Revisiting acoustic correlates of pharyngealization in Jordanian and Moroccan Arabic: Implications for formal representations

This exploratory study of Jordanian and Moroccan Arabic (JA and MA) aims to evaluate whether pharyngealization is associated with an epilaryngeal constriction which causes ‘retraction’ and tense voice quality in surrounding vowels, following the Laryngeal Articulator Model (LAM) (Esling, 2005). Twenty male speakers (10 per dialect) produced vowels preceded by /d or dʕ/. Thirteen acoustic correlates obtained at the onset and midpoint were used to assess this type of constriction. A predictive modeling approach was used; starting with Bayesian Generalized Linear Mixed Effects modeling followed by Conditional Random Forest for classification. Vowels in the pharyngealized context were more open (higher F1, Z1-Z0), more back (lower F2, higher Z3-Z2), more compact (lower Z2-Z1), and showed spectral divergence (higher Z3-Z2). Voice quality results showed these vowels to be produced with a tense voice. High classification rates of 93.5% for JA and 91.1% for MA were obtained and variable importance score showed formant-based measures outperform voice quality ones. This suggests pharyngealization has ‘retraction,’ with a back and down gesture, as a primary correlate followed by [+constricted glottis]. The implications of these results provide strong support for LAM, the feature [+cet], and the use of the epilarynx to describe pharyngealization.


Introduction
Pharyngealization (or emphasis) in Arabic is generally assumed to involve retraction of the tongue dorsum towards the upper pharyngeal area, which leads to a lowering of the second formant in the surrounding vowels (e.g., Bin-Muqbil, 2006;Ghazeli, 1977;Watson, 2007;Zawaydeh, 1999;Zawaydeh & de Jong, 2011). Although these characteristics are mostly agreed upon, pharyngealization is also associated with a retracted epiglottis, a raised larynx, a pressed/tense voice quality, and/or a protruded lip posture (see e.g., Cantineau, 1960;Hess, 1998;Laufer & Baer, 1988;Lehn, 1963;Zeroual & Clements, 2015;Zeroual et al., 2011, among others). Although located near the constriction observed for 'true' pharyngeals, authors claimed that the two, i.e., 'true' pharyngeals and pharyngealization, share the same place but vary in degree of constriction (e.g., Laufer & Baer, 1988). Hence, and following the Laryngeal Articulator Model (Esling, 2005), both will share an epilaryngeal constriction that may be exhibited differently. An epilaryngeal constriction causes the tongue root and body to be pulled back and down in a one combined gesture and causes the vowels to be produced with the 'retracted' quality (Esling, 2005;Moisik, 2013a;Sylak-Glassman, 2014a, b). A secondary consequence of an epilaryngeal constriction causes a change in the voice source with an increase in harmonic amplitude especially in high frequencies (Halle & Stevens, 1969;Laver, 1980;Moisik, 2013b;Stevens, 1977Stevens, , 1998Story, 2016). This increase is generally associated with a tense/creaky/harsh voice quality (Edmondson & Esling, 2006;Kuang & Keating, 2012, 2014Moisik, 2013b). Our aim in this study is to provide a complete analysis of the acoustics of Arabic pharyngealized plosives in order to evaluate the presence of acoustic evidence for epilaryngeal constriction. In addition to the traditional use of absolute formant frequencies, we will investigate Bark-difference metrics as an alternative correlate to the articulatory 'retracted' vowel quality that is associated with an epilaryngeal constriction (Esling, 2005). Acoustic correlates of voice quality via harmonic differences will also be investigated to assess the presence of a tense voice quality as a consequence of an epilaryngeal constriction.
The paper is organized as follows. The sections following this introduction provide an overview of the literature on the phonetics of pharyngealization, with special attention on the consequences of epilaryngeal constrictions, before outlining the goals of the current study. Section 2 presents the method used in this study including speakers, dialects, and the corpus from which they were drawn, as well as the acoustic analyses and statistical design. Section 3 presents the results, starting with the most typically used acoustic correlates-absolute formant frequencies-before turning to the Bark-difference metrics, and acoustic correlates of voice quality. Finally, this section presents an exploratory classification technique (random forests) to evaluate primary and secondary correlates in both dialects. The last section ends with a discussion of the results and their implications for the current descriptions of pharyngealization. It is hoped that these accounts will open the door to further research into the role of the laryngeal activity that is associated with pharyngealization and pharyngeals in general.

Correlates of pharyngealization
Pharyngealization is a secondary articulation that involves a constriction located in the pharyngeal area that causes retraction of the body and root of the tongue towards the pharyngeal wall (Laver, 1994). Catford (1977, p. 163) defines this type of articulation as linguo-pharyngeal whereby the root of the tongue, including the epiglottis, moves backwards to narrow the pharynx in the front-back dimension. In many, if not all, languages, a narrowing of the pharyngeal passage near the tip of the epiglottis, and a raised larynx are the direct consequences of this constriction (see e.g., Catford, 1977, p. 193;Ladefoged & Maddieson, 1996, p. 307). Although this secondary articulation is pertinent to vowels (Ladefoged & Maddieson, 1996, p. 306), in Arabic and many Semitic languages, however, it is generally associated with consonants directly influencing the surrounding vowels (Laver, 1994).
Given the proximity between F1 and F2 observed in 'true' pharyngeals and pharyngealization, it was used as an acoustic correlate of the auditory impression of 'darkness' or 'heaviness' reported in the literature, with pharyngeals obtaining Z2-Z1 (or F2-F1 in Bark) distance below the 3.5 Z threshold for formant merging, whereas pharyngealized consonants were slightly above this threshold (close to 4.5 Z) (e.g., . This proximity between formants was reported in perceptual studies based on synthetic stimuli as a major correlate to the separation between pharyngealized and non-pharyngealized consonants in Arabic, either at the midpoint in Jordanian and Moroccan Arabic (Al-Tamimi, J., 2002; Al-Tamimi, J. & ; Barkat-Defradas et al., 2003) or at the onset in Moroccan Arabic (Yeou, 2001). We will be using this (and other) formant proximity measures as correlates of pharyngealization.
Although not specifically focusing on pharyngealization, work by Alwan (1986Alwan ( , 1989) investigated the acoustic and perceptual correlates responsible for distinguishing between uvular and pharyngeal consonants. At the onset of the vowel, pharyngeals were associated with a higher F1 and lower F2 frequencies, an increased bandwidth of the F2 (through an estimated A2-H1 value), and a smaller distance between F2-F1, whereas uvulars had a relatively stable F1 and slight raising of F2, a larger distance between F2-F1 and an increased bandwidth of F1 (through an estimated A1-H1 value). These same results were confirmed in perception in an /a:/ context, although the F2 onset and the widened F1 bandwidth were not found to contribute as much to the distinction. Bandwidth differences as estimated through harmonic differences seem to be an important correlate to the separation between pharyngeals and uvulars and thus we will be using it as a potential correlate for pharyngealization.

