Phrase-Level ATR Vowel Harmony in Anum—A Case of Recursive Prosodic Phrasing

Abstract

(1) Like many other Kwa languages, Anum employs a pattern of [ATR] vowel harmony that is regressive and [+ATR] dominant (RVH). This paper analyses RVH as a phrasal process which takes into account recursive phonological phrases. The proposal argues for an application of the process within and across non-maximal phonological phrases (φ) and a blocking of application across maximal phonological phrases (φ_max). (2) Investigating RVH in Anum in more detail, the size of constituents and the complexity of sentence structures are varied. Target sentences were recorded and transcribed for [ATR] vowel harmony. (3) The empirical data show that RVH applies frequently between words that belong to either the same or to different syntactic constituents, but is blocked between two verb phrases of a serial verb construction and between any word and a following sentence-final time adverbial. Interestingly, RVH occurs between a sentence-initial subject constituent and a following verb or verb phrase, independent of the size of the subject constituent and the remaining number of words in the sentence. (4) The proposed OT analysis accounts for RVH within syntax-phonology Match Theory and addresses both word-level and phrase-level harmony. The special behaviour of subject constituents that prosodically phrase together with verbs and with constituents of the verb phrase (VP) is discussed. Either a phonological well-formedness constraint or a syntactically distinct input may account for phrasing effects with subject constituents in Anum.

1. Introduction

Vowel harmony is widespread across the languages of the world and is usually analysed as a phonological harmony process at the word level. In this paper, I show that vowel harmony applies in larger domains than the word such that the harmonizing feature [+ATR] spreads regressively across word boundaries (RVH). I argue that vowel harmony across word boundaries can be analysed as a phrasal phenomenon. In line with a growing body of research acknowledging this phrasal approach to vowel harmony (Kimber 2011; Kügler 2015; Nilsson and Downing 2019; Obiri-Yeboah and Rose 2022; Downing and Krämer n.d.), this study discusses the actual phrasal domain of application in Anum, investigating RVH in different syntactic contexts. A closer examination of phrasal vowel harmony in Anum shows that across word boundaries, harmony varies; while it applies between nearly all words, certain exceptions are found where RVH is blocked. I argue that Anum phrasal vowel harmony can be captured by assuming a model of the syntax-prosody interface (Selkirk 2011) where prosodic well-formedness constraints account for the variation found in the data. I propose that RVH in Anum applies within and between phonological phrases, but is blocked at a maximal phonological phrase boundary (φ_max), thus acknowledging recursive phonological phrasing (Itô and Mester 2012).

Anum is a two-tone Kwa language spoken in Ghana. It belongs to the Guang branch of the Kwa languages, which branches into the North, North-East, Central, and South Guang languages; Anum belongs to the South Guang languages (also called Gwa) (Kropp Dakubu 1988, p. 79). As in many other Kwa languages, ATR vowel harmony is productive in Anum (Painter 1971; Obeng 1995; Ofori 2013). While vowel harmony is usually recognized as a word-level phonological phenomenon, vowel harmony can occur across word boundaries as well (for an overview, see Downing and Krämer n.d.). For instance, both Painter (1971) and Obeng (1995) report on vowel harmony between words in Anum, which is regressive and restricted to the [+ATR] feature. In (1), subject-verb sentences are shown. The [+ATR] high front vowel of the verb in (1a) regressively affects the adjacent [–ATR] vowel of the preceding word, changing its [–ATR] value into [+ATR]. The vowels that undergo cross-word harmony appear in boldface throughout the paper. In (1b), the final [–ATR] vowel of the subject noun is not affected because the verb contains a [–ATR] vowel.

1.	Cross-word vowel harmony in Anum (Obeng 1995, p. 149)
	a.	/kwamɪ	dʑi/	[kwami dʑi]
		Kwame	eat
		‘Kwame eats.’
	b.	/kwamɪ	na/	[kwamɪ na]
		Kwame	walk
		‘Kwame walks.’

This type of regressive [+ATR] vowel harmony between words (RVH) is a common phonological process in further Kwa languages like in the neighbouring dialect of Anum, Gua (Obiri-Yeboah and Rose 2022), in Akan (Kügler 2015), Nkami (Akanlig-Pare and Asante 2016), Tafi (Bobuafor 2013), Efutu (Obeng 2008), Igbira (Scholz 1976), Chumburung (Snider 1989), or Nawuri (Casali 2002), as well as in other (West-)African languages like Dεg (Gur; Crouch and Herbert 2003), Vata (Kru; Kaye 1982; Kimber 2011), Somali (Cushitic; Nilsson and Downing 2019), or Kinande (Bantu; Archangeli and Pulleyblank 2002). RVH has been analysed as a phrase-level process that operates within phonological phrases (Kügler 2015; Nilsson and Downing 2019; Obiri-Yeboah and Rose 2022): In Akan, the process of RVH usually applies between words, except for words that are divided by a maximal phonological phrase boundary (Kügler 2015). In particular, RVH in Akan is blocked between a verb and a preceding subject, assuming that a VP constitutes a case of a maximal phonological phrase in prosodic constituency. In Gua, on the other hand, RVH applies in phonological phrases that preferably consist of two or three words (Obiri-Yeboah and Rose 2022). Rhythmic size constraints on the size of prosodic constituents thus determine the domain of application of RVH in Gua. For Somali, Nilsson and Downing (2019) argue that the domain of cross-word harmony is a complex word group, which constitutes a prosodic unit smaller than the phrase. In all cases, the essential conclusion is that the domain of RVH is larger than a prosodic word.

The assumptions underlying the phrasal analysis of RVH proposed here are the following: First, syntactic constituents like lexical words, syntactic phrases, and clauses are matched to prosodic constituents like the prosodic word (ω), the phonological phrase (φ), and the intonation phrase (ι) (Selkirk 2011). The syntax-phonology match thus determines the relevant prosodic domains in which phonological processes like RVH apply. In the Match model proposed by Selkirk (2011), phonological well-formedness constraints may, however, adjust or restructure prosodic phrasing with the result that prosodic constituency is no longer isomorphic to syntactic constituency. Second, prosodic constituents are assumed to be recursive, a fact that follows from Match Theory alone. Itô and Mester (2012) elaborate on recursive prosodic constituents and suggest that prosodic constituents represent minimal or maximal constituents, i.e., they are dominated by or dominate another constituent of that domain. Phonological processes can then be analysed as being domain-sensitive, relating to a particular level of prosodic phrasing. For instance in Akan, RVH applies regularly within and between non-maximal φ-phrases, but RVH is blocked at maximal phonological phrase boundaries (φ_max) (Kügler 2015). Like in Akan, RVH in Gua also applies within φ-phrases. Unlike in Akan though, in Gua, prosodic phrasing is achieved by a Binarity constraint requiring that φ-phrases consist of at least two prosodic words (Obiri-Yeboah and Rose 2022). Hence, a prosodic well-formedness constraint determines the domain of phrase vowel harmony in Gua.

The proposal argued for in this paper is that Anum RVH is a phrasal phonological process like the one in Akan and Gua. I assume an isomorphic syntax-phonology match for Anum, recursive prosodic constituency, and language-specific assumptions for subject constituents to account for the RVH patterns found in Anum. The domain of application of RVH is identical to that of Akan, i.e., RVH occurs within maximal φ-phrases. The crucial, language-specific fact is that in Anum, subject NPs seem to be phrased together with a following verb. This particular phrasing can be accounted for by either applying a prosodic well-formedness constraint that restructures prosodic phrasing, or by assuming a distinct syntactic input for the syntax-phonology match. Both options are discussed in Section 4.5.

The paper is structured as follows. Section 2 provides a brief background on the phonology and syntax of Anum. Section 3 presents the empirical data on RVH in Anum in different syntactic structures. Section 4 presents an OT analysis of RVH that includes both word-level and phrase-level regressive [+ATR] vowel harmony, and Section 5 concludes.

2. Aspects of Anum Phonology and Syntax

2.1. Anum Vowels and Vowel Harmony

Anum has nine oral vowels and seven nasal vowels as shown in (2). The distribution of the nasal vowels is limited mainly to word-final position according to Ofori (2013). The nine oral vowels are divided into two harmonic classes as shown in (3). There is no contrastive [+ATR] counterpart to /a/. However, there is an allophonic variant of /a/, which is the front vowel [æ] (Obeng 1995). This allophone only appears in word-initial position, and both mid and high vowels trigger the appearance of the allophone, e.g., as in/akpu/ [ᴂkpu] ‘chin’. Note that the [+ATR] a-allophone behaves like all [+ATR] phonemes in terms of vowel harmony processes; it neither blocks vowel harmony within the word, nor does it block cross-word harmony (see e.g., (8b) below).

2.	Anum oral vowels		Anum nasal vowels
	i, ɪ	u, ʊ	ĩ, ɪ̃	ũ, ʊ̃
	e, ε	o, ɔ	ε̃	ɔ̃
		a		ã

3.	Anum harmonic vowel classes
	a.	Class 1 [+ATR]:	/i e o u/
	b.	Class 2 [–ATR]:	/ɪ ε ɔ ʊ a/

From an articulatory point of view, tongue root advancement manifests itself as variation in the ‘relative size of the pharynx’ (Lindau 1975, p. 80; Dolphyne 1988; Tiede 1996). Articulatory studies have shown that there is almost no difference in the degree of primary constriction, which means that the tongue dorsum maintains its height in the oral cavity, with the crucial difference arising through an increase in the size of the pharyngeal cavity in the case of [+ATR] vowels. This is different from the [±tense] distinction, which is mainly rooted in variation in the height of the tongue dorsum (e.g., Lindau et al. 1972; see Hoole and Mooshammer 2002 for this distinction in German vowels). According to Painter (1971, p. 243), Anum rather exhibits a decrease in the relative size of the pharyngeal cavity for [–ATR] vowels, which is the opposite effect to in Akan, for instance, where [+ATR] vowels show an increase in pharyngeal cavity size; note, however, that this is an impressionistic observation as there are no articulatory data on Anum vowel harmony available. Acoustically, an increase in the size of the pharyngeal cavity results in a lowering of the first formant (F1) (Lindau 1979; Hess 1992; Tiede 1996). At the same time, the F1 dimension mirrors the degree of tongue height, which may be why the features [±ATR] and [±tense] are often used interchangeably in phonological theory.

Systematic [±ATR] vowel harmony in Anum is a word-level phenomenon (Painter 1971). The vowels of a word consist of either of the two harmonic classes, as is shown in (4). In (4a), all the vowels in the words are [−ATR], while in (4b), all the vowels are [+ATR].

