Asymmetry of Kazakh velar and uvular consonants

Little descriptive work has been done on the place and voicing restrictions of the asymmetrical velar and uvular consonant inventory in Kazakh. In Kazakh, velar and uvular consonants are restricted depending on their neighbouring vowel. Velars appear in front vowel environments and uvulars appear in back vowel environments (place restriction). Voiced and voiceless velars and uvulars are restricted depending on their position in the word. At the morpheme boundary, velars and uvulars are voiceless in the word-final position and voiced in the stem-final position, when followed by a vowel-initial suffix (voicing restriction). The results from elicitation-based production experiments with six native Kazakh speakers reveal that the place restriction is not productive from real words to nonce words but the voicing restriction is. The data suggests a derived-environment effect where the resulting voicing process is conditioned morphologically. A theoretical analysis within Optimality Theory captures the voicing pattern using an indexed-markedness constraint and Local Conjunction.


Introduction
Most Turkic languages have velar and uvular consonants. These consonant inventories can vary and many of them exhibit restrictions on place of articulation and voicing. Little descriptive work has been done on these types of assimilation processes in Turkic languages (e.g., Comrie 1981;Nevins and Vaux 2004;Pattillo 2013). This paper aims to bridge the gap in the literature by looking at velar and uvular consonants in Kazakh. This language has velar and uvular consonants, but its inventory is asymmetrical.
To illustrate, let me briefly discuss what is already known about velar and uvular consonants in Kazakh. First, velar and uvular consonants are restricted with respect to their neighbouring vowels. Velars appear in words containing front vowels and uvulars appear in words containing back vowels, as shown in (1).
The existing descriptions contain rather limited amounts of data, which does not allow generalisations to be drawn about the full distribution of dorsal consonants. Many questions remain unanswered. Are the velar and uvular consonants contrastive intervocalically? What happens to [x]: does it alternate with [q], [ɡ] or [ʁ]? Is the restriction on velars next to front vowels and uvulars next to back vowels productive? Does "next to" the vowel mean before, after, or either? Finally, how do we account for the distribution of velars and uvular consonants theoretically?
To answer these questions, I conducted three elicitation-based production experiments with six native Kazakh speakers. The results revealed the following. In real words, participants produced the target velar and uvular consonants. Velar consonants were produced in words constraining front vowels and uvular consonants were produced in words containing back vowels. There was a voicing alternation in the stem-final position whereby voiceless velar and uvular consonants were produced word-finally and voiced velar and uvular consonants were produced intervocalically. The first generalisation found in real words was not extended to nonce words: the velar and uvular consonants appear unrestricted in stems beside front and back vowels. However, the second generalisation was found to be productive: word-final voiced targets were devoiced and stem-final voiceless targets were voiced intervocalically.
In the theoretical analysis, I focus on the word-final and the stem-final position as opposed to the stem-internal positions. The data suggests that the voicing alternation can be explained as a morphologically-derived environment effect (Burzio 2011;Hall 2006;Kula 2008). I offer an analysis within Optimality Theory Smolensky 1993/2004;Kager 2004;McCarthy 2008) using an indexed-markedness constraint (à la Pater 2007) and Local Conjunction (Itô and Mester 2003;Kager 2004;Smolensky 2006;Crowhurst 2011) to capture the pattern.
The paper is structured as follows: Section 2 discusses the background; Section 3 summarises the results from the elicitations; Section 4 offers some discussions points regarding the results; Section 5 provides an analysis within Optimality Theory focusing on the word-final and the stem-final dorsal consonants; and Section 6 briefly concludes.

2 Background
This section provides a more detailed overview of what is already know about velar and uvular consonants in Kazakh. This consonant inventory is asymmetrical. The place restriction on consonants /k/ and /g/ exhibit velar and uvular variants that interact with the neighbouring vowels. Native Kazakh words exhibit vowel harmony so velars appear in words containing front vowels and uvulars appear in words containing back vowels. The voicing restriction depends on whether the velar and uvular consonants appear in the word-final position or the stem-final position with a following vowel-initial suffix. Voiceless velar and uvular consonants appear word-finally and voiced velar and uvular consonants appear stem-finally intervocalically. Nothing is known about what happens to the fifth velar consonant [x] found in loanwords so it is unclear whether this segment exhibits any place or voicing restrictions.
This section is organised as follows. Section 2.1 discusses the velar and uvular consonant inventory, Section 2.2 discusses the place restriction and Section 2.3 discusses the voicing restriction.

