x1D719}_{1+n}$ (Stage 3 to 4 in Table 28). In other words, if a change in one feature (such as a change from [ $-$ nas] to [ $+$ nas]) entails the change of another, automatic (redundant) feature (e.g. [ $-$ son] to [ $+$ son]), or if one of the changes in the Blurring Process is a perceptual change that can target two features simultaneously and precisely those two features play a role in the Blurring Process, then the unnatural alternation might arise from two changes, as we know that the minimality principle holds only for non-automatic non-redundant features and might be violated in some cases (see Section 2.2). We expect such cases, if they exist at all, to be extremely rare because of the restriction that the change of two features in the Blurring Process be automatic/redundant or simultaneous.

Second, the MSCR might not hold in rare cases in which a phonetic change allows a reinterpretation of the phonological status of certain segments. Such cases are, however, expected to be rare and, more importantly, probably always allow alternative analyses. Let us imagine a language in which voiceless stops shorten in word-final position or in coda position more generally. Post-vocalically, voiceless stops universally have some degree of voicing into closure as it is difficult to cease voicing immediately after the vowel (Westbury & Keating Reference Westbury and Keating1986, Davidson Reference Davidson2016). If the closure duration of such voiceless stops shortens significantly in the coda position, the percentage of voicing into closure will increase. Such shortened voiceless stops can be analyzed either as shortened voiceless stops or as voiced stops, because of the higher percentage of voicing into closure due to their shorter duration.Footnote ^[38] In such cases that do not undergo the Blurring Process, however, the conditioning factor for a change is still present in the synchronic system. In the presence of this conditioning factor, alternative analyses are possible. Finally, even if unnatural patterns can arise outside of the Blurring Process, they are comparatively rare, as this paper argues. All unnatural processes known to me for which no alternative explanations exist result from the Blurring Process (Beguš & Nazarov Reference Beguš and Nazarov2017, Reference Beguš and Nazarovto appear; Beguš Reference Beguš2018). Under this less strong position, then, the MSCR is at least a strong tendency and the rest of the argumentation that follow from the MSCR still holds.

The MSCR is also limited in scope to segmental phenomena. Prosodic/ suprasegmental processes cannot be represented by feature matrices as in Table 28 and require an independent tier under most approaches to prosodic phonology (Goldsmith Reference Goldsmith1976, Hayes Reference Hayes1995). Because the changes on a prosodic tier can be independent from changes in the segmental tier, unnatural prosodic processes can arise from less than three sound changes. Unnatural prosodic phenomena are beyond the scope of this paper. A more detailed investigation of what constitutes an unnatural change and whether the lack of the MSCR restriction in the prosodic domain translates into a higher rate of unnatural processes would be desired.

4.2.2 Phonotactic restrictions

It is important to note that fewer than three sound changes are required to produce unnatural static phonotactic restrictions. For instance, Kiparsky (Reference Kiparsky2006) provides several trajectories of two or more sound changes leading to the unnatural process of final voicing. Most of his scenarios, however, would result in a phonotactic restriction against voiceless stops word-finally and not in FV as an alternation. Even a single sound change might produce an unnatural phonotactic restriction: if a sound change targets voiceless stops to the exclusion of voiced stops, the resulting phonological system would allow only voiced stops in word-final position. Similarly, post-vocalic voicing of stops would result in a system that would restrict word-final voiceless stops and allow voiced stops in word-final position. While such a system would better be analyzed as a natural restriction against post-vocalic voiceless stops, another sound change reintroducing prevocalic voiceless stops into the system would create an unnatural phonotactic restriction where voicing of stops would be contrastive initially and word-medially, but not word-finally – where only voiced stops would surface.

For the purposes of modeling the diachronic development of unnatural processes, this paper distinguishes between static phonotactic restrictions and alternations. The difference is that the latter require a segment to alternate in a given phonetic environment within the same morphological unit. In other words, the MSCR holds for alternations in which a segment B surfaces as a segment A in an environment X within the same morpheme.