4.	a. [−ATR]		b. [+ATR]
	ɔdɪdε	‘metal’	eteemi	‘money’
	fʊkyɪ	‘sweep’	onni	‘meat’
	ɔlɔfε	‘hundred’	hurobui	‘grinding stone’

In the case of morphological derivation, the ATR feature of the word stem determines the ATR feature of the affixes. In (5), the pronominal prefix paradigm is shown to depend on the ATR specification of the verbal stem. The pronominal prefixes appear as their [−ATR] variants in (5a) with the [−ATR] verb mɔ ‘to kill’. In (5b), in contrast, the prefixes appear with a [+ATR] vowel in combination with the [+ATR] verb ku ‘to dig’. Note that there is a difference between the singular and plural pronouns. While the singular pronouns prefix to the verb, the plural pronouns present separate words that are not prefixed to the verb. Hence, the singular pronouns undergo vowel harmony, agreeing in their ATR specification with that of the verb stem. The plural pronouns in fact show a case of vowel harmony across word boundaries, to be discussed in Section 3, in which only the vowel that is immediately adjacent to the verb changes its [ATR] specification in the case of a [+ATR] verb stem.

5.	a. [−ATR]		b. [+ATR]
	mɔ	‘to kill’	ku	‘to dig’
	mɪmɔ	‘I kill’	miku	‘I dig’
	wʊmɔ	‘you kill’	wuku	‘you dig’
	mʊmɔ	‘she/he kills’	muku	‘she/he digs’
	εnɪ mɔ	‘we kill’	εni ku	‘we dig’
	εnɪ mɔ	‘you.PL kill’	εni ku	‘you.PL dig’
	εmʊ mɔ	‘they kill’	εmu ku	‘they dig’

In verbal morphology, pronominal singular prefixes (as in (5)), future and perfect tense markers, and completive aspect markers are prefixed to the verb stem and harmonize with the ATR specification of the verb stem (Obeng 1995). These facts suggest a unidirectional, right-to-left spreading vowel harmony in Anum (Painter 1971; Ofori 2013). Although there are suffixes in Anum, like the agentive marker hʊ, the ATR value of the suffix does not undergo vowel harmony in the case of a [+ATR] noun, as in æbili-hʊ ‘a speaker’ (Ofori 2013).1 At first sight, one might be tempted to analyse Anum vowel harmony as a case of root-controlled vowel harmony, similar to other Kwa languages like Akan (Clements 1985; Dolphyne 1988; but see Casali 2012 for a [+ATR] dominance view in Akan). However, following the arguments in Casali (2012) for Akan, it seems more convincing to analyse vowel harmony in Anum as a regressive [+ATR] dominance effect. Like in Akan, Anum affixes come in their [−ATR] variant in isolation, which suggests an underlying representation of the [−ATR] feature; the presence of a non-changing [−ATR] suffix also points in this direction. Although leftward spreading of [+ATR] between roots in compound words does not seem to occur in Anum (e.g., akyɪ-nuu ‘elder woman’), a proper analysis of compounding in Anum is lacking, which might shed light on some crucial differences in word formation compared to Akan. For instance, plural is marked on each part of the compound, which may indicate that each word stem of the compound is phrased separately, joining as a compound at a higher level. An in-depth analysis of compounding in Anum goes beyond the scope of the present study and is left for future research. Nevertheless, the distribution of [ATR] in affixes as well as in cross-word [+ATR] harmony suggests an analysis of vowel harmony in Anum as [+ATR] dominant (cf. Casali 2012).

Like in Akan, in Anum disharmonic roots exist (6). Note that in (6), the word-initial syllables carry the [+ATR] feature followed by [−ATR] syllables. There are no words with the reverse pattern, corroborating the [+ATR] dominant leftward direction of vowel harmony in Anum. If the [−ATR] feature were to trigger vowel harmony, and given the leftward direction of vowel harmony discussed above, these disharmonic words would be predicted to consist of [−ATR] vowels only. If, in contrast, dominant [+ATR] vowel harmony were bidirectional, no disharmonic words would be expected to exist in Anum. Since neither is the case, we can conclude that vowel harmony in Anum is a [+ATR] dominant, right-to-left directional harmony process.

6.	humε	‘swell’
	ᴂhibɔhʊ	‘hunter’
	ibiεsʊ	‘market’

2.2. Anum Word Order Properties

To account for vowel harmony patterns between words in terms of a phrasal approach, the word order properties of Anum are relevant. Like most other Kwa languages, Anum is an SVO language (Ofori 2013). Usually, topic and focus constituents are fronted to the sentence-initial position, and focused object constituents are regularly resolved in-situ by a resumptive pronoun. Adjuncts or adverbials are placed after the object in sentence-final position. At the phrase level, Anum also allows for serial verb constructions (SVC), where at least two VPs including their object arguments follow each other (Kropp Dakubu 1988; Shluinsky 2017).

The structure of the NP requires modifiers to follow their head noun except for possessive pronouns, which precede the noun (Ofori 2013). If there is more than one modifier, they appear in a strict order; the adjective is closest to the head noun, followed by a quantifier, and a final definite article.

3. Regressive Vowel Harmony across Word Boundaries

As introduced above, vowel harmony in Anum is not only a word-level phenomenon but also takes place across word boundaries (7). Like for word-level harmony, the trigger of cross-word harmony is the [+ATR] feature. As a necessary prerequisite for this phonological process, two words with distinct ATR specifications where a [−ATR] word precedes a [+ATR] word need to be adjacent (Obeng 1995). This is identical in all the languages cited in the introduction. The [+ATR] feature triggers assimilation in the adjacent vowel of the preceding [−ATR] word. Like in Akan (Kügler 2015) and Gua (Obiri-Yeboah and Rose 2022), but unlike in Vata (Kimber 2011) and Nawuri (Casali 2002), it is only the final vowel of the [−ATR] word that is affected. The process is thus limited in its scope.

7.	a.	kwamɪ	dʑi	[kwami dʑi]
		‘Kwame’	‘eat’	‘Kwame eats.’
	b.	dansʊ	dʑi	[dansu dʑi]
		‘Danso’	‘eat’	‘Danso eats.’
	c.	ɔsεkɪrε	dʑi	[ɔsεkɪre dʑi]
		‘Ɔsεkrε’	‘eat’	‘Ɔsεkrε eats.’
	d.	anɔ	dʑi	[ano dʑi]
		‘Ano’	‘eat’	‘Ano eats.’
	e.	ansa	dʑi	[ansæ dʑi]
		‘Ansa’	‘eat’	‘Ansa eats.’
	f.	ansa	jeli	[ansæ jeli]
		‘Ansa’	‘stand’	‘Ansa stands.’	(from Obeng 1995, p. 149)

The data on Anum RVH presented in Obeng (1995, p. 149) show that each of the five [−ATR] vowels of Anum are affected similarly, assimilating to a [+ATR] vowel (7). When comparing (7e) and (7f), it becomes obvious that both high and mid [+ATR] vowels trigger RVH. The syntactic structure of the examples in (7) is identical, and the clause consists of a subject noun and a verb.

The phonological process of RVH applies in a number of contexts in Anum, which suggests that it is a regular assimilation process in this language. However, RVH is blocked in certain other contexts. Before presenting the data on regularly occurring RVH in Anum in Section 3.2, Section 3.1 introduces the speech material used for analysis and its elicitation. Section 3.3 discusses contexts in which RVH is blocked.

3.1. Speech Materials

The main part of the speech data for the analysis of RVH in Anum presented here was elicited on a fieldwork trip to Ghana in spring 2019, while some preliminary data was elicited on a fieldwork trip to Ghana in spring 2018. The complete data set was developed based on information from the grammar of Anum by Ofori (2013), and by working with Richard Lah, a native speaker of Anum, as a language informant. Word lists were created that contained words with distinct vowel harmony patterns and verb morphology paradigms. Care was taken to include data on all vowel distinctions. In addition, lists were created with sentences that contained the relevant environments to test for phrasal vowel harmony. For this purpose, complex noun phrases were embedded into sentences, and intransitive and transitive sentences with differing numbers of objects as well as sentences including time adverbials and serial verb constructions were created.

In order to test phrasal vowel harmony, five native speakers of Anum (4 male, 1 female) were recruited and recorded at the University of Education, Winneba. The speakers were seated in a quiet room. They were equipped with a Rode Lavalier condenser microphone, which was plugged into a Zoom H4n Pro recorder. The speakers completed different tasks including reading the word and sentence lists from which the current data is taken. Participants were asked to read each sentence silently and then produce the sentence as in a natural conversation.

Two linguistics students at the University of Education, Winneba, annotated parts of the data in Praat (Boersma and Weenink 2022) at the word level and with glosses, and transcribed the target words according to the IPA. The data annotation was checked with Dr. Kwasi Adomako, a phonologist and native speaker of Akan. Although Anum is a tone language, I have refrained from giving the tonal transcription since lexical tone does not play a role in RVH. For parts of the data, a formant analysis was conducted in Praat. This was done using the default settings in the formant menu.

3.2. Regularly Occurring RVH in Anum

This section provides data on regularly occurring RVH in Anum. Phrasal and sentential contexts within which RVH applies are the NP, the VP, and the clause containing a subject NP and a VP.

3.2.1. RVH within the NP

The first data set comprises NPs of varying complexity. In (8), all nouns are modified, and generally, RVH occurs between a noun and a modifier. The modifier varies, being either a quantifier (8a) or an adjective (8b, c). Note that in (8b), the low vowel allophone [æ] of the adjective occurs due to word-level harmony triggered by the high front vowel. When the adjective ‘slippery’ is combined with the noun ‘ground’, ATR spreads from the low vowel allophone across the word boundary; it does not block RVH.

Figure 1 illustrates the difference in the first and second formants of the final vowel for the noun kʊtʊ ‘hat’ in (8a), spoken by the same speaker. Spoken in isolation, in Figure 1a F1 is at 420 Hz and F2 at 950 Hz. The assimilation due to RVH shown in Figure 1b lowers F1 to 310 Hz and F2 to 760 Hz. Lowering of F1 clearly indicates the assimilation to [+ATR] [u].

In complex NPs containing two modifiers, RVH also occurs between modifiers (8d-e). The quantifier bebree ‘many’ affects the final vowel of the preceding adjective, triggering its assimilation to the [+ATR] vowel. In the case of (8d), the adjective kɔɔ ‘red’ consists of two syllables. According to Ofori (2013), in sequences of two identical vowels, each vowel projects its own syllable. In line with RVH in other cases, RVH targets only the immediately adjacent vowel of the preceding word.