Epiglotto-laryngeal articulation
Although it is assumed that pharyngealization is accompanied by retraction of the tongue body, a few studies reported a constriction much lower in the vowel tract. Using various articulatory techniques, researchers have shown pharyngealized consonants to be produced with the epiglottis forming a constriction with the pharyngeal wall in the vicinity of the middle or lower pharynx that is also accompanied by tongue root/epiglottis retraction and larynx raising (e.g., Laufer & Baer, 1988;Zeroual & Clements, 2015;Zeroual et al., 2011). It is not clear then whether these differences are caused by dialectal and/or individual differences, or whether all pharyngealized consonants in Arabic have tongue root retraction and larynx raising that leads to a tense/pressed voice quality. In her comparative study of tongue root activity in various languages including Arabic using factor analyses of Ghazeli's (1977) x-ray data, Hess (1998) showed that the pharyngealized set in Arabic is better characterized by an upper pharyngeal constriction. However, her results showed that the pharyngealized set (both consonants and surrounding vowels) is also accompanied by tongue root retraction (through the Laryngopharyngeal Constriction Factor), partial constriction at the pocket of the epiglottis (through the Radical Constriction Factor), raising of the larynx, and widening of the pharyngeal wall (through the Pharynx Shifting Factor) (Hess, 1998, p. 233). In fact, the combined effect of pharyngealization was discussed by Lehn (1963, pg. 29-31), who explained that in addition to the "slight retraction, lateral spreading, and concavity of the tongue and raising of its back (what has been called velarization)," emphatic consonants are associated with "faucal and pharyngeal constriction (pharyngealization)," "slight lip protrusion or rounding (labialization)," and/or "increased tension of the entire oral and pharyngeal musculature resulting in the emphatics being noticeably more fortis than the plain segments." More recently, the same patterns of tongue root/epiglottis retraction and larynx raising of pharyngealized consonants were observed by Al-Tamimi, F. and Heselwood (2011) on multiple speakers of Jordanian Arabic using nasoendoscopic, videofluoroscopic, and acoustic data and by Zeroual et al. (2011) and Zeroual and Clements (2015) on 2-3 speakers of Moroccan Arabic using nasoendoscopic, ultrasound, EMA, and acoustic data. However, Hassan and Esling (2011) report observing differences between pharyngeal and pharyngealized consonants while examining emphasis spread in the productions of 1 Iraqi Arabic speaker using fibreoptic nasal laryngoscope. According to the authors, 'true' pharyngeal consonants showed aryepiglottic sphinctering, tongue root retraction, and larynx raising as a direct consequence of a laryngeal constriction, while pharyngealized consonants only showed tongue body retraction and raising accompanied by larynx lowering. It is possible then that some dialectal and/or idiosyncratic differences are the reason for these discrepancies. Another possibility is related to the fact that the pharyngealized set is produced with retraction of the tongue dorsum and/or root, whereas pharyngeal sets have a retraction of the tongue root as an enhancement of the epilaryngeal constriction and thus both sets are different in their 'pharyngeal' component (see e.g., Esling, 2005;Moisik, 2013a;Sylak-Glassman, 2014b). A slight larynx raising and/or constriction in the pharyngealized vowels and consonants in Arabic compared to pharyngeals may still be observed, and indicates a difference between the two categories (Hess, 1998). The exact degree of larynx raising and/or constriction in pharyngeal and pharyngealized sets is an important point to investigate further both articulatorily and acoustically in order to shed light to the exact nature of the production of these consonants. Table 1 provides a summary of the main articulatory and acoustics correlates as observed in the literature, with the additional acoustic correlates as investigated in this study. It aims at providing a correlation between articulatory and acoustic correlates in order to highlight the missing acoustic correlates of the laryngeal activity.
The various accounts of the effects of pharyngealization on surrounding vowels in Arabic show that its articulatory correlates form a complex picture with retraction of part(s) of the tongue (dorsum and/or root), retraction of the epiglottis, raising and/or constriction of the larynx leading to a more tense or pressed voice quality, and lip rounding/protrusion. However, the various acoustic accounts of the effects of pharyngealization seem to be restricted to the analyses of the first three formants of the vowels surrounding these consonants with variable outcomes as to their effects. To the best of our knowledge, there do not seem to be any accounts of the acoustic correlates of the tense or pressed articulation, and/or raised larynx, which are usually measured in terms of articulation (although see our own research Al-Tamimi, J., 2014. The next section introduces more in-depth account of the consequences of an epilaryngeal constriction following the Laryngeal Articulator Model.