Velar and uvular consonant inventory
Kazakh has an inventory of 5 velar and uvular consonants. The inventory is shown in (3) (Bekturova and Bekturov 1996)), whereby the alternation differs in both place and manner of articulation.

Place restriction
Velar and uvular consonants are restricted in the place of articulation depending on neighbouring vowels. Like in many Turkic languages, vowel harmony is a key feature in native Kazakh words, whereby all vowels are either front or back. This affects whether a dorsal consonant within a word is velar or uvular. Each word usually contains only front vowels or back vowels, as shown in (4).  (6) (Kara 2002;Muhamedowa 2016).
Notice in the above examples that these words do not contain only front vowels or back vowels.

Voicing restriction
Velar and uvular consonants are restricted in voicing depending on what position the stem-final consonant appears in a word. As in many Turkic languages, the voicing alternation is also a key feature at the morpheme boundary of Kazakh words. This affects whether the dorsal consonant appears as voiced or voiceless. This paper examines two phonological phenomena: word-final devoicing and stem-final intervocalic voicing when followed by a vowel initial suffix, as shown in (7) Notice in the above examples that the voiceless velar and uvular consonants appear in the word-final position and the voiced velar and uvular consonants appear intervocalically with a following vowel-initial suffix (Kara 2002). Voiceless and voiced pairs or velar and uvular consonants alternate with one another but it is ambiguous whether it is word-final devoicing or intervocalic voicing. Nothing is known about what happens to [x]. Due to insufficient explanations in the literature regarding the properties of Kazakh velar and uvular consonants, I conducted three elicitation-based experiments with 6 native Kazakh speakers. The following section discusses the results from those elicitations.

Results
This section discusses the results from the elicitations. I present the results going through the place restriction and the voicing restriction. There are several main findings. First, the pattern whereby velar consonants are produced in front vowel environments and uvular consonants are produced in back vowel environments was not extended to nonce words. All five dorsal consonants were unrestricted in the wordinitial, word-medial, and word-final positions in nonce words containing front or back vowels. Second, the pattern whereby voiceless velar and uvular consonants appear word-finally and voiced velar and uvular consonants appear intervocalically with a following vowel-initial suffix was extended to nonce words. Target voiced velar and uvular consonants were devoiced in the word-final position and target voiceless velar and uvular consonants were voiced intervocalically. Third, the voiceless velar fricative [x] alternated with the voiced uvular fricative [ʁ] in the word-final and stem-final positions.
This section is organised as follows. Section 3.1 presents the place restriction results and Section 3.2 presents the voicing restriction results.

Place restriction
The place restriction observed in real words was not extended to nonce words. In real words, participants produced target velar consonants in front vowel environments and target uvular consonants in back vowel environments, as shown in (8) In nonce words, we expected participants to produce velar consonants in front vowel nonce words even if the letter corresponded to a uvular consonant, and the reverse situation for uvular consonants in written stimuli. However, this did not turn out to be the case. Participants produced the target velar and uvular consonants regardless of the vowel. This included the fifth dorsal consonant [x]. All five consonants were produced invariably, regardless of whether a front or back vowel appeared in the nonce words. Thus, the results suggest that nonce words patterned with loanwords, where the distribution of dorsal consonants is not limited in its place of articulation. This shows that the place restriction does not seem to be productive.