8.	a.	kʊtʊ	bebree		→	[kʊtu bebree]
		‘hat’	‘many’			‘many hats’
	b.	sɪrε	/afi/[æfi]		→	[sɪre æfi]
		‘ground’	‘slippery’			‘slippery ground’
	c.	aklεnɪ	pebi		→	[aklεni pebi]
		‘chicken’	‘small’			‘small chicken’
	d.	kʊtʊ	kɔɔ	bebree	→	[kʊtʊ kɔo bebree]
		‘hat’	‘red’	‘many’		‘many red hats’
	e.	kʊtʊ	ahwε	bebree	→	[kʊtʊ ahwe bebree]
		‘hat’	‘new’	‘many’		‘many new hats’

3.2.2. RVH within the VP

The next set of data concerns RVH within the VP. The process is found between an object and a preceding verb (9a) and within a complex direct object (9b). In (9a), the [+ATR] feature associated with the direct object triggers regressive assimilation in the verb, turning the [−ATR] back mid vowel of the verb stem into a [+ATR] vowel. Note that the assimilation process is not iterative. The surface [+ATR] vowel of the verb does not cause assimilation in the adjacent final vowel of the subject kwamɪ. Hence, the monosyllabic verb as a target of RVH is not a trigger for further assimilation, which is a crucial difference to Vata, where RVH is iterative across sequences of monosyllabic words (Kimber 2011).

9.	a.	[kwamɪ	so	kokosi]
		/kwamɪ	sɔ	kokosi/
		Kwame	take	coconut
		‘Kwame takes coconut.’
	b.	[kofi	æædʑi	mɪkẽ	tei]
		/kofi	a.adʑi	mɪkε̃	tei/
		Kofi	COMPL.eat	my.wife	food
		‘Kofi has eaten my wife’s food.’

In (9b), RVH occurs within the complex direct object ‘my wife’s food’. The [+ATR] possessee ‘food’ triggers regressive assimilation in the preceding [−ATR] possessor. RVH affects the final nasal front mid vowel of the possessor ‘my wife’, turning it into a [+ATR] vowel. In (9b), the subject contains [+ATR] vowels. If it were the [−ATR] subject kwamɪ, RVH would apply between the [+ATR] verb and the subject. The [+ATR] allophone of the low vowel does not block RVH, as shown in (8b) above.

3.2.3. RVH within the Clause

The data presented in this section illustrate the following three facts. First, within the clause, RVH applies between a subject and a verb, between a verb and an object, and within simple and complex subject and object NPs. Second, multiple instances of RVH occur within a sentence, though not iteratively. Third, the occurrence of RVH within or between a constituent does not depend on the number of words within a constituent or a clause.

The data set in (10) extends the data in (7) cited from Obeng (1995) on subject-verb clauses. In (7), we saw that all sentences consisted of two lexical words, a subject and a verb, and RVH applied between the two. (10a) repeats the sentence (7a) for reference. The remaining data in (10) are more complex, increasing the number of words in the clause. (10b) consists of four words, of which the determiner is a function word cliticizing to the adjective. RVH takes place between the adjective and the subject noun, and between the verb and the preceding determiner. Likewise, the clause in (10c) contains three words, and RVH applies between the verb and the preceding adjective of the subject NP. Hence, RVH applies independently of the complexity of the subject constituent.

10.	a.	[kwami	dʑi]
		/kwamɪ	dʑi/
		Kwame	eat
		‘Kwame eats.’ (see (11))
	b.	[kwami	teente	æ	dʑi]
		/kwamɪ	teente	a	dʑi/
		Kwame	tall	DET	eat
		‘The tall Kwame eats.’
	c.	[kwamɪ	kɔo	dʑi]
		/kwamɪ	kɔɔ	dʑi/
		Kwame	fair	eat
		‘Fair Kwame eats.’

The next type of data in (11) shows more complex subject NPs consisting of two modifiers. The complex NPs appear in a sentence with a direct object following the verb. Both sentences contain five words. RVH applies between the subject constituent and the verb in (11a), affecting the final vowel of the second modifier kɔɔ ‘red’. In addition, RVH applies between the subject noun and its modifier. In (11b), RVH applies twice, once within the subject NP, and once between the direct object and the verb.

11.	a.	[kwami	timi	kɔo	dʑi	kokosi]
		/kwamɪ	timi	kɔɔ	dʑi	kokosi/
		Kwame	short	fair	eat	coconut
		‘Short fair Kwame eats coconut.’
	b.	[kwamɪ	kɔo	tũũ	so	kokosi]
		/kwamɪ	kɔɔ	tũũ	sɔ	kokosi/
		Kwame	fair	dark	take	coconut
		‘Fair dark-spotted Kwame takes coconut.’

The data in (12) illustrate multiple occurrences of RVH within a clause. RVH is independent of the number of words, which successively increases from four to six in (12). In all cases, RVH applies between the first modifier and the subject noun, as well as between the verb and the direct object. (12c) shows a case where RVH additionally applies between the sentence-final numeral and the preceding modifier of the object NP.

12.	a.	[kwami	teente	so	kokosi]
		kwamɪ	teente	sɔ	kokosi
		Kwame	tall	buy	coconut
		‘Tall Kwame buys coconuts.’
	b.	[kwami	teente	so	kokosi	blɔblɔ]
		kwamɪ	teente	sɔ	kokosi	blɔblɔ
		Kwame	tall	buy	coconut	sweet
		‘Tall Kwame buys sweet coconuts.’
	c.	[kwami	teente	so	kokosi	blɔblo	du]
		kwamɪ	teente	sɔ	kokosi	blɔblɔ	du
		Kwame	tall	buy	coconut	sweet	ten
		‘Tall Kwame buys ten sweet coconuts.’

3.3. Blocking of RVH in Anum

So far, the occurrence of RVH appears to be a regular phonological process in Anum, applying in all contexts within a sentence. However, in certain syntactic configurations the regular process of RVH is blocked. The blocking of RVH is indicated by ‘||’. Two blocking contexts are discussed here. First, RVH is blocked between a VP and a following time adverbial. Second, RVH is blocked between two VPs in a serial verb construction.

The data in (13) present a clause that contains a subject, verb, object, and an additional, sentence-final time adverbial dudu ‘on Monday’. The time adverbial is a [+ATR] word. In both examples, the word preceding the time adverbial is a [−ATR] word. The structural prerequisite for RVH is thus met. However, RVH is blocked between the time adverbial and the preceding word. The final [−ATR] vowel of the adjective blɔblɔ ‘sweet’ in (13a) and that of the object noun ŋkatɪ ‘groundnut’ in (13b) do not assimilate to [+ATR]. If the domain of RVH were the entire clause, the prediction would be that RVH would occur in (13), which is, however, not the case. Note that independently of the blocking, RVH occurs within a subject NP and between an object and a verb in (13a), as shown before in the previous sections.

13.	a.	S(ω ω)		V(ω)	O(ω ω)		||	Adv (ω)
		[kwami	teente	so	kokosi	blɔblɔ		dudu]
		kwamɪ	teente	sɔ	kokosi	blɔblɔ		dudu
		Kwame	tall	buy	coconut	sweet		Monday
		‘Tall Kwame buys sweet coconuts on Monday.’
	b.	S(ω)	V(ω)	O(ω)	||	Adv (ω)
		[kofi	dʑi	ŋkatɪ		dudu]
		kofi	dʑi	ŋkatɪ		dudu
		Kofi	eat	groundnut		Monday
		‘Kofi eats groundnuts on Monday.’

The second type of data showing a blocking of RVH comes from serial verb constructions. Verb serialization is a common feature of Kwa languages (Kropp Dakubu 1988; Shluinsky 2017). The characteristic property of a serial verb construction (SVC) is that two finite verbs optionally including their complements appear within a single clause (see Schachter 1974 for Akan SVCs), and that the two verbs together present a single event rather than two consecutive events (for an overview, see Shluinsky 2017). Syntactically, each of the verbs is contained in its own VP, and the two VPs are sisters within a higher functional projection (Baker 1989). Typically, the verbs in a serial verb construction share identical tense and aspect markers, which according to Baker (1989) is an argument that each of the verbs functions as a head of its own VP. The process of RVH regularly takes place within a VP, as was shown above in Section 3.2.2. However, the process of RVH is blocked between two VPs as shown in (14). Both VPs contain a verb and a direct object as a complement. The complement of the first VP is a [−ATR] word and the verb of the second VP is a [+ATR] word. Hence, the structural prerequisite for RVH is met. Nevertheless, RVH does not affect the [−ATR] vowel of the direct object of the first VP. Note that the meaning of the two verbs of the serial verb construction in (14) combines to the meaning of ‘to play’. Thus, to ‘use’ a thing and ‘eat game’ means to ‘play’ with that thing.

Serial verb constructions may only consist of a sequence of two verbs. According to a recent study by Tyler and Kastner (2022) on SVCs in Degema, an Edoite language spoken in Nigeria, there is pressure for the two verbs to form a single prosodic phrase. If that were the case in Anum, too, RVH would be expected to occur between the two verbs. This has to be left open for future research since we currently do not have data for this particular case.

14.		S	(VP	)	||	(VP	)
	a.	[kofi	blε	ɔkɔtɔ		odʑi	ohili]
		kofi	blε	ɔkɔtɔ		odʑi	ohili
		kofi	use	crab		eat	game
		‘Kofi plays with the crab.’
	b.	[kofi	blε	kʊtʊ		odʑi	ohili]
		kofi	blε	kʊtʊ		odʑi	ohili
		kofi	use	hat		eat	game
		‘Kofi plays with the hat.’

4. RVH as a Phrase-Level Process in Anum—An OT Analysis

This section provides an OT-analysis of both word-level (Section 4.2) and phrase-level (Section 4.3) vowel harmony in Anum. The first section (Section 4.1) introduces the syntax-phonology interface framework in which the analysis is couched.