Epilaryngeal constriction
Both classic and more recent articulatory reports of pharyngealization and 'true' pharyngeals sharing the same constriction location lead to the conclusion that articulatorily speaking, pharyngealization in Arabic is produced by constricting the epilaryngeal tube. This, in turn, has a direct consequence on retracting the tongue body due to tongue root retraction, with a concomitant larynx raising and/or constriction. This is the view developed in the Laryngeal Articulator Model on the type of constriction seen in the epilarynx (Esling, 2005). This model was extensively modeled by Moisik (2013a) and typologically and formally evaluated in Sylak-Glassman (2014a). According to Moisik (2013a, p. 84), the lower vocal tract "is bounded inferiorly by the glottis and superiorly by the oropharyngeal isthmus and velo-pharyngeal port." The epilarynx is located within the lower vocal tract above the larynx and has the ventricular folds as its lower part and the rim formed by the epiglottis and aryepiglottic folds as its upper part (Moisik et al., 2012). Pharyngeals in general (and potentially pharyngealized consonants in particular) are produced by sphincterally constricting the epilarynx through constricting the intrinsic or the extrinsic laryngeal muscles (Esling, 2005;Moisik, 2013a;Sylak-Glassman, 2014a, b); tongue retraction is caused by this constriction and is seen as a facilitator and an enhancer of the pharyngeal articulation (Esling, 2005;Moisik, 2013a;Sylak-Glassman, 2014a, b). Constriction of the hyoglossus muscle draws the tongue as a whole backward and downward and leads to retraction of the tongue root and dorsum (Moisik, 2013a, p. 372-373;Sylak-Glassman, 2014b, p. 5). Vowels produced with this configuration are 'retracted'; they are produced by a back and down gesture, which partially matches the consequences we see with pharyngealization.
The constriction as seen in the laryngeal area, and particularly in the epilarynx, has direct consequences on the quality of the sounds produced; ventricular folds couple with the vocal folds and are brought down to the glottis to allow for creaky phonation to occur (Moisik, 2013a;Sylak-Glassman, 2014a, b) and when constricted they are associated with laryngealization, tense and harsh voice quality (Edmondson & Esling, 2006;Stevens, 1977). Aryepiglottic fold constriction is associated with an enhanced and clearer voice quality especially in singing due to an increased energy in the higher frequencies as well as 'ringing' seen in specific types of singing (Story, 2016;Titze & Story, 1997). Aryepiglottic fold constriction is observed in the production of 'true' pharyngeal consonants (Esling, 2005) and a few studies reported 'trilling' as a direct consequence of 'true' pharyngeal articulation (Esling, 1999;. The difference between the two types of constrictions may be seen as a direct correlate of the difference between 'true' pharyngeals and pharyngealization; the former is produced by constricting the aryepiglottic folds as a primary feature whereas the latter by constricting the ventricular folds as a secondary feature. Pharyngeals were described as having tongue root retraction and lowering, epiglottal retraction, in addition to larynx raising as one component (Esling, 2005;Hassan & Esling, 2011;Moisik, 2013a)-descriptions that seem to match the few reports of pharyngealization in Arabic (see Section 1.1.2). This seems to suggest that the whole guttural class, including pharyngeal, pharyngealized, and uvular consonants, is produced with an epilaryngeal constriction, albeit to varying degrees of stricture and phonation (Moisik, 2013a;Sylak-Glassman, 2014a, b).