Voicing restriction
The voicing restriction observed in real words was extended to nonce words. In real words, participants produced voiceless velar consonants in front vowel environments and voiceless uvular consonants in back vowel environments in the word final position, as shown in (10). (10) Dorsal consonants in word final position a. VELAR: [købelek] 'butterfly' b. UVULAR: [qɑsɯq] 'spoon' Notice in the above examples that the word-final segment is voiceless whether it be a velar or a uvular consonant. Participants also produced voiced velar consonants in front vowel environments and voiced uvular consonants in back vowel environments intervocalically with a following vowel-initial suffix, as shown in (11).
Notice in the above examples that the stem-final segment is voiced intervocalically with a following vowel-initial suffix. Also, intervocalic voicing is extended to the stem-final [x], as shown in (12) In nonce words, we expected participants to devoice voiced targets word-finally. We anticipated participants to neutralise voiced velar /g/ to voiceless [k] and the voiced uvular [ʁ] to voiceless [q]. Four of the six participants devoiced all voiced targets in the word-final position while the other two preserved the word-final voiced consonant when reading from written stimuli. However, all participants produced word-final voiced targets as devoicing from auditory stimuli. This included the fifth dorsal consonant [x], which was produced faithfully (since it is voiceless already). We also expected participants to voice stemfinal voiceless velar and uvular consonants intervocalically with a following vowel-initial suffix. All participants produced the stem-final target voiceless velar and uvular consonants as voiced intervocalically.
The surprising results concerned the velar and uvular fricatives [x] and [ʁ]. First, the voiced velar /ɡ/ neutralised to [k] but the voiced uvular [ʁ] did not neutralise to [q], as expected, but partially neutralised to [χ] when using written stimuli but to [x] when using auditory stimuli. The manner of articulation was maintained, even when it resulted in an illicit segment. Second, the stem-final /x/ voiced to [ʁ] intervocalically when followed by a vowel-initial suffix. This matches the real word production. Thus, the results suggest that nonce words patterned with real words where the distribution of stem-final velar and uvular consonants is limited in its voicing pattern. This shows that the voicing restriction seems to be productive.

Discussion
This section is a discussion of the results that were explained in the previous section. The results found that the place restriction was not productive but the voicing restriction concerning their extension from real words to nonce words. The voicing restriction contributed the most to the description of Kazakh and provides the greater basis for some interesting discussion points.
First, we determined that the place restriction found in real words was not extended to nonce words. Real words show velar consonants appearing in front vowel environments, and uvular consonants appearing in back vowels environment, but this strict pattern was not found in nonce words. There are several possible explanations for this. It could be that the restriction on the place of articulation is no longer productive. The pattern observed in native words would not be extended to new words. It could be that participants treated nonce words as loanwords. Since the place restriction is not found in Kazakh loanwords, participants would not neutralise place of articulation in loanwords and would instead pronounce them faithfully. We know that loanwords are often exceptional cross-linguistically (see Itô and Mester 1999, 2001and Kang 2011 for an overview) so this is not surprising. It could be that the place neutralisation was not obtained due to the task effect. When reading, participants fail to neutralise, thus neutralisation could be described as incomplete (e.g., Slowiaczek and Dinnsen 1985; Charles-Luce and Dinnsen 1987; Warner et al. 2004;Dmitrieva et al. 2010;Kharlamov 2012Kharlamov , 2014Roettger et al. 2014 and many others).
Second, the voicing facts are more complex. We have also determined that the voicing restriction found in real words was extended to nonce words. For word-final devoicing, recall that two of the participants invariably produced voiced segments in the word-final position faithfully. Also, all participants invariably neutralised [ʁ] with [x], instead of the expected [q], in the word-final position. Even though participants were asked to treat nonce words as if they were real words, it is possible that participants treated nonce words as loanwords. In many languages, loanwords may allow less restrictive phonotactic requirements than for native words, This phenomenon can be modelled using a co-phonology approach (Orgun 1996;Inkelas et al. 1997;Antila 2002;Inkelas and Zoll 2007) or indexed constraints Mester 1999, 2001;Pater 2007;Flack 2007). The co-phonology approach captures the diversity between native and loanword phonology by associating a lexical or morphological class with a different phonological grammar. The indexed constraint approach captures the diversity under a single constraint ranking for the entire language but a constraint within the ranking is indexed to an individual morpheme.
It is also possible that neutralisation was incomplete. The reading task using written stimuli confirmed that incomplete neutralisation could stem from the influence orthography has on written speech. If spelling suggests a difference, participants may make a distinction during careful speech. This was demonstrated by no devoicing in the word-final position and the participants' ability to produce a velar beside a back vowel and a uvular beside a front vowel.
For stem-final intervocalic voicing with a following vowel-initial suffix, the voicing alternation occurs at the local domain of the suffix and does not occur in the identical environment found wordinternally, as shown in (13).
The indexed constraint theory of lexical exceptions offers the opportunity to propose a morphemespecific markedness constraint. These types of constraints are limited to a specific set of suffixes and apply to entire outputs in which the indexed constraints occur. Pater (2007) discusses how a phonological process applies when the conditioned environment contains a portion of an affix in a morphologicallyderived environment. The scope is limited to the string of segments that contains some portion of the morpheme.
The voicing restriction contributed the most to the description of Kazakh since this phonological process is productive across real and nonce words. The place restriction was not productive and will not be discussed further since this phenomenon has been previously explored (see Bekturova and Bekturov 1996;Kara 2002;Batayeva 2013;Muhamedowa 2016). The following section will provide an Optimality Theory analysis that accounts for final devoicing and intervocalic voicing.