4.1. Background on the Syntax-Phonology Interface

In prosodic phonology, the formation of prosodic structure is usually assumed to depend on syntactic structure, which roughly corresponds to prosodic structure (see Bennett and Elfner 2019 for an overview). Theories on the syntax-phonology interface entertain two major approaches: in one approach the relation between syntactic and prosodic constituency is assumed to be direct, while in the other approach, which I will follow in this paper, it is assumed to be indirect (see e.g., Elordieta 2008; Selkirk 2011; Truckenbrodt 2011; Féry 2017; Bennett and Elfner 2019 for an overview). Following the latter approach, the present study is couched within Match Theory (Selkirk 2011), which proposes that syntactic constituents are mapped into corresponding prosodic constituents, and vice versa. In Optimality Theory (OT) terms, the Match constraints are faithfulness constraints2 that, very generally, match the syntactic constituents ‘word’, ‘phrase’, and ‘clause’ to the corresponding prosodic constituents ‘prosodic word’ (ω), ‘phonological phrase’ (φ), and ‘intonation phrase’ (ι). These interact with prosodic markedness constraints that require certain restrictions on prosodic structure formation. For instance, many languages require binary prosodic constituents with the result that a single prosodic word that constitutes a phonological phrase on its own according to the Match constraints would be merged together with another prosodic word in order to obey Binarity at the phrase level (for Binarity, see Selkirk 2000).

In order to account for word-level and phrase-level harmony, the relevant Match constraints on word and phrase creation are shown in (15) (Selkirk 2011). These input-output Match constraints ensure that strict isomorphy exists between syntactic constituents and prosodic constituents. Hence, each lexical word is matched to a corresponding prosodic word, and each lexically headed syntactic phrase is matched to a corresponding phonological phrase.

15.	a.	Match(LexWd, ω)
		The left and right edges of a lexical word in the input syntactic representation must correspond to the left and right edges of a prosodic word ω in the output phonological representation (Selkirk 2011). Assign one violation mark for every lexical word whose left and right edges do not correspond to the left and right edges of ω.
	b.	Match(LexP, φ)
		The left and right edges of a lexical phrasal projection in the input syntactic representation must correspond to the left and right edges of a phonological phrase φ in the output phonological representation (Selkirk 2011). “Suppose there is a syntactic phrase (XP) in the syntactic representation that exhaustively dominates a set of one or more terminal nodes α. Assign one violation mark if there is no phonological phrase (φ) in the phonological representation that exhaustively dominates all and only the phonological exponents of the terminal nodes in α.” (Elfner 2012, p. 185).

The original proposal of the Match constraints referred to general Match constraints that would match any syntactic word (X⁰) and any syntactic maximal projection (XP) to their corresponding prosodic constituents (Selkirk 2011). The constraint in (15a) is a more restricted Match constraint that takes the general distinction between lexical and function words into account (e.g., Truckenbrodt 1999; Selkirk 1996). In (15b), I assume here a lexically specific phrasal Match constraint that restricts the relevant syntactic input to phrases that are lexically headed (Elfner 2012). This assumption implies the widely held distinction between lexical and functional projections (e.g., Truckenbrodt 1999). As a consequence, functionally headed phrases are ignored in prosodic phrase formation (see Elfner 2012 for a departure from the original Match constraints).

Given that recursivity frequently occurs in syntactic structure, a faithful match of syntactic constituency to prosodic constituency involves recursive prosodic structure by definition. Itô and Mester (2012, 2013) elaborate on this built-in effect of Match Theory, proposing recursive prosodic constituents that appear as a minimal or a maximal constituent (see Figure 2). According to their model, a constituent of the same level dominates a minimal constituent. Conversely, a maximal constituent dominates a constituent of the same level, but is not dominated by a category of the same level. In addition, intermediate constituents dominate and are dominated by a constituent of the same level, rendering this sub-category as either non-minimal or non-maximal. Relevant for the present discussion is the maximal phonological phrase (φ_max), which is not dominated by any further φ, meaning that in syntactic input structure there is no further lexically headed projection.

The proposal I am arguing for here is that in Anum, both word-level harmony and cross-word harmony (RVH) are [+ATR] dominant and regressive in direction. The difference between the two is in the domain. While word-level harmony takes the entire prosodic word as its domain, cross-word harmony is a bounded and domain-sensitive phrase-level process. RVH applies within φ_max, but it is blocked from occurring between φ_max-phrases.

4.2. Word-Level Vowel Harmony in Anum

The approach to modelling word-level vowel harmony taken here is couched within Match Theory (Selkirk 2011) and contains the following components. First, an input-output Match constraint ensures prosodic word formation such that lexical words are matched to corresponding prosodic words (ω) (15a). In addition to the interface constraint, a number of faithfulness and markedness constraints account for the word-level vowel harmony in Anum. Second, a sequential markedness constraint that drives vowel harmony (16a) (Pulleyblank 2002; Mahanta 2008; see also Obiri-Yeboah and Rose 2022 for vowel harmony in Gua). As with the faithfulness constraint, our adaption of this markedness constraint is to propose a domain-sensitive constraint referring to prosodic constituent structure, i.e., the prosodic word (ω). Third, a positional faithfulness constraint that requires feature identity within minimal prosodic words (ω_min) (16b). Fourth, a general faithfulness constraint requiring identity in the feature [+ATR] (16c).

16.	a.	*([−ATR][+ATR])ω
		Suppose a feature sequence of [−ATR] followed by [+ATR] within a prosodic word. Assign a violation mark to [+ATR] segments preceded by [−ATR] segments within the prosodic word (adapted from Mahanta 2008; Pulleyblank 2002).
	b.	Ident-OI[ATR]ωmin
		Every vowel of a minimal prosodic word in the output with the value [αATR] must have a corresponding vowel in the input with the value [αATR]. Assign a violation mark for a vowel in ωmin that does not correspond in its [ATR] value with that vowel of the lexical word in the input.
	c.	Ident-IO[+atr]
		Every vowel in the input with the value [+ATR] must have a corresponding vowel in the output with the value [+ATR]. Assign a violation mark for each input [+ATR] vowel that has no corresponding [+ATR] vowel in the output.
	d.	Ranking of word-level vowel harmony constraints in Anum
		Match(LexWd, ω), Ident-OI(atr)ωmin, *([−ATR][+ATR])ω >> Ident-IO(+ATR)

The sequential markedness constraint *([−ATR][+ATR])ω (16a) is a specific version of an Agree constraint (Baković 2000) encoding directionality (Mahanta 2008). The drawback of assuming a general Agree constraint is its symmetrical nature, which favours both [−ATR] and [+ATR] sequences. The fact that disharmonic words exist in Anum (see (6) above) would, however, require a decision on which of the feature values should be agreed on. As (6) showed, disharmonic words only have the structure of [+ATR] [−ATR], meaning that [+ATR] does not spread progressively, and, at the same time, [−ATR] does not spread regressively targeting [+ATR]. This fact provides an argument to assume [+ATR] dominance in Anum (see Casali 2012 for [+ATR] dominance in Akan). Breaking up Agree into sequential markedness constraints ensures that all vowels in a prosodic word agree in their [ATR] feature, and that directionality is encoded in the assimilation process. A violation of this constraint occurs at a juncture of two distinct feature specifications, i.e., between a sequence of [−ATR] [+ATR] or vice versa. This binary decision is different from a gradient version employed in Gua, where the analysis relies on a sequential markedness constraint that targets vowel sequences at any distance (Obiri-Yeboah and Rose 2022). This way, the analysis for Gua allows for gradient violation rather than assigning a penalty only at disharmonic junctures. Alternative approaches to vowel harmony include Alignment constraints, which express directionality equally well. Like the gradient sequential markedness constraint proposed for Gua (Obiri-Yeboah and Rose 2022), an Alignment constraint counts violations of feature identity up to a prosodic constituent boundary, hence evaluating vowel harmony in a gradient way (see Walker 2012 for an overview). Our analysis for Anum favours a non-gradient vowel harmony trigger.

The positional faithfulness constraint (16b) makes reference to prosodic constituency, preventing the [ATR] feature value of a minimal prosodic word from changing compared to the corresponding vowel of a lexical word in the input. Applying it in an Output-Input correspondence ensures that the vowel in the output structure mapped from a lexical word in syntactic representation into a corresponding (minimal) prosodic word in phonological representation is identical to the corresponding vowel in the lexical word. For positional faithfulness in relation to vowel harmony, see for instance Beckman (1997).

Finally, a general faithfulness constraint requiring identity of the feature [+ATR] (16c) militates against harmony. Ranking it below the harmony-driving markedness constraint (Markedness ≫ Faithfulness) captures any process of assimilation within OT (Mahanta 2008; Walker 2012). The positional faithfulness constraint, however, necessarily dominates the general faithfulness constraint (16d).

Turning to the analysis, the evaluation of the [+ATR] word in (5b), ku ‘to dig’, is given in Tableau (17). The 3.SG pronoun mʊ has a [−ATR] vowel (see Section 2.1). According to Selkirk (1996), three possible prosodic structures with respect to clitization of the function word can be assumed: First, affixial clitization (mʊ-(ku)_ω)_ω, which leads to a recursive prosodic word structure but keeps the left and right edges of the lexical word in prosodic structure. Second, internal clitization (mʊ-ku)_ω, which leads to a mismatch between syntactic and prosodic structure, yet no recursivity. Third, a non-integration of the function word in prosodic word structure, free clisis (mʊ-(ku)_ω)_φ, where the function word is parsed by a φ–phrase. This latter candidate can be ruled out by a general parsing constraint requiring exhaustive parsing, which is omitted in the present discussion for the ease of simplicity. The Match constraint requires that the word stem of the lexical word ku be mapped into a prosodic word (ω), and that the pronoun, by virtue of it being a function word, is not mapped into a prosodic word. Cliticization to the ω occurs due to a general constraint requiring exhaustive parsing, which is not shown here. The resulting affixial clitization structure is shown in candidates (17a–d) where the entire prosodic word contains both the verb stem and the pronoun. Candidate (17e) shows internal clitization of the pronoun.

The optimal candidate (17a) obeys all of the constraints. Positional faithfulness is fulfilled since the stem vowel keeps its [+ATR] feature in the matched minimal prosodic word. The sequential markedness constraint *([−ATR][+ATR])ω requires regressive assimilation of the [+ATR] feature on the pronominal prefix, which is fulfilled. Finally, feature assimilation to [+ATR] does not violate the general Ident constraint since the prefix contains a [−ATR] feature. In the case of a candidate being faithful to the input, (17b), the harmony-driving sequential markedness constraint is violated. If the [ATR] feature value of the stem vowel in the minimal phonological word (17c, d) is changed, the positional faithfulness constraint Ident-OI[ATR]ω_min is violated. Both candidates also violate the general Ident constraint because the stem vowel is not identical in its [+ATR] value of the input. Candidate (17e) violates the Match constraint since the left and right edge of the prosodic word do not match with the left and right edge of the lexical word.