Acoustic consequences
From an acoustic point of view, and given that the vowels in the vicinity of pharyngealized consonants are better described as 'retracted' (Sylak-Glassman, 2014a, b), we will advocate the use of the proximity between formants as an (psycho-)acoustic correlate. Bark-difference formant frequencies are a well-established method for vowel recognition, perception, and normalization (e.g., Syrdal & Gopal, 1986;Thomas & Kendall, 2007, inter alia). The proximity between Z1-Z0 (that is F1-f0 in Bark) correlates well with the openness dimension (i.e., [±high]) (Fahey et al., 1996;Hoemeke & Diehl, 1994;Syrdal & Gopal, 1986;Traunmüller, 1981), with more close ([+high]) vowels showing a Z value lower than 3 Bark, and more open ([-high]) vowels around 5 Bark (Syrdal & Gopal, 1986, p. 1090). Z2-Z1 correlates well with compactness of the specturm (Sylak, 2011;Syrdal & Gopal, 1986) and is highest for front vowels, and lowest for (mid-)open back vowels (Syrdal & Gopal, 1986, p. 1090. And finally Z3-Z2 shows the highest difference for (mid-) open back and back vowels and the smallest for front vowels and correlates well with the backness/retraction of vowels (Syrdal & Gopal, 1986, p. 1090). Z3-Z2 correlates well with 'spectral flatness,' or divergence of spectral peaks of F2 and F3 as seen in pre-palatal /i/ to distinguish it from /y/ (Wood, 1986). Thus we also hypothesize that a large Z3-Z2 difference between pharyngealized and non-pharyngealized contexts is a by-product of formant merging observed in Z2-Z1 due to the pharyngeal constriction. Hence we expect the vowels in the vicinity of pharyngealized consonants to show a higher Z1-Z0, a lower Z2-Z1, and a higher Z3-Z2. Voice source changes occur with an epilaryngeal constriction. Constricting the ventricular folds leads to a tense, pressed, or laryngealized voice quality with an overall lowered or flatter spectral tilt, due to the increased energy above the first harmonic (Halle & Stevens, 1969;Hanson et al., 2001;Klatt & Klatt, 1990;Laver, 1980Laver, , 1994Moisik, 2013b;Stevens, 1977Stevens, , 1998Sundberg & Askenfelt, 1981). When an epilaryngeal constriction is caused by an aryepiglottic fold constriction an 'enhanced' voice quality, especially in singing, can be seen Story, 2016;Titze, 2008;Titze & Story, 1997). Samlan and Kreiman (2014) described the acoustic and perceptual consequences of constricting the epilarynx at either the aryepiglottic or the ventricular folds. Their results suggest no major differences in the increase in energy in the high frequencies, but perceptually the voice qualities were different in both types of constrictions. We will be investigating spectral slope measures that are widely used as an acoustic evaluation of phonation and voice qualities which have been successfully used to distinguish non-modal from modal phonation (e.g., Garellek, 2012;Hanson et al., 2001;Keating et al., 2015;Klatt & Klatt, 1990;Kuang & Keating, 2012;Ladefoged & Maddieson, 1996, among others).
The various metrics used refer to the differences between the harmonics closest to f0, 2*f0, F1, F2, and F3 and are expressed as H1, H2, A1, A2, and A3 respectively. Garellek (2012) and Kuang and Keating (2012) provide a comprehensive summary of the various acoustic correlates of phonation and spectral slope. Given that the tense/pressed voice quality was reported in the context of pharyngealization, we will restrict the explanations below to the creaky/tense/pressed voice quality correlates as compared to breathy voice quality. A tense/pressed voice tends to show a lower H1-H2 as the main acoustic correlate (e.g., Garellek, 2012;Hanson et al., 2001;Keating et al., 2015;Klatt & Klatt, 1990;Kuang & Keating, 2012;Ladefoged & Maddieson, 1996, among others). Spectral tilt measures, i.e., H1-A1, H1-A2, and H1-A3 seem to be directly correlated with the abruptness of vocal fold closure (Garellek, 2012;Hanson et al., 2001). H1-A1 correlates well with the bandwidth of F1; as the bandwidth of F1 increases, the amplitude of the harmonic closest to F1 decreases and thus a lower H1-A1 indicates a creaky/tense voice (Garellek, 2012;Hanson et al., 2001;Kuang & Keating, 2012). H1-A2 and H1-A3 are also correlated with creaky/tense phonation, showing lower values (Garellek, 2012;Klatt & Klatt, 1990;Kuang & Keating, 2012). Hanson & Chuang (1999) and Hanson et al. (2001) described how an abrupt closure of the glottis can yield a change in the source spectrum; a decrease in spectral tilt around F3 is observed, and hence a lowered H1-A3 would be expected. If the glottis is closed off simultaneously from the front to the back along the length of the vocal folds, then an increase in spectral tilt would be seen and a higher H1-A3 is obtained. This should not be seen as an indication of an increased noise, as an increased noise as seen in breathy voice and/or glottal opening will increase spectral tilt further and H1-A3 will be much higher (for more detail, see e.g., Hanson & Chuang, 1999;Hanson et al., 2001;Klatt & Klatt, 1990;Stevens, 1998). Moving to the higher frequencies, the amplitude difference A1-A2 is directly related to phonation types, and thus a creaky/tense voice tends to have lower A1-A2 (Aralova et al., 2011;Fulop et al., 1998;Guion et al., 2004;Kang & Ko, 2012). And finally, constricting the epilarynx on its own or through retraction of the tongue root leads to an increase in the energy of the harmonics above F1. This increase can either be extensive in that the energy in A2 and A3 are higher than that of A1; consequences of an extreme epilaryngeal constriction that leads to an 'enhanced' voice (Story, 2016). In this case, a lowered or even negative A1-A3 and A2-A3 is obtained due to the concentration of energy around F3, F4, and F5 as seen in the 'singer's' formant (Story, 2016). However, when the epilaryngeal constriction is minimal and/or is associated with a primary pharyngeal constriction, we would expect a change in the pattens. This would cause an increase in the amplitude differences of A1-A3 and A2-A3, due to the decrease in energy around F3. This again correlates well with the acoustic consequences of an abrupt closure of the vocal folds (Hanson et al., 2001).

Formal representation
As we saw above, an epilaryngeal constriction yields both a 'lingual' and a 'laryngeal' effect. This combined effect of an epilaryngeal constriction is seen as a fundamental shift from considering lingual or laryngeal constrictions separately and can be used to describe the post-velars as one set of combined articulations (for more detail on the glottocentric vs linguocentric view, see Moisik, 2013a;Moisik et al., 2012;Moisik & Esling, 2011;Sylak-Glassman, 2014a, b). This allowed for the introduction of a new set of laryngeal features to account for this combined effect; [±cet] (for [±Constricted Epilaryngeal Tube]) (Moisik et al., 2012;Moisik & Esling, 2011) and, as an extension, [±ce] (for [±Constricted Epilarynx]) (Sylak-Glassman, 2014a). Following the Lower-Vocal-Tract Phonological Potentials (Moisik, 2013a), [+cet] causes surrounding sounds to be retracted, constricted and to have raised larynx voice and tense phonation; all as a combined unit (Moisik, 2013a;Moisik et al., 2012;Moisik & Esling, 2011). 'Retracted' is due to tongue and epiglottis retraction with a back and down gesture; 'constricted' of certain intralaryngeal muscles and 'raised larynx' leading to tense voice quality (Moisik et al., 2012, p. 5). The feature [+ce] is used to distinguish between pharyngeals/epiglottal and pharyngealized/epiglottalized categories with the former being assigned this feature as a primary place feature, whereas the latter receives it as a secondary feature (Sylak-Glassman, 2014a, p. 136-138). Following Moisik et al. (2012) and Moisik and Esling (2011), 'true' pharyngeals (and other categories) are assigned this feature as an indication of the primary constriction of the epilarynx; pharyngealized consonants are assigned [+ce] to their primary specification (Sylak-Glassman, 2014a, p. 138). Whether all or parts of these combined articulatory consequences are to be used depends on the category of sounds (Moisik et al., 2012, p. 5-6), hence pharyngealized consonants may show tongue root retraction and/or intralaryngeal muscle constriction but not larynx raising, etc. (for more detail, see e.g., Moisik & Esling, 2011,

Aims of the current study
This exploratory study aims at acoustically investigating whether an epilaryngeal constriction is associated with pharyngealization in Arabic. We take the views developed in the Laryngeal Articulator Model that an epilaryngeal constriction leads to a 'retracted' vowel in the pharyngeal area with both tongue backing and lowering that causes a tense/pressed/laryngealized voice quality due to constricting the larynx. Acoustically speaking, we expect the vowels in the pharyngealized context to show a raised F1 and Z1-Z0, a lowered F2 and Z2-Z1 as a direct consequence of 'retraction' as defined in the Laryngeal Articulator Model, as well as a raised Z3-Z2 signaling both backness and spectral divergence as an enhancing correlate to the already lowered Z2-Z1. F3 can potentially show a lowered value if pharyngealization is produced in the mid-low pharynx for front vowels or a raised F3 for back vowels due to the tighter constriction. Spectral slope and voice quality correlates are expected to correlate well with the tense/pressed voice quality with an overall lowered H1-H2, H1-A1, H1-A2, H1-A3, A1-A2, and with an increased energy in the high frequencies leading to an overall flatter spectrum and a relatively lowered A1-A3 and A2-A3.