Analysis
This section provides an analysis within Optimality Theory (OT) Smolensky 1993/2004;Kager 2004;McCarthy 2008) that accounts for both final devoicing and intervocalic voicing of the velar and uvular consonants. To capture the voicing pattern, standard featural assumptions are used whereby all segments are fully specified for their voicing, place and manner. The feature Pater 2007) and Local Conjunction (see Itô and Mester 2003;Smolensky 2006;Crowhurst 2011) are used to account for the voicing patterns.

[±voice] differentiates between voiced ([ɡ] and [ʁ]) and voiceless ([k], [q] and [x]) segments, the feature [±high] differentiates between velars ([k], [ɡ] and [x]) and uvular ([q] and [ʁ]) segments and the feature [±continuant] differentiates between fricative ([x] and [ʁ]) and stop ([k], [q] and [ɡ]) segments. 4 A combination of an indexed-markedness constraint (see
The analysis is organised as follows. Section 5.1 presents word-final devoicing and Section 5.2 presents intervocalic voicing. Since voiced velar and uvular consonants invariable devoice in the word-final position, I employ the word-final devoicing constraint, as shown in (14). This constraint prevents a change in the VOICE specification of a segment from input to output. Let us first consider /ɡ/ neutralising to [k] in the word-final position. Voicing changes while maintaining place and manner of articulation. Consider the tableau in (16). Candidate (b) fatally violates the top-ranked constraint FINDEV, which the winning candidate (a) does not.  This constraint prevents a change in the continuant [high] of a segment from input to output. IDENT constraints fall under McCarthy and Prince's (1995) correspondence constraints whereby a constraint checks that input and output segments agree in feature specification. Since manner is maintained of articulation is maintained of place of articulation when [ʁ] neutralised to [x], instead of to [q], in the word-final position, IDENT(cont) must be ranked higher than IDENT(hi). I also employ a higher ranked markedness constraint *χ that ensures the voiceless uvular fricative does not surface. This is an instance of Structure Preservation (e.g., Kiparsky 1985) whereby a phonological rule will not introduce a novel sound that is not part of the phoneme inventory. The segment [χ] is not found within the Kazakh consonant inventory and should not be a possible output under any circumstance. We also know that FINDEV must rank above IDENT(voi) and IDENT(cont) must be ranked above IDENT(hi).