17.
		[mʊ-[ku]]    3.SG dig    (see (5b))	Match(LexWd, ω)	Ident-OI[ATR]ωmin	*([−ATR][+ATR])ω	Ident-IO[+ATR]
	a.	(mu-(ku)ωmin)ω
	b.	(mʊ-(ku)ωmin)ω			*!
	c.	(mʊ-(kʊ)ωmin)ω		*!		*
	d.	(mu-(kʊ)ωmin)ω		*!		*
	e.	(mʊ-ku)ω	*!		*

The Tableau in (18) illustrates that disharmonic words can occur in Anum as well, however with the structural restriction that any [+ATR] syllable must precede a [−ATR] syllable. Given the sequential markedness constraint *([−ATR][+ATR])ω driving [+ATR] harmony in Anum, a disharmonic word such as humε ‘swell’ in (18b) does not violate the markedness constraint. The harmony trigger simply does not occur on the right (see Mahanta 2008 for similar facts in Assamese). In the remaining candidates (18a, c, d), the change of the [ATR] of one syllable incurs a violation of the positional faithfulness Ident-OI[ATR]ω_min constraint and/or the general Ident constraint. Candidate (18d) has the reverse [ATR] pattern of both stem vowels and thus violates Ident-OI[ATR]ω_min twice. All candidates obey the Match constraint since a lexical word is matched into a minimal prosodic word. There are no clitics as noun class prefix or the like that would need to be taken into account to arrive at a candidate deviating in prosodic structure that would violate the Match constraint.

18.
		[humε] swell (see (6))	Match(LexWd, ω)	Ident-OI[ATR]ωmin	*([−ATR][+ATR])ω	Ident-IO[+ATR]
	a.	(hume)ωmin		*! (σ2)
	b.	(humε)ωmin
	c.	(hʊmε)ωmin		*! (σ1)		*
	d.	(hʊme)ωmin		*!* (σ1+2)	*	*

4.3. Phrase-Level Analysis of RVH

This section provides an analysis of RVH in terms of prosodic phrasing. The proposal is that RVH is a domain-sensitive phonological assimilation process that operates within φ-phrases. The properties of RVH are identical to those of word-level harmony. In particular, the process is regressive, hence unidirectional, and it is [+ATR] dominant, i.e., [+ATR] is the triggering feature. The approach to phrasal RVH contains the following components. First, the phrasal Match constraint in (15b) requires lexically headed syntactic phrases to be matched to corresponding phonological phrases.

Second, the sequential markedness constraint in (16b) triggering vowel harmony is adapted to refer to prosodic word structure. The adaptation concerns a ban on a [−ATR][+ATR] sequence across prosodic word boundaries (19a). The direct reference to prosodic constituent structure fits into the syntax-phonology match approach proposed here given that prosodic word and phrase structure emerges through the Match constraints in (15). With this, it is an adaptation of the proposal of phrasal vowel harmony for Akan (Kügler 2015) in that the sequential markedness constraint specifically refers to prosodic structure, which is a result of the syntax-phonology interface anyway. Note that the sequential markedness constraint evaluates the juncture between a [−ATR] and [+ATR] sequence, requiring a prosodic word boundary between the two feature specifications. Alternative formulations of this type of constraint have been proposed by Pulleyblank (2002), who suggests a sequential markedness constraint that evaluates any [−ATR] vowel preceding a [+ATR] vowel in a distal fashion, extending the sequential prohibition to long-distance environments. This way, the sequential prohibition scopes over all preceding [−ATR] vowels and assigns violations in a gradient fashion for all instances of a preceding [−ATR] vowel. See also Obiri-Yeboah and Rose (2022), who suggest the gradient version of the sequential markedness constraint for Gua.

The constraint formalized in (19a) triggers regular RVH as it penalizes the non-application of RVH across a prosodic word boundary. This prosodic markedness constraint is reminiscent of a constraint on tone spreading proposed in Kula and Bickmore (2015). The authors account for inter-word H tone doubling in Copperbelt Bemba, where an H tone on a word-final TBU of one word spreads onto the first TBU of the following word. They however do not relate their account of tone spreading to prosodic words.

The third component is a constraint that militates against cross-word harmony (19b). I suggest using a constraint that requires sharp alignment of the [+ATR] feature with the boundary of a prosodic word. This way, the constraint demands that the [+ATR] feature not be shared across a prosodic word juncture. The requirement is fulfilled by a CrispEdge constraint whereas classic Alignment constraints would fail (Itô and Mester 1999) because in a situation of cross-word harmony, the [+ATR] feature would be aligned with the right edge of the triggering word, although it also reaches across the boundary. As Itô and Mester (1999) argue, the CrispEdge constraint is independent of general alignment.

19.	a.	*[−ATR])ω ω([+ATR]
		Assign a violation for a vowel at the right edge of a prosodic word ω that is [−ATR] and is followed by a vowel at the left edge of a following prosodic word ω that is [+ATR]. (adapted from Kügler 2015, p. 184; Obiri-Yeboah and Rose 2022, p. 187)
	b.
		Assign a violation mark for a sequence of a shared feature across a prosodic word boundary if the [+ATR] feature is not crisply aligned with left edge of a prosodic word ω.

Alternative analyses of phrasal vowel harmony include constraints that demand explicit restriction of iterative spreading of the feature [+ATR] further than the prosodic word juncture (Kimber 2011; Obiri-Yeboah and Rose 2022). Kimber proposes a serial harmony account using a constraint of the S_HARE family (McCarthy 2011). S_HARE(F) requires a violation mark for every pair of adjacent elements that are not linked to the same token of [F]. The sharing of a feature accounts for word-level vowel harmony in her account. To account for cross-word harmony, Kimber suggests an extension of the domain of the S_HARE constraint from the domain of the prosodic word to the domain of the prosodic juncture. This move requires identical features on both sides of the prosodic juncture. Similarly, Obiri-Yeboah and Rose (2022) propose a constraint that targets the prosodic juncture, allowing only the vowels at the juncture to be identical in the [+ATR] feature. They capture their analysis by proposing a co-phonology account within harmonic span theory (McCarthy 2004). Their constraint Adjacency[+ATR] starts from the triggering [+ATR] head vowel and requires a violation of all preceding non-head [+ATR] vowels except for the immediately adjacent vowel at the juncture. Their analysis requires a co-phonology account because the Adjacency[+ATR] constraint has to be low ranked in word-level harmony but has to dominate the triggering sequential markedness constraint at the phrasal level.

The general occurrence of RVH, either within an NP or within a VP, is illustrated with these constraints in Tableaux (20), (21), and (22). Both constraints in (19) together limit the spreading of [+ATR] to the immediately left-adjacent syllable.

The data in Tableau (20), taken from (8a) above, show the application of RVH within the NP. The noun kʊtʊ ‘hats’ and the modifier bebree ‘many’ form an NP, and are matched accordingly to a φ-phrase. Note that we assume that an adjective forms a head of its own adjective phrase, which is adjoined to the NP projection (see e.g., Adger 2003). As a head of an AdjP it would be matched to a separate φ-phrase yielding a structure like [_NP [_AdjP modifier] [_NP noun]]. Adjoining the two phrases forms an NP on its own. For reasons of simplicity, we do not illustrate the NP-internal structure here since it does not bear on the analysis. RVH applies within this φ-phrase, more specifically between the modifier and the noun. The optimal candidate (20a) obeys the higher-ranked sequential markedness constraint *[−ATR])ω ω([+ATR] (19a) triggering cross-word harmony but violates the CrispEdge constraint since the [+ATR] feature of the modifier is not crisply aligned with the left edge of ω but spreads regressively across the word boundary. The feature assimilation incurs a violation of the positional faithfulness constraint Ident-OI[ATR]ω_min for the noun. Cross-word harmony leads to a surface disharmonic word, violating the word-level harmony-triggering sequential markedness constraint *([−ATR][+ATR])ω. This shows that the phrasal harmony-triggering sequential markedness constraint necessarily needs to dominate the word-level harmony-triggering sequential markedness constraint: *[−ATR])ω ω([+ATR] >> *([−ATR][+ATR])ω.

Candidate (20b), being faithful to the input, shows no RVH. It thus violates the high-ranked sequential markedness constraint *[−ATR])ω ω([+ATR] that requires cross-word harmony. Candidate (20c) shows unbounded or iterative application of RVH and is ruled out because of the violations of the general faithfulness constraint requiring input-output identity with respect to [+ATR] segments. Although the word-level Ident constraint that requires faithfulness with respect to [ATR] within the minimal prosodic word is violated twice, the equal ranking of the CrispEdge, the Ident-OI[ATR]ω_min and the sequential markedness constraint *([−ATR][+ATR])ω takes no decision on this double violation. As a lexical word, the noun represents a minimal prosodic word, and since both vowels are assimilated to [+ATR], cross-word harmony incurs two violations.3 Compared to the optimal candidate (20a), candidate (20c) violates the low-ranked general Ident constraint, even twice. Note that both candidates (20a) and (20c) violate the CrispEdge constraint because of cross-word harmony. Iterative assimilation of the [+ATR] feature is thus excluded by a word-level harmony constraint. Candidate (20d) shows an instance of two φ-phrases, each word projecting its own φ-phrase, violating the high-ranked Match constraint. Candidate (20e) shows progressive [−ATR] spreading to the modifier. This candidate is ruled out based on the low-ranked general Ident constraint requiring identity in [+ATR].

20.
		[NP kʊtʊ bebree] hats many (see (8a))	Match(LexP, φ)	*[−ATR])ω ω([+ATR]	CrispEdge[+ATR]ω	Match(LexWd, ω)	Ident-OI[ATR]ωmin	*([−ATR][+ATR])ω	Ident-IO[+ATR]
	a.	((kʊtu)ωmin (bebree)ωmin)φ			*		*	*
	b.	((kʊtʊ)ωmin (bebree)ωmin)φ		*!
	c.	((kutu)ωmin (bebree)ωmin)φ			*		**		*!*
	d.	((kʊtʊ)ωmin)φ ((bebree)ωmin)φ	*!	*
	e.	((kʊtʊ)ωmin (bεbree)ωmin)φ			*		*	*	*!

The next Tableau (21) illustrates cross-word harmony between an object and a preceding verb. The object and the modifier form an NP, and together with the verb, the NP is embedded in the VP. The data are taken from (12b) above. The faithful candidate (21a) violates the sequential markedness constraint *[−ATR])ω ω([+ATR] since no cross-word harmony takes place. The optimal candidate (21b) shows cross-word harmony, thus obeying the sequential markedness but violating CrispEdge since [+ATR] is not crisply aligned at the left edge of the ω of the noun. It also violates the lower-ranked positional faithfulness constraint Ident-OI[ATR]ω_min since the [ATR] of the verb changed its specification. Candidate (21c) shows an instance of a mismatch between syntactic and prosodic constituent structure and is ruled out because of the high-ranked Match constraint.