Speakers
Twenty Jordanian and Moroccan male speakers (10 of each dialect), aged 20 to 30, participated in this experiment. Jordanian Arabic speakers originated from Irbid in the north of Jordan whereas Moroccan Arabic speakers come from Mohammedia (near Casablanca). They all reported no history of articulatory or hearing disorders, and shared the same sociolinguistic background: At the time of recording, they were students at university, and all lived in the city (i.e., spoke an urban variety). Some Jordanian Arabic speakers had some knowledge of French and/or English (beginner to advanced levels), while Moroccan Arabic speakers were non-Berber, and had knowledge of French. Both groups were proficient in the standard and dialectal forms of Arabic.

Dialects
Although Jordanian and Moroccan are dialects of the same language, some important phonetic-phonological, rhythmical, and lexical differences exist between these varieties. They belong to the eastern and western zones respectively, and some researchers have shown that these two dialectal zones exhibit substantial differences (Versteegh, 2001). Vowel inventories are reduced in western dialects compared to the eastern ones (Marçais, 1977) and this has a direct effect on vowel space size which is reduced in western dialects (Al-Tamimi, J., 2007a;Al-Tamimi, J. & Ferragne, 2005;Barkat, 2000). A more complex syllable structure seems to operate in western dialects (Cohen, 1962), which has a direct effect on their rhythmic structure; they are described as having a 'faster' and 'jerkier' (or more halting) rhythmic structure than eastern dialects (Barkat, 2000;Ghazali et al., 2002Ghazali et al., , 2007. From an intonational point of view, both regions seem to show differences in placement of stress, although their patterns in producing the statements that are used here seem to follow the rise-fall pattern (Ghazali et al., 2007). Lexical differences between the two zones are important enough for the dialects to be mutually unintelligible, which leaves the question open as to whether these constitute different languages or dialects.
A key aspect of the current study is to evaluate whether the phonetic implementation of pharyngealization in these two dialects is different or not. From a cross-dialectal perspective, Bellem (2007) suggested that pharyngealization is implemented in the same manner, although some dialects can show more of a 'guttural' quality than others. It is not clear however, what is meant by a more 'guttural,' whether it is more mid-to-low pharyngeal constriction or other qualitative differences. On coarticulatory patterns, Embarki et al. (2011) showed that both dialects displayed different patterns of the effects of pharyngealization with Moroccan Arabic speakers having the mostly distinctive locus equation slopes, followed by Jordanians (and then the two other dialects, Kuwaiti and Yemeni). And in our own research, we have shown the two dialects to behave differently; steeper formant slopes of F1, F2, and F3 were obtained in JA (Al-Tamimi, J., 2007a, b). Our aim is then to provide a clearer picture on the potential differences as observed between these two dialects in how pharyngealization is implemented.

Material and recordings
The material used in this exploratory study comes from a larger corpus on bilabial, alveolar, and velar stops that was used to investigate the role of dynamic correlates (i.e., formant slopes) in production and perception (for more details, see Al-Tamimi, J., 2007a, b). The real words used in this study are listed in Table 2. Voiced alveolar pharyngealized and non-pharyngealized target consonants were embedded in ˈC 1 V 1 C, ˈC 1 V 1 CV, ˈC 1 V 1 CVC, or CVˈC 1 V 1 C syllable structures, where C 1 = /d or dˤ/ and V 1 = /iː ɪ eː aː ɐ oː ʊ uː/ in Jordanian Arabic or /iː aː ə ʊ uː/ in Moroccan Arabic (Al-Tamimi, J., 2007a). Due to the environment used, it was not possible to obtain minimal sets for the two varieties nor comparable words with /d dˤ/ followed by /oː/ in JA. As can be seen, the target consonants were sometimes either in initial or medial positions, or with other 'guttural' sounds present due to constraints on finding real words. However, it is generally agreed on that even with cross-dialectal variations, emphasis spread is greater from coronal pharyngealized consonants, i.e., 'true' emphatics compared to other 'guttural' sounds (Hellmuth, 2013;Watson, 2007). In addition, rightward emphasis spread is more common than leftward (Hellmuth, 2013), although Bellem (2007) seems to suggest small variations between dialects in how this aspect is implemented. Given these restrictions, our aim is to evaluate how these additional acoustic correlates can be used on such a corpus.
The speakers were seated in front of a computer in a sound attenuated room (for Jordanian Arabic speakers) or in a very quiet room (for Moroccan Arabic speakers), and the recordings took place via a computer assisted user interface developed specifically for this task. After a training phase, and for purposes of other experiments using this dataset, speakers were asked to produce each item as realized in the target word, in the target CV syllable and then in the target isolated vowel, while trying to keep the production of each vowel constant across realizations and having at least 0.5 sec gap between each sequence. This particular task is aimed at evaluating the role of contextual information on the degree of vowel hypo-vs hyper-articulation and in vowel perception (for more details see Al-Tamimi, J., 2007a). The words were randomly presented with five repetitions in an adapted carrier sentence using Modern Standard Arabic (MSA) script without vocalization in order not to influence the production of each speaker, and to obtain the real dialectal realization of each word. The carrier sentence had an important role here as it was used as a way to convey the meaning of the unvocalized word, and hence reduce the influence of the MSA script. In less than 2% of the time, some speakers produced a non-dialectal form; in these cases, the experimenter clicked on a button (N for no) that enabled the word to be added to the end of the set and the speaker was asked to reproduce it. In this second round of production, most speakers reproduced the words in a dialectal form. The quality of the productions was assessed by the author and in the case of Moroccan Arabic by a native speaker of the variety. Speakers were asked to produce each word, while not moving or modifying the distance to the microphone, the loudness, intonation, or their speech rate. Each sound file was then digitized directly on the same computer, with a sampling frequency of 22.05 kHz, 16-bit quantization, in mono channel using a Sony MS 907 microphone (distance 15-20 cm from the speakers' mouths). Given that the corpus used here is part of a larger study that investigated the role of dynamic specification in production and perception of vowels, the length of all the experiments (production and perception) was about 2 hours per speaker (for more details see Al-Tamimi, J., 2007a). For the current study, vowels produced in the word condition are used to evaluate the differences observed in vowel realizations as a function of the pharyngealized vs non-pharyngealized environments; the total number of words produced by the speakers for this study was 700 for Jordanian and 500 for Moroccan Arabic (henceforth JA and MA).