Final devoicing
Consider the tableau in (19) for /ʁ/ neutralising to [x] word-finally. Voicing and place of articulation change while maintaining manner of articulation. Candidate (a) and (b) fatally violate the top-ranked constraints *χ and FINDEV, respectively. Both candidate (c) and (d) violate IDENT(voi) but candidate (d) fatally violates IDENT(cont), which candidate (c) does not. Since voiceless velars and uvulars voice at the stem-final boundary when followed by a vowel-initial suffix, I employ the indexed-markedness constraint in (20) which does not allow morpheme specific voiceless segments to appear intervocalically.
This constraint prevents a morpheme-specific string where a voiceless segment is not allowed to surface intervocalically in the output and will apply when any part of the string falls within the suffix. In other words, a suffix vowel will affect the stem consonant but is blocked from applying in the identical steminternal environment. The voicing pattern is captured using one constraint ranking while being able to account for the morphologically-conditioned phonology. Since voiceless velar and uvular consonants voice intervocalically at the morpheme boundary, *VTVSUFFIX must be ranked higher than IDENT(voi). A voicing change is favoured over being faithful to [voice]. Let's first consider /k/ voicing to [ɡ] intervocalically when followed by a vowel-initial suffix. Voicing changes while maintaining place of articulation and manner of articulation. Consider the tableau in (21). Candidates (a), (c) and (d) fatally violate the higher ranked constraint *VTVSUFFIX. Recall that this constraint applies to the stem-final consonant because of the locality effect (see Pater 2007). Both candidates (b) and (e) violate IDENT(voi) but candidate (e) fatally violates IDENT(cont), which candidate (b) does not. lɪbsɪɡ-ɪ * c.
lɪbsɪq-ɪ *! * * Finally, let's consider /q/ voicing to [ʁ] intervocalically when followed by a vowel-initial suffix. Voicing and manner of articulation change while maintaining place of articulation. An additional constraint is required to account for both /q/ and /x/ voicing to [ʁ]. In the /x/ alternation, the manner of articulation is respected but not place of articulation; In the /q/ alternation, the place of articulation is respected but not manner of articulation. At this point, the current constraint ranking favours /q/ alternating with [ɡ] over [ʁ] To rule out unattested alternations, I employ a locally-conjoined constraint. This type of constraint creates a single composite constraint by conjoining two constraints. Itô and Mester (2003) and Smolensky (2006) assume that Universal Grammar makes a combinatorial operation "&" available to individual languages. These languages may choose to activate "&" to derive a complex constraint on a languagespecific basis. Smolensky (2006: 43) defines Local Conjunction as follows in (23). A constraint conjunction expands the expressive power of Optimality Theory. Conjoined constraints also increase the available set of constraints and improve the precision of the evaluation process of possible candidates (Crowhurst 2011). Locally-conjoined constraints will incur a violation if and only if both members of the complex constraint are violated within a specified domain (Kager 2004). Therefore, I employ the composite constraint in (24), which requires place of articulation to be maintained and, at the same time, disfavours a morpheme-specific intervocalic velar consonant.
lɪbsɪx-ɪ *! * * * To summarise, the change in voicing may result in illicit consonants in Kazakh, whereby the next best option is chosen by changing place of articulation, rather than manner of articulations. This is seen by /x/ voicing to [ʁ], instead of voicing to [ɡ], while /ʁ/ devoices to [x], instead of devoicing to [q]. Under the constraint ranking IDENT(voi), IDENT(cont) >> IDENT(hi), one would expect /q/ to voice to [ɡ], but instead it voices to [ʁ] unexpectedly. The local conjunction that created the single composite constraint IDENT(hi)&*VC[+hi]VSUFFIX prevents the change into an intervocalic [ɡ] when followed by a vowel-initial suffix.
The constraint ranking in the above analysis is not always obvious and is somewhat complex. The relationship between constraints can be visualised in the following diagram in (26). The analysis above accounted for word-final devoicing and intervocalic voicing when followed by a vowel-initial suffix. The voicing pattern of velar and uvular consonants in Kazakh suggests a derivedenvironment effect (Burzio 2011;Hall 2006;Kula 2008) and is captured within Optimality Theory Smolensky 1993/2004;Kager 2004;McCarthy 2008) using an indexed-marked constraint to account for the morphologically-conditioned phonological process (Pater 2007) and a Local Conjunction constraint (Itô and Mester 2003;Kager 2004;Smolensky 2006;Crowhurst 2011)

Conclusion
This paper looked at the asymmetrical inventory of velar and uvular consonants in Kazakh. Existing descriptions contain limited amounts of data, which made it difficult to draw any generalisations regarding the full distribution of velar and uvular consonants in Kazakh. The aim of this research was to fill the gap in the literature by conducting elicitation-based experiments using real and nonce words.
The results provided several generalisations. In real words, participants produced target velar consonants in front vowel environments and target uvular consonants in back vowel environments. However, in nonce words, the velar and uvular consonants were unrestricted in front and back vowel environments. While the place restriction was not productive and did not extend from real words to nonce words, the voicing restriction was productive. Participants produced word-final voiced targets as devoiced word-finally and stem-final voiceless targets as voiced when followed by a vowel-initial suffix.
A better understanding of Kazakh velar and uvular consonants was achieved from the elicitationbased production experiments, which increased the amount of data on Kazakh. The voicing patterns, word-final devoicing and intervocalic voicing, contributed the most to the description since the voicing alternations were productive across real and nonce words. Focusing on these two phenomena, the data suggests a derived-environment effect whereby the phonological process is conditioned morphologically. A theoretical analysis of the voicing alternations was captured within Optimality Theory using an indexed-markedness constraint and Local Conjunction. In short, this presents one of the most extensive works done on Kazakh yet.