21.
		[VP sɔ [NP kokosi blɔblɔ]] buy  coconut sweet (see (12b))	Match(LexP, φ)	*[−ATR])ω ω([+ATR]	CrispEdge[+ATR]ω	Match(LexWd, ω)	Ident-OI[ATR]ωmin	*([−ATR][+ATR])ω	Ident-IO[+ATR]
	a.	((sɔ)ω ((kokosi)ω (blɔblɔ)ω)φ)φ		*!
	b.	((so)ω ((kokosi)ω (blɔblɔ)ω)φ)φ			*		*
	c.	(((sɔ)ω)φ ((kokosi)ω (blɔblɔ)ω)φ)φ	*!	*

The next example illustrates that cross-word harmony is bound to the adjacent word, and does not iterate further to yet another word. As was shown in Tableau 18, cross-word harmony is bound to the immediately adjacent vowel of the preceding prosodic word and thus does not iteratively affect further vowels, unlike for instance in Vata (Kimber 2011) or in Nawuri (Casali 2002). The SVO sentence in (22) consists of a subject NP and a VP that contains a verb and an object NP. These are matched to two prosodic phrases, one corresponding to the subject NP and one to the VP, which contains a further φ-phrase that corresponds to the object NP. Candidate (22a) is faithful to the input, showing no cross-word harmony between the verb and the subject. As before in (20) and (21), the sequential markedness constraint *[−ATR])ω ω([+ATR] is violated because no cross-word harmony takes place. The optimal candidate (22b) obeys the sequential markedness constraint but violates the lower-ranked CrispEdge constraint, again because here the [ATR] of the object is not crisply aligned with the left edge of the prosodic word that contains the object noun. The candidate in (22c) shows iterative cross-word harmony through the verb to the final vowel of the subject noun. In this case, the CrispEdge constraint is violated twice since the [ATR] of the object is not crisply aligned with the prosodic word twice, each instance counting as an alignment violation. Candidate (22d) illustrates a mismatch between syntactic and prosodic constituent structure in that the subject NP is integrated in the prosodic phrase that is matched to the corresponding VP. Note that the issue of subjects and their phrasing becomes important below as cross-word harmony regularly occurs between a subject and a verb (see e.g., the data in (7), (10), and (11) above).

22.
		[NP kwamɪ] [VP sɔ [NP kokosi]] Kwame  buy coconut (see (9a))	Match(LexP, φ)	*[−ATR])ω ω([+ATR]	CrispEdge[+ATR]ω	Match(LexWd, ω)	Ident-OI[ATR]ωmin	*([−ATR][+ATR])ω	Ident-IO[+ATR]
	a.	((kwamɪ)ω)φ ((sɔ)ω ((kokosi)ω)φ)φ		*!
	b.	((kwamɪ)ω)φ ((so)ω ((kokosi)ω)φ)φ			*		*
	c.	((kwami)ω)φ ((so)ω ((kokosi)ω)φ)φ			**!		**	*
	d.	((kwamɪ)ω (sɔ)ω (kokosi)ω)φ	*!	*

Having established the general occurrence of RVH in Anum, further issues have to be discussed in order to account for the RVH patterns shown in Section 3. First, the blocking of RVH shown in Section 3.3 needs to be accounted for (Section 4.4). The analysis proposed so far shows that RVH applies between words in general. To account for the blocking cases, I argue for an extension of the domain-sensitive constraint (19b) that penalizes the association of the [+ATR] feature across prosodic constituents.

Second, the data from Obeng (1995) shown in (7) needs an extended analysis to address the special role of subject NPs in Anum (see Obiri-Yeboah and Rose 2022 for the special role of subject NPs in the neighbouring dialect Boso). The syntax-phonology interface requires that a subject NP and a VP each be matched to a separate φ-phrase. The analysis of blocking of RVH takes φ-phrase boundaries into account. As shown in Section 4.5, the blocking account discussed in the next Section 4.4. wrongly predicts blocking of RVH between a subject and a verb.

4.4. Analysis of Blocking of RVH

The data discussed in the previous Section 4.3 suggest a phrase-level analysis of RVH. The blocking contexts (see Section 3.3) show that RVH is blocked at the edges of particular syntactic phrases, i.e., the AdvP in (13) and the VP in (14). Assuming the formation of prosodic phrases using Match Theory (Selkirk 2011) (Section 4.1) in combination with a theory on recursive prosodic phrasing (Itô and Mester 2012), the proposal is that RVH is blocked at the edge of a maximal phonological phrase (φ_max). In terms of OT, this blocking is achieved by a prosodic markedness constraint CrispEdge that is domain-sensitive with respect to φ_max. This constraint is formalized in (23), and it represents an extension of the CrispEdge constraint referring to the level of the prosodic word (19b). Generally, the CrispEdge constraints avoid multiple linking of features between prosodic constituents (Itô and Mester 1999; Selkirk 2011). The constraint in (23) has been shown to account for blocking of RVH in Akan (Kügler 2015).

23.	CrispEdge[+ATR]φmax
	Spreading of the feature [+ATR] is prohibited across the edge of a maximal phonological phrase. Assign a violation mark if [+ATR] is not crisply aligned with φmax. (Kügler 2015, p. 198)

The constraint in (23) makes reference to recursion-based subcategories like φ_max. The question remains though as to why it is not just a phonological phrase boundary that represents the relevant prosodic category? In other words, why does the analysis refer to a maximal phonological phrase? The data on VP-internal RVH discussed in Section 3.2.2 suggest that RVH crosses φ-phrase boundaries. Consider the VP of example (9a), which is repeated in (24a). The VP contains a verb plus a following direct object NP. Application of the Match constraints (15) yields a φ-phrase containing the object NP, and a second φ-phrase containing the verb plus object NP. The application of RVH then crosses the left edge of a φ-phrase. Compare this case with the blocking case of a time adverbial in (13), repeated in (24b). Match phrase (15b) predicts that the object NP, the VP, and the AdvP each match to a φ-phrase. The VP and the AdvP are syntactically adjacent while the object NP is embedded in the VP. In (24b), RVH does not apply between the time adverbial dudu ‘Monday’ and the preceding object ŋkatɪ ‘groundnut’. The non-application of RVH cannot be accounted for if one assumes the presence of just a φ-phrase boundary. (24a) shows that RVH applies in spite of the presence of a φ-phrase boundary; (24b) shows that RVH does not apply. Hence, there must be a distinction between the types of boundaries in (24a) and (24b).

24.	a.	[VP sɔ	[NP kokosi]]
		(so	(kokosi)φ)φ
		buy	coconut
	b.	[VP dʑi	[NP ŋkatɪ]]	[AdvP dudu]
		(dʑi	(ŋkatɪ)φ)φ	(dudu)φ
		eat	groundnut	Monday

The solution adopted here is to assume recursion-based subcategories (Itô and Mester 2012) (see Figure 2 above). The syntax-phonology match is illustrated in (25). There is no syntactic analysis of the time adverbial available for Anum. For Akan, Saah (1994) suggests an analysis of a time adverbial as an adjunct of TP. Let us assume this adjunct-of-TP analysis for Anum as well, thus a time adverbial takes scope over the whole sentence. Given this syntactic structure, both the VP and the AdvP have no further lexical phrasal projection above them. Following the Match constraints (15), the VP and the AdvP form a φ_max. This means that all daughters of TP which contain a lexically headed phrase are mapped into a φ_max. The prosodic structure arrived at clearly makes a distinction between φ_max and φ_non-max. A non-maximal φ-phrase is not crucial for RVH as it may cross φ-phrases that are dominated by a further φ-phrase (e.g., (12b) and its analysis in (21)). However, the crucial context for the blocking of RVH is met at a maximal φ-phrase boundary, in the example here between the time adverbial and the adjacent, VP-final word. Note that the match between CP or TP and its corresponding intonation phrase (ι) is not discussed in depth here but follows a Match constraint like (15) that targets the clause (Selkirk 2011).

Turning to an OT analysis of the blocking contexts, the ranking of the domain-sensitive markedness constraint CrispEdge in (23) needs to be higher than the sequential markedness constraint *[−ATR])ω ω([+ATR] that triggers RVH because CrispEdge forces the blocking of RVH at φ_max. Candidate (26b) shows RVH and thus incurs a violation of the high-ranked domain-specific markedness constraint. The high-ranked phrasal Match constraint requires faithful prosodic phrasing, resulting in isomorphic syntactic and prosodic phrases. Any candidate violating faithful syntax-phonology match in any possible way is penalized; as an example, (26c) presents a prosodic phrasing that is not faithful to the input in that it integrates the time adverbial into the phonological phrase of the VP. The optimal candidate (26a) violates the sequential markedness constraint *[−ATR])ω ω([+ATR] triggering cross-word harmony since no RVH has occurred. Such candidates have been losing candidates in the previous Tableaux, as RVH was mandatory. Here, in the case of the blocking context, the CrispEdge constraint targeting maximal phonological phrases demands no cross-word harmony across a maximal phonological phrase boundary. Hence the lower-ranked harmony-triggering constraint is violated.

26.
		[VP dʑi [NP ŋkatɪ]] [AdvP dudu] eat  groundnut on Monday (see (13))	Match(LexP, φ)	CrispEdge[+ATR]φmax	*[−ATR])ω ω([+ATR]	CrispEdge[+ATR]ω	Match(LexWd, ω)	Ident-OI[ATR]ωmin	*([−ATR][+ATR])ω	Ident-IO[+ATR]
	a.	(dʑi (ŋkatɪ)φ)φmax (dudu)φmax			*
	b.	(dʑi (ŋkati)φ)φmax (dudu)φmax		*!		*		*	*
	c.	(dʑi (ŋkati)φ (dudu)φ)φmax	*!	*		*		*	*

Like the time adverbial, in serial verb constructions each of the VPs form their own φ_max and hence, blocking of RVH is predicted between the two VPs (see the data in (14) above). I follow Baker (1989), who argues that in syntactic representation the two VPs are sisters of a higher functional projection. Each verb functions as the head of its own VP, and as a head each verb receives tense and aspect markers. Syntactically there are thus two adjacent VPs that are not dominated by any further lexical projection. Hence, the phrasal Match constraint (15b) predicts that each VP forms a maximal φ-phrase. The blocking of RVH between the two VPs is then predicted by the CrispEdge constraint that targets maximal phonological phrases (23).