Data processing and acoustic analyses
Data were segmented manually using Praat (Boersma & Weenink, 2009). The onset/offset of a vowel corresponded to the first/last periodic pulse in the waveform and coincided with the vertical homogeneity of the first four formants in the 0 to 5000 Hz display of a wide-band spectrogram. In the cases where a sonorant followed the vowel, intensity drop and visual inspections were used to determine the boundary position of the vowels.

Acoustic analyses
Acoustic analyses were performed automatically using a Praat script designed by the author and adapted from Al-Tamimi, J. and Khattab (2015). Before performing the analyses, measurement frame positions of the onset and midpoint of a vowel were estimated following Al-Tamimi, J. (2004, 2007b) and Al-Tamimi, J. and Khattab (2015) in order to obtain accurate measurements and reduce errors from the automatic analyses performed by Praat. The acoustic periodicity of voiced frames in vowels was first estimated by computing f0 (see below for more details) for each sound file and by performing a PointProcess (cross-correlation) analysis. Following this, the average length of a complete glottal cycle was estimated for each speaker and each sound file; this glottal cycle ranged over 8-10 ms for our male speakers. Starting off from the initial estimates of measurement times as obtained from the TextGrids (following the segmentation as described above), these were then adjusted to match the time of maximum intensity occurring within the length of an average complete glottal cycle. Following this, they were left-aligned to the original onset estimate, and centered at the original midpoint estimate. The intensity values, computed every 5 ms, were interpolated before computing the maximum; the adjustments that resulted from this process were up to 2-3 ms around the original positions. All the reported measurements are obtained at the estimated positions. Formant frequencies: Formant frequencies (F1, F2, and F3) were measured at the vowel onset and midpoint of each vowel. These were obtained from a 25-ms Kaiser2 (Gaussianlike) window with a 5-ms time step and interpolation. A maximum of five formants was requested in the formant analysis using the default Burg algorithm for formant estimation with a maximum frequency of 5 kHz for male speakers. Following formant estimation, Praat's Formant track function was used to reduce the errors in automatic formant estimation. Formant frequencies were then verified manually to prevent potential errors obtained from automatic extraction. When formant tracks obtained through the initial analysis were weakened, as in the case of /ʊ uː/ vowels, a second analysis was run using an LPC smoothed curve obtained from a 256-point zero-padded DFT spectrum computed from a 10-ms Kaiser2 window left-aligned at the onset or centered at the midpoint of the vowel, after down-sampling the sound file to 10 kHz, low-pass filtering with an antialiasing filter with a cut-off frequency of 5 kHz, and pre-emphasizing it by a factor of 0.98. Both the FFT and the smoothed LPC displays were used to estimate the position of a particularly weakened formant.
Bark-difference formant frequencies: Once the first three formant frequencies were estimated, these were converted to the psychoacoustic Bark scale following Traunmüller's (1990) formula 1, where Z n is the Z value (i.e., critical bandwidth) of the formant n , and f n is the frequency in Hz of the formant n (including f0 in both cases), and any Z values lower than 2 Bark, were corrected using formula 2, where Z reflects the original Z value and Z c is the corrected one: The fundamental frequency f0 was also estimated and only used in Bark-difference results.
Overall, the results showed non-significant changes according to the consonant and vowel quality; there was however a pattern of lower f0 in JA and higher in MA. The estimation followed the procedure in Al-Tamimi, J. and Khattab (2015) and used the two-pass method following (Hirst, 2011). For the first pass, Praat's default settings were used: 5-ms step, 40-ms Kaiser2 window, floor and ceiling = 75-500 Hz respectively. This allowed to obtain the pitch contour for each speaker. The first and third quartiles (i.e., 25-75% values) were obtained for each speaker and these were then multiplied by a coefficient; 0.75 and 1.5 respectively. These were then used as the new floor and ceiling in the second pass and the actual f0 computations was done using a 5-ms step and an effective Gaussian window length of 30 ms. The floor and ceiling values for speakers ranged between 75-100 and 150-200 Hz respectively. The frequencies obtained in Hz were converted to the Bark scale using formulas 1 and 2. Once the Bark transformation applied on F1, F2, F3, and f0 in both positions of the vowel, the Bark-difference Z values were then computed to reflect Z1-Z0, Z2-Z1, and Z3-Z2 at the onset and at the midpoint of the vowel. Voice quality: To obtain voice quality measures at both onset and midpoint of the vowel, the sound files were low-pass filtered with an anti-aliasing filter which had a cut-off frequency of 5 kHz, down-sampled to 10 kHz, and pre-emphasized by a factor of 0.98. Intervals 40 ms long were defined to allow for at least 4 to 5 complete glottal cycles to be used in estimating spectra of vowels. From each sound file, one interval was left-aligned at the onset and a second centered at the midpoint of the vowel; these were then windowed using a Kaiser2 window function. From each windowed interval, a 256-point zero-padded DFT spectrum was computed and the logarithmic power spectral density, with a bin size of 19 Hz, was computed. Following Al-Tamimi, J. and Khattab (2015), the amplitudes of the first and second harmonics and of the first to the third formants were automatically obtained by detecting the highest peaks for a particular harmonic; maximum amplitude was obtained from f0*0.9 to f0*1.1 and from 2*f0*0.95 to 2*f0*1.05 for H1 and H2, respectively. For the amplitude of the harmonics closest to the three first formants, we estimated the bandwidth frequencies using the formula proposed by Hawks and Miller (1995) instead of using the automatically estimated ones in Praat due to many errors that prevented manual correction. Then maximum amplitudes were obtained in the region from F1 -0.5*Bandwidth1 to F1 + 0.5*Bandwidth1 for A1. The same procedure was applied for A2 and A3 (and using Bandwidths 2 and 3 respectively). The automatic detection of formant frequencies, and highest peaks were manually checked to prevent errors. Instead of using inverse-filtering to estimate these harmonics, we relied on the normalization procedure as developed by Iseli et al. (2007) and implemented in our Praat script to obtain the 'corrected' versions of these harmonics that removes the boosting effects of surrounding formants. H1, H2, A1, and A2 were normalized by correcting for the effects of F1 and F2, whereas A3 was normalized by correcting for the effects of F1, F2, and F3. Then amplitude differences were obtained for H1*-H2*, H1*-A1*, H1*-A2*, H1*-A3*, A1*-A2*, A1*-A3*, and A2*-A3* at both onset and midpoint of the vowel.