One of the advantages of constraint-based approaches to phonology is that static phonotactics and alternations are modeled simultaneously with the same theoretical devices (Prince & Smolensky Reference Prince and Smolensky1993/2004; Hayes Reference Hayes, Kager, Pater and Zonneveld2004; Pater & Tessier Reference Pater, Tessier, Slabakova, Montrul and Prévost2006). It is likely that productive phonotactic restrictions and productive alternations have the same underlying grammatical mechanisms and should therefore not be distinguished in synchronic grammars (Pater & Tessier Reference Pater, Tessier, Slabakova, Montrul and Prévost2006). There are three main reasons for why such a distinction is justified when modeling unnatural processes diachronically. Alternations provide considerably more and more reliable evidence for learners. In the case of alternations, learners are faced with surface forms within the same morpheme that contain both A and B. If a combination of sound changes does not result in an alternation, learners are faced with either categorical or gradient distributions. There is no doubt that phonotactic restrictions can be active phonological processes, but the likelihood of learners not acquiring a distribution of segments as an active process is much higher if evidence for the distribution does not surface in the same morpheme (e.g. Becker et al. Reference Becker, Ketrez and Nevins2011). Because of this increased likelihood that a combination of sound changes that provides only distributional evidence will not result in a productive process, combined with the fact that the typology of unnatural phonotactics is more challenging to establish (see below), I make a distinction between alternations and static phonotactics when modeling the diachronic development of unnatural processes. In fact, the two unnatural phonotactic restrictions which I argue elsewhere show signs of productivity, Berawan and Tarma Quechua (Beguš & Nazarov Reference Beguš and Nazarov2017, Reference Beguš and Nazarovto appear; Beguš Reference Beguš2018), did indeed arise from a combination of sound changes that would also give evidence for an alternation if all changes operated categorically. The fact that evidence for the unnatural phonotactic restriction was morphological (which resulted from the Blurring Process) likely contributed to the productivity of the unnatural phonotactic restrictions.

Second, phonotactic restrictions that do not arise from a Blurring Process might allow alternative analyses. In the scenario outlined above in which post-vocalic voicing of stops is followed by a sound change that reintroduces voiceless stops in post-vocalic (but not final) position, a potential analysis could treat the newly introduced stops as a distinct series, especially if some phonetic properties of the segments the new series develops from are preserved. While such an analysis is undesirable, it is available in the case of phonotactic restrictions and impossible in the case of alternations: one cannot assume a different phoneme within the same morphological unit.

Finally, I distinguish phonotactic restrictions from alternations for practical purposes. It appears that many unnatural phonotactic restrictions go unnoticed in the literature or are not analyzed as unnatural, probably precisely due to the reasons described above. Identifying unnatural distributions/phonotactic restrictions is a harder task not only for the learner, but also for linguists describing languages. For this reason, it is considerably more difficult to create typological studies of unnatural phonotactic restrictions. For example, I have found two languages which potentially allow only voiced but not voiceless stops in word-final position and which have not been discussed in the literature on unnatural processes. Ho and some dialects of Spanish might have a restriction against final voiceless stops while allowing word-final voiced stops for at least a subset of places of articulation. In Ho, for example, voiceless stops do not appear word-finally (except in some loanwords), but voiced labial and retroflex stops do surface word-finally (Pucilowski Reference Pucilowski2013). Word-final voiceless stops and labial and velar voiced stops in Spanish are restricted to loanwords. Word-final /d/, however, surfaces in native vocabulary. In a variety of Madrid Spanish, word-final /d/ is most frequently deleted or surfaces as a voiceless fricative, but in citation form, /d/ is recorded as surfacing as a voiced stop [d] in 30% of cases (Burkard & Dziallas Reference Burkard, Dziallas, Belz, Mooshammer, Fuchs, Jannedy, Rasskazova and Żygis2018).Footnote ^[39] It needs to be acknowledged that these cases are preliminary descriptions and detailed phonetic data is needed to confirm these restrictions. If the stops in word-final position are indeed phonetically real, these cases would suggest that a phonotactic restriction that allows word-final voiced stops while banning voiceless stops might be more common than previously thought. This would align well with the distinction between unnatural alternations that follow the MSCR and unnatural phonotactics that do not.

In sum, the MSCR states that unnatural segmental alternations always arise from at least three sound changes. Unnatural phonotactic restrictions or prosodic alternations can arise from fewer sound changes. This paper distinguishes between alternations and phonotactic restrictions because the first are more likely to result in productive processes, lack alternative analyses, and are easier to detect. While the MSCR does not apply to unnatural phonotactic restrictions, two cases I describe elsewhere (Berawan and Tarma Quechua; Beguš & Nazarov Reference Beguš and Nazarov2017, Reference Beguš and Nazarovto appear; Beguš Reference Beguš2018) that did result in productive unnatural phonotactic restrictions actually did arise from the three sound changes in the Blurring Process. A further study of unnatural phonotactics and their diachronic developments would reveal how readily restrictions arising from fewer than three sound changes result in productive processes.

We saw that, while a single sound change is constrained to follow phonetic naturalness, a combination of sound changes appears unconstrained: a number of single instances of sound change can operate on each other (limited of course by the timeframe of active operation), and even if such a combination results in unnatural alternation, the synchronic grammar can still incorporate it (Section 3.5); the final sound change will not be blocked (pace Kiparsky Reference Kiparsky2006, Reference Kiparsky and Good2008). However, if we assume that the combination of sound changes is unconstrained, we still need to explain why unnatural processes (B  ${>}$  A  $/$  X) are rare – as in the case of PND – or even unattested – as in the case of final voicing. Section 5 discusses implications of the Blurring Process and the MSCR for deriving typology within the Channel Bias approach.