The analysis of phrasal vowel harmony so far assumes that RVH applies regularly within φ-phrases and across non-maximal φ-phrases, and that it is blocked at the edges of φ_max. However, one crucial issue still remains to be discussed: the special role of subject NPs. Their prosodic phrasing leads to an analytical dilemma under the current analysis. The next section presents this analytical conundrum and discusses different options to address the particular syntactic role subject constituents play in Anum.

4.5. The Role of Subject NPs in Prosodic Phrasing in Anum

The original data of Obeng (1995) presented in (7) show that RVH occurs between a subject and a verb. The analysis so far suggests that a subject and a verb are prosodically phrased together, i.e., (subject verb)φ, as RHV occurs within phonological phrases. However, if one follows a general syntactic analysis in which the subject NP occupies a higher projection outside the VP (e.g., SpecTP; [_TP [_NP subject] [_VP verb]]), the phrasal Match constraint (15b) predicts that a subject NP is matched to a corresponding maximal φ-phrase since the subject NP has no further lexically headed projections above it. The verb also forms a separate maximal φ-phrase for the same reason. Under this syntactic analysis, the hitherto proposed phrasal approach to cross-word harmony runs into the analytical dilemma illustrated in Tableau (27). The prosodic phrase formation of two adjacent φ_max-phrases predicts blocking of RVH between the subject and the verb, as illustrated by the winning candidate (27a). The alleged winner (27a) obeys the Match constraint, resulting in the prosodic phrasing of two separate φ_max-phrases, and no RVH. The blocking of RVH is achieved by the CrispEdge requirement, which predicted the correct blocking behaviour of cross-word harmony in the case of adverbial phrases and serial verb constructions. In contrast, the actual attested data are represented in candidate (27d), which, however, violates the high-ranked phrasal Match constraint. The unfaithful prosodic phrasing rules out the attested candidate and candidate (27c). If RVH applies across the edges of maximal φ-phrases as shown in candidate (27b), the CrispEdge constraint demanding no feature assimilation across maximal φ-phrases is violated. The high ranking of the faithfulness constraint excludes candidates in which prosodic structure is not isomorphic to syntactic structure. Hence the analytical conundrum.

27.
		[NP kwamɪ] [VP dʑi]  Kwame  eat ‘Kwame eats.’ (see (7a))	Match(LexP, φ)	CrispEdge[+ATR]φmax	*[−ATR])ω ω([+ATR]	CrispEdge[+ATR]ω	Match(LexWd, ω)	Ident-OI[ATR]ωmin	*([−ATR][+ATR])ω	Ident-IO[+ATR]
	a.	(kwamɪ)φmax (dʑi)φmax			*
	b.	(kwami)φmax (dʑi)φmax		*!				*	*
	c.	((kwamɪ)φ (dʑi)φ)φmax	*!		*
	d.	((kwami)φ (dʑi)φ)φmax	*!					*	*

The crucial question now remains of how to account for a non-maximal φ-phrase boundary between the subject NP and the verb as in candidate (27d). There are in principle two possible solutions. The first one concerns prosodic well-formedness constraints that may restructure faithful syntax-phonology phrasing, leading to a non-isomorphic prosodic structure. The second one concerns the nature of the input itself. The two solutions to the analytical dilemma are each discussed in the following.

Concerning prosodic well-formedness constraints, it would be necessary to assume a constraint that is ranked above the faithfulness Match constraint which would solve the analytical problem by allowing a prosodic phrasing that is deviant from faithful syntactic structure. Based on the data in (7), one might be tempted to assume a constraint on the size of prosodic constituents. A requirement on prosodic constituency size has been proposed for different languages (Selkirk 2000, 2011) and can be formulated in terms of a requirement that φ-phrases be binary (see Obiri-Yeboah and Rose 2022 for such an account for phrasing in Boso). For the present purpose, it suffices to assume a general constraint on Binarity as formulated in (28). Ranking of (28) above the faithfulness Match constraint leads to a restructuring of φ-phrases, resulting in prosodic constituency that is no longer isomorphic to syntactic constituency.

28.	Binarity–φ
	A φ-phrase consists of two ω. Assign a violation mark if a φ-phrase consists of less or more than two ω’s.

As the reanalysis of (27) in (29) shows, the addition of a markedness constraint like Binarity–φ renders the previously winning candidate (27a) non-optimal—it violates Binarity–φ twice since both φ-phrases consist of only one ω each. Candidate (27c) obeys Binarity–φ but does not show regular RVH and thus fatally violates the sequential markedness constraint *[−ATR])ω ω([+ATR] that requires RVH to occur. Restructuring prosodic phrasing according to Binarity–φ makes candidate (27d) the winner, which is the optimal output in (29). The addition of the prosodic well-formedness constraint thus accounts for the data in (7).

29.
		[NP kwamɪ] [VP dʑi]   Kwame  eat   ‘Kwame eats.’ (see (7a))	Binarity–φ	Match(LexP, φ)	CrispEdge[+ATR]φmax	*[−ATR])ω ω([+ATR]	CrispEdge[+ATR]ω	Match(LexWd, ω)	Ident-OI[ATR]ωmin	*([−ATR][+ATR])ω	Ident-IO[+ATR]
	a.	(kwamɪ)φmax (dʑi)φmax	*!*			*
	b.	(kwami)φmax (dʑi)φmax	*!*		*				*	*
	c.	(kwamɪ dʑi)φmax		*		*!
	d.	(kwami dʑi)φmax		*			*		*	*

However, the Binarity–φ constraint would also require binary prosodic constituents in the case of longer sentences, for instance in the case of a more complex subject NP. Consider for instance the sentence in (10c) containing a complex subject NP, consisting of a noun and following modifier, and a verb. Tableau (30) shows that in any of the candidates, there is inevitably a violation of Binarity–φ since three prosodic words cannot be distributed evenly across two φ-phrases. Three candidates (30c, d, e) are not isomorphic to syntactic structure and thus violate the high-ranked faithfulness Match constraint. The remaining two isomorphic candidates obey the Match constraint. (30a) shows the correct cross-word harmony, but fatally violates the CrispEdge constraint since in the case of isomorphic phrasing, cross-word harmony crosses a maximal φ-phrase. The winning candidate (30b) shows no cross-word harmony, being faithful to the input. Nevertheless, the winning candidate (30b) is not the attested candidate, but rather (30e). Hence, Binarity does not solve the analytical dilemma of cross-word harmony between a verb and a subject NP as we might conclude from (29).

30.
		[NP kwamɪ kɔɔ] [VP dʑi]   Kwame fair  eat   ‘Fair Kwame eats.’ (see (10c))	Binarity–φ	Match(LexP, φ)	CrispEdge[+ATR]φmax	*[−ATR])ω ω([+ATR]	CrispEdge[+ATR]ω	Match(LexWd, ω)	Ident-OI[ATR]ωmin	*([−ATR][+ATR])ω	Ident-IO[+ATR]
	a.	(kwamɪ kɔo)φmax (dʑi)φmax	*		*!		*		*	*
	b.	(kwamɪ kɔɔ)φmax (dʑi)φmax	*			*
	c.	(kwamɪ)φmax (kɔo dʑi)φmax	*	*!	*		*		*	*
	d.	(kwamɪ)φmax (kɔɔ dʑi)φmax	*	*!		*
	e.	(kwami kɔo dʑi)φmax	*	*!	*		*		*	*

For the neighbouring dialect of Anum, Boso, Obiri-Yeboah and Rose (2022) acknowledge the special role of subject NPs as well. Their account of prosodic phrasing is based on a combination of different rhythmic constraints that regulate the size of prosodic constituents, like Binarity, BinMin(φ,ω)-Edge, and Tern-Max. In particular, the positional Binarity constraint referring to the edge of an intonation phrase (BinMin(φ,ω)-Edge) could be of interest for the present data as it requires a φ-phrase to consist of minimally two ω’s at either edge of an intonation phrase. Tern-Max restricts the total number of ω’s in a φ-phrase to three. Applying the positional Binarity constraint to the Anum data instead of the general Binarity constraint (28) would solve the analytical dilemma of (30). If ranked highest, candidates (30a-d) would all violate the positional Binarity constraint because either the φ-phrase to the left edge of the intonation phrase (30c, d) or the φ-phrase to the right edge of the intonation phrase (30a, b) consists of one ω, failing to fulfil the Binarity requirement. The winning candidate would integrate all three words into one φ-phrase (30e), and cross-word harmony could apply, as is indeed the case for candidate (30e). However, the positional Binarity constraint runs into trouble if we go back to Tableau (26), which illustrated the blocking behaviour of a maximal φ-phrase boundary in the case of a sentence-final time adverbial. Consider the data in (13) and the analysis in (26), where the time adverbial is one lexical word that is matched to one corresponding prosodic word. BinMin(φ,ω)-Edge requires a sentence-final binary φ-phrase with the result that the time adverbial and the previous object would be phrased together. If this were the case, cross-word harmony between the time adverbial and the preceding object noun would be predicted, which is however not attested. Thus, while the positional Binarity constraint would solve the analytical conundrum at the left edge of the ι-phrase, it would cause a different dilemma at the right edge of the ι-phrase. Further, note that in Boso, cross-word harmony is blocked between a subject and a verb in certain constellations. For Anum, this subject-verb boundary blocking is not attested, which would speak against an adoption of the analysis of Obiri-Yeboah and Rose (2022) for Anum. In Boso, they found that unary φ-phrases might occur within an intonation phrase, yet not at the edges of an ι-phrase. This is a crucial difference from the data presented here, since in Anum a unary maximal φ-phrase is allowed to occur at the right edge of an ι-phrase, e.g., in case of a time adverbial.

In accounting for the positional peculiarities of the Anum subject NP, a possible Binarity constraint referring to the next higher level of the prosodic hierarchy, i.e., the intonation phrase (ι-phrase), would need to be restricted to the left edge of the ι-phrase. Hence the BinMin(φ,ω)-Edge constraint of Obiri-Yeboah and Rose (2022) would need to be reformulated as in (31). While the targeting of a particular edge of a prosodic constituent is a conceivable extension of prosodic well-formedness constraints, this particular constraint (31) is, to my knowledge, not attested for any other language. Hence, its formulation is specifically tailored to the present data and needs more justification.

31.	BinMin(φ,ω)-LeftEdge (ι)
	A φ-phrase at the left edge of an ι-phrase minimally consists of two ω’s. Assign a violation mark for each φ-phrase at the left edge of an ɪ-phrase that contains only one ω.