Statistical analyses
A total of 30,966 measurements (17,992 in JA and 12,974 in MA) were obtained from the 13 different acoustic correlates summarized above (Section 2.2.1), measured at both onset and midpoint (i.e., a total of 26 measures). Our main aim in this study is to evaluate the degree to which a particular acoustic correlate can be used to successfully predict the difference between the pharyngealized and non-pharyngealized consonants, hence, we adopted a predictive approach (Baguley, 2012;Hastie et al., 2009;Kuhn & Johnson, 2013). First, we started by using Generalized Linear Mixed-Effect Modeling (GLMM) (Quené & Van den Bergh, 2008) in a confirmatory way that was followed by an exploratory classification technique, namely Conditional Random Forests that allowed us to evaluate which acoustic correlate(s) are the most predictive of the two consonant categories. All analyses were run using the statistical software r (version 3.3.1) (R Core Team, 2016).

Generalized Linear Mixed-Effects Modeling (GLMM)
Before running the GLMMs, we started by examining the correlation levels in our data. We used correlation matrices ordered by hierarchical clustering obtained with the function hclust from the package reshape2 (Wickham, 2007), and generated via a modified R script (Melike, 2016). These correlation matrices are presented in Appendix 1 for JA and in Appendix 2 for MA. The correlation matrices showed that out of the total combinations of 650 correlations in each dialect, 138 turned out to be significant in JA and 172 in MA. Correlation coefficients are judged as statistically significant at p < 0.01, i.e., any absolute r 2 value >0.16 was significant (correlation and p values obtained with the function rcorr from the package Hmisc (Harrell Jr, 2016)). In both dialects, voice quality and formantbased measures are negatively correlated with each other, and all formant measures (being absolute or Bark-difference) are positively correlated with each other (except in JA, where Z3-Z2 is negatively correlated with Z2 and Z2-Z1). Given the recommendations in Baayen (2008, p. 181-183), it was not advisable to use all these measures together in a regression analysis as they will be giving either the same outcome (in the case of positively correlated ones) or cancel each other out (for negatively correlated cases). Hence we decided to use separate regression analyses on each of the individual acoustic correlates and compared these via predicted probabilities. In addition, due to the high predictive outcome of some acoustic correlates (e.g., F2, Z3-Z2, see below), it was not possible to use one GLMM model combining all the acoustic correlates, as these were canceling each other out, resulting in model non-convergence.
Due to these constraints (multicollinearity and high predictive power), we used an alternative approach. Normally distributed acoustic correlates were used (by transforming the absolute formant frequencies to the Bark scale; Bark-difference and spectral slope measures are already on logarithmic scales). We obtained the descriptive statistics using the package psycholing (Fraundorf, 2015), with the means and standard deviations for the original and z-scored values. The z-scored values were computed to a mean of 0 and a standard deviation of 1, separately for each dialect (using the means and standard deviations presented in column "All" in Table A3 in Appendix 3 that provides a summary of the combined contexts in JA and MA and for each of /d/ and /dˤ/ contexts, with the mean and standard deviations in the original and the z-scored scale. We also included the differences in each of the acoustic cues between /dˤ/ and /d/ to allow for comparability of the outcome with the results of the GLMMs).
Following this, and to avoid multicollinearity, we ran separate GLMMs on the individual z-scored acoustic correlates following Schielzeth (2010) as a simple and meaningful way to compare individual models and to assess their effect sizes. We started by running GLMMs with the 'consonant' as a binomial response category (dummy coded with /d/ = 0 and /dˤ/ = 1), and the separate acoustic correlates as predictors (i.e., a Mixed-Effect Logistic Regression). The dummy coding of consonant is a requirement of the GLMM and allows for a meaningful interpretation of the results with the β of the Intercept representing the average in the /d/ environment whereas the predictor's β is that in the /dˤ/ environment. Vowels (contrast coded) and repetitions (centered) were added as additional predictors; however, they did not improve the fit of the GLMM (through likelihood ratio comparison), thus these were dropped from the final models.
For the random part of the model, we used speakers and items as crossed random factors . Within item variation with respect to repetitions was included to allow for this random variation to be taken into account. Both by-speaker and by-item random slopes were used following a maximal specification model (Barr et al., 2013). These allowed both speakers and words to vary with respect to the within variation that is due to the acoustic correlates. By-item random slopes were necessary given that we only had one item per consonant and vowel combination; this allowed the acoustic correlates to vary within each word, otherwise the model would inaccurately overestimate this variation. In many occasions, random intercepts for words were not successful at providing a clear picture of the results, and hence were dropped, i.e., only a by-item random slope was used.
To prevent quasi-and complete separation of the data that was obtained with lme4 package (Bates et al., 2015), we used a Bayesian Generalized Linear Mixed-Effects Models as implemented in the package blme (Chung et al., 2013). The function bglmer requires defining a prior for both the fixed and the random effects; we used a multivariate normal distribution with a diagonal variance-covariance matrix for the fixed effect prior, i.e., fixef.prior = normal(cov = diag(2.5,2)), with 2 being the 2 fixed-effect parameters and 2.5 representing a variance of 2.5 which is equivalent to a 1.58 SD that is close to our actual SD for the z-scored data (being 1SD) (following the recommendations of Gelman et al., 2008Gelman et al., , p. 1366 for numerical predictors). The random effect prior was kept at its default, (i.e., cov.prior = wishart, for more detail, see Chung et al., 2013).