Another solution to the analytical dilemma discussed above might be an examination of the nature of the syntactic input. The standard assumption in syntax concerning a subject NP is that it represents a syntactic phrase moved out of the VP (e.g., Adger 2003). Given such a syntactic structure, a subject NP is not lexically headed by any further phrase (32a) and would thus constitute a case of a maximal φ-phrase when applying the Match constraint in (15b). If we were to assume a different syntactic input, an alternative would be to assume that a subject NP is part of the VP (32b), presumably the specifier of VP after the application of Merge to satisfy the selection of an agent θ-role (see Adger 2003). In such a syntactic constellation, the subject NP and the VP would constitute syntactic sisters, each of which would be matched to a corresponding φ-phrase. As a result, no maximal φ-phrase boundary would occur between the subject and the verb, allowing for cross-word harmony.

32.	a.	Simplified syntactic structure of a subject outside VP
		[CP [NP subject] [VP verb]]
		and its corresponding prosodic structure after applying Match(LexP, φ) (15b)
		( (subject)φ-max (verb)φ-max )ι
	b.	Simplified syntactic structure of a subject within a VP
		[CP [VP [NP subject] [VP verb]]]
		and its corresponding prosodic structure after applying Match(LexP, φ) (15b)
		( ( (subject)φ-non-max (verb)φ-non-max )φ-max )ι

The crucial question for an account like (32b) is whether we can find independent arguments in favour of such a syntactic analysis. Two facts may at least point towards this direction. Of course, future research on the Anum syntactic constituent structure is required to show whether a subject-within-VP assumption turns out to be motivated.

First, a number of Bantu languages show a phrasing pattern in which subject NPs are integrated into prosodic phrases that also contain the verb (e.g., Southern and Northern Sotho, Khoali 1991; Zerbian 2007). Comparing the prosodic phrasing of Northern Sotho with that of other Bantu languages, Zerbian (2007, p. 259, Table 1) lists some further languages that integrate a subject NP into the prosodic phrase with the verb. The particular cues indicating phrase boundaries in Northern Sotho are High Tone Spread and penultimate phrase lengthening. The lack of these cues between a subject NP and a verb then indicates that there is no φ-phrase break between the two constituents. The fact that the occurrence of RVH between a verb and a subject in Anum indicates no strong or maximal φ-phrase boundary is thus not simply an exceptional phrasing pattern.

Similarly, not only Bantu languages but a language like European Portuguese seems to pattern with Anum (Elordieta et al. 2005). In contrast to Spanish, Elordieta et al. (2005) argue that SVO sentences in European Portuguese are usually phrased in a single ι-phrase, the subject NP being part of the verbal domain, whereas in Spanish, a subject constituent regularly is phrased in a separate phrase. They argue that the prosodic difference is a result of different syntactic positions of a subject constituent in European Portuguese and Spanish: In European Portuguese, basically, the subject and the verb are syntactic sisters appearing in the same syntactic domain, which applying Syntax-Phonology Match constraints results in a single prosodic phrase. This representation for European Portuguese would be similar to the simplified syntactic assumption for Anum stated in (32b). Presumably, an SVO sentence in European Portuguese would constitute a maximal φ-phrase with the individual constituents being phrased into non-maximal φ-phrases. The main point of the European Portuguese facts for our discussion here is that a distinct syntactic structure for a subject NP is assumed that results in a distinct prosodic phrasing grouping the subject and verb together in one prosodic constituent.

Second, there is an asymmetry in the realisation of resumptive pronouns in Anum. In the case of subject focus, an obligatory resumptive pronoun appears as a clitic to the verb of the matrix clause, and the focused subject appears in a sentence-initial focus phrase that ends in a focus marker ‘ne’ (33).

33.	a.	Who is standing on the table?
	b.	alebi	ne	a.yeli	ɔkplɔ	a	so
		child	foc	pro.stand	table	det	on
		‘It is child that stands on the table.’

Similarly, an all-new context requires a resumptive pronoun as a clitic to the verb of the matrix clause. The resumptive pronoun indicates that the subject appears in a separate, sentence-initial phrase (34). This sentence-initial phrase may optionally end in a topic marker ‘leε’, rendering the subject the topic of the clause. Both with (34b) and without the topic marker (34c), an obligatory resumptive pronoun referring to the subject appears as a clitic to the verb.

34.	a.	What is going on?
	b.	alebi	a	a.yeli	ɔkplɔ	a	so
		child	det	pro.stand	table	det	on
		‘The child stands on the table.’
	c.	alebi	a	leε	a.yeli	ɔkplɔ	a	so
		child	det	top	pro.stand	table	det	on
		‘The child stands on the table.’

Now, consider a case of VP-focus (35). The subject is in the background of the sentence. It is neither topicalized nor focused. Syntactically, no resumptive pronoun is required. The lack of the resumptive pronoun in this case clearly indicates that the subject has not moved into a particular position higher up in the tree. Of course, this empirical fact does not directly show that the subject stays as a specifier of VP. Yet the asymmetry in resumptive pronouns can at least give an indication that the subject constituent has not moved. Further syntactic evidence needs to be shown in future research to substantiate this claim.

35.	a.	What did the child do?
	b.	alebi	yeli	ɔkplɔ	a	so
		child	stand	table	det	on
		‘Child stands on the table.’

If we assume a representation in which the subject NP is the specifier of the VP in the syntactic structure, the syntax-phonology Match constraint in (15b) requires that the subject constituent be matched to a non-maximal φ-phrase (32b). Note that other syntactic structures might be possible, specifically similar to European Portuguese for instance. The exact syntactic structure is however irrelevant for the point of discussion here. Central to the argument, the assumption is that subject NP and VP are syntactic sisters. Note that under this assumption, the different syntactic analysis is unproblematic for the cases discussed before.

Given the distinct input structure of (32b) compared to (32a), the analysis proposed for Anum above works successfully. The winning candidate (36c) obeys both the faithfulness Match constraint and the sequential prosodic markedness constraint *[−ATR])ω ω([+ATR] triggering cross-word harmony. Since there is no φ_max between the subject and the verb, an analysis along the lines of a markedness constraint requiring RVH to occur between words suffices to account for the optimal output form. The word-level Ident constraint is violated, which is similar to other regularly occurring cases of RVH shown above.

36.
		[VP [NP kwamɪ kɔɔ] [VP dʑi]]	Match(LexP, φ)	CrispEdge[+ATR]φmax	*[−ATR])ω ω([+ATR]	CrispEdge[+ATR]ω	Match(LexWd, ω)	Ident-OI[ATR]ωmin	*([−ATR][+ATR])ω	Ident-IO[+ATR]
	a.	(kwamɪ kɔo)φmax (dʑi)φmax	*!	*				*	*	*
	b.	( (kwamɪ kɔɔ)φ (dʑi)φ )φmax			*!
	c.	( (kwami kɔo)φ (dʑi)φ )φmax				*		*	*	*

5. Discussion and Conclusions

The proposal argued for in this paper is that RVH in Anum can be analysed as a phrase-level phenomenon: [+ATR] vowel harmony applies regressively between words as shown above. In some contexts, however, the process of RVH is blocked, and this is where prosodic phrasing plays a crucial role in the analysis. As argued for above, a maximal phonological phrase boundary (φ_max) blocks RVH. The analysis crucially takes in account recursive prosodic constituency, assuming maximal and non-maximal φ-phrases. If a syntax-phonology match is applied straightforwardly, prosodic phrasing is isomorphic to syntactic structure. Prosodically, then, there is a difference between a non-maximal φ-phrase that is dominated by a further φ-phrase and a maximal φ-phrase that is not dominated by a further φ-phrase. A φ_max arises if the corresponding syntactic phrase is not headed by any further lexical projection.

The analysis proposed here deviates from previous analyses of phrasal vowel harmony. Analyses like Kimber (2011) or Obiri-Yeboah and Rose (2022) assume different OT frameworks for their analyses: harmonic span theory for Vata (Kimber 2011) and co-phonologies for Gua (Obiri-Yeboah and Rose 2022). Their way of restricting phrasal vowel harmony as a bounded process targeting only the adjacent vowel of the preceding word involves constraints that specifically allow or require that vowels at a word juncture share a feature. The analysis of Anum proposed here is more like the one proposed for Akan (Kügler 2015). The syntax-phonology matching of words, phrases, and clauses is the basis for the formation of prosodic constituent structure, and the bounded RVH process is accounted for by applying a CrispEdge constraint that restricts feature sharing at a prosodic constituent boundary rather than allowing it.

The phenomenon of RVH itself is quite similar between Anum and many further Kwa languages: the process is regressive, it is only [+ATR] which triggers this kind of cross-word harmony, and the harmony targets only one immediately adjacent vowel. The harmony process is thus bounded. Other Kwa languages seem to differ in the properties of their RVH. For instance, in Nkami and in Nawuri the harmony process may span over a larger domain than just the immediately adjacent vowel (Akanlig-Pare and Asante 2016; Casali 2002), or in Vata, RVH can be iterative (Kimber 2011). Like Anum, Nkami is a Guang language and thus belongs to the same branch of the Kwa languages. According to Akanlig-Pare and Asante (2016), in Nkami RVH affects all vowels of the preceding word that occurs before a triggering [+ATR] word (37). Hence, the harmonizing process can be considered to be unbounded.

37.	ɔkʊ	yɪrɪ	obu	amʊ	sʊ	→	[yiri obu]
	someone	stand	house	DET	on
	‘Someone is standing on the house.’						(Akanlig-Pare and Asante 2016, p. 36)

Despite the similarities between Anum and other Kwa languages, there is one crucial difference: While a subject NP in a language like Akan forms a separate φ_max and RVH is blocked between a subject NP and a verb, a subject NP in Anum phrases together with the verb and the VP. The occurrence of RVH between a subject constituent and a verb speaks in favour of the fact that prosodically, both constituents appear in one φ-phrase. An analysis in terms of constraints on the size of prosodic phrases like Binarity needs a tailored constraint that requires Binarity of φ-phrases at the left edge of an intonation phrase. Alternatively, if a subject NP is represented as the specifier of a VP, under such an assumption a faithful syntax-phonology match yields an isomorphic structure, and RVH is predicted to occur within such a φ-phrase. Depending on the results of future research into phrase-level harmony, either of these possible analyses may turn out to be motivated. Nevertheless, the fact that different OT approaches to phrasal vowel harmony exist (Kimber 2011; Kügler 2015; Obiri-Yeboah and Rose 2022) shows that further research into this topic is needed to gain a better understanding of its modelling in terms of a comprehensive single OT account.