Random Forests via Conditional Inference Trees
Following the individual GLMMs, we wanted to use a classification technique as a predictive model (Hastie et al., 2009;Kuhn & Johnson, 2013). Random Forests are one of the most versatile machine learning algorithms as they do not require many tunings of their settings. They have been applied to sociolinguistic data (for a detailed description of these methods, see Tagliamonte & Baayen, 2012), and also to acoustic cue weighting in perception (Brown et al., 2014). Random Forests were originally proposed by Breiman (2001) as an ensemble learning algorithm that uses independent classification and regression trees in growing a forest. The independence stems from the randomness of the selection process in the algorithm: By randomly selecting a subset of the data and of the predictors, a decision tree is constructed. This is then repeated several times (for the total number of trees in a forest), and a majority of votes is then taken for prediction (Liaw & Wiener, 2002). After the forest is grown, one can estimate the prediction accuracy as well as the ranking of the most important predictors. This model can also be used to predict the outcome on new unseen data, e.g., newly collected production/perception data or on the testing set (out-of-bag set). The algorithm splits the data into learning and testing sets; the former constitute around 66.7% of the data and named as an in-bag set, while the latter constitutes the remaining data of 33.3% and named as an out-of-bag set. Then subsampling without replacement is used in growing the forest (for more details, see Strobl et al., 2009).
Instead of using the original implementation of Random Forests available in the R package randomForest (Liaw & Wiener, 2002), we used Random Forests grown from Conditional Inference Trees as implemented in the package party (Hothorn et al., 2006;Strobl et al., 2008Strobl et al., , 2007. Strobl et al. (2008Strobl et al. ( , 2007 found that the original randomForest provided biased estimates of Random Forests' variable importance (see below) as it was biased towards variables with multiple categories and multiple cut-points, and also overestimated variable importance measures when correlated data is used (as is the case in our study). To guard against this bias, they developed an unbiased selection process (subsampling without replacement). This form of random forests, i.e., based on conditional inference trees, are well suited to deal with collinear variables (Strobl et al., 2008(Strobl et al., , 2007Tagliamonte & Baayen, 2012), with "'small n large p' case, where the number of predictor variables p is greater than the number of subjects n" (Strobl et al., 2009, p 323) and they allow to evaluate the importance of the predictors used in classification by ranking them after controlling for interactions and collinearity.
Random Forests were run using the party package with the function cforest on the combined 26 different acoustic correlates for each dialect separately and by using the recommended cforest_unbiased control with mtry = 5 (rounded square root of 26 predictors). Number of trees ntree can be tuned to allow for less computation time, and we followed the density-based metrics developed by (Oshiro et al., 2012) to estimate the density of our dataset using the formula 3, with a = number of predictors, n = number of observations and c = number of classes.
For JA and MA, the density-based metrics were equal to 1.79 and 1.69 respectively. According to Oshiro et al. (2012), these values constitute a low density database. The authors found that a large number of trees close to and above 2000 is only needed with high density datasets that exceed a density-based value of 3. A low number of trees (ntree = 128 trees) was needed for the majority of their datasets to achieve a high predictive accuracy. Hence, we implemented the same procedure as that suggested by Oshiro et al. (2012). We started by building 15 different random forests for each dialect by using ntree from 100 to 1500 in a 100 trees increment. Then for all generated random forests, we checked their predictive power, by using the function predict using the out-of-bag set as a cross-validation (using OOB = TRUE). Then we used an AUC (for Area Under the Curve) based comparison using the package pROC (Robin et al., 2011), by generating an ROC curve (for Receiver Operating Characteristics) and then by performing a non-parametric Z test of significance on the correlated ROC curves using the function roc.test, following DeLong et al. (1988). The results of this comparison showed that for JA, 400 trees were enough to reach the highest predictive accuracy, whereas for MA, 300 trees were enough. Hence we ran random forests with these ntree values. Then we used an AUC-based estimation of the variable importance, varimpAUC as it takes into account both accuracy and error of estimation and used conditional permutation tests with conditional = T (for more detail, see specification in Strobl et al., 2008Strobl et al., , 2009). We then evaluated how well correlated are the random forest results with the actual data, and the percentage correct classification by using an out-of-bag cross-validation.
After this initial random forest using all of the acoustic correlates, we ran six additional exploratory random forest analyses that will be used for their predictive accuracy. These aimed at evaluating whether the predictive accuracy would be different depending on the type of acoustic correlates used. We compared formant-based metrics only, i.e., absolute + Bark-difference formants, absolute formants, or Bark-difference formants only). The aim here is to evaluate the strength of these metrics at separating the two categories. This was followed by the combination of absolute formants or Bark-difference formants in addition to voice quality, in order to assess whether there is an increase/decrease in predictive power when voice quality is used. And finally, we assessed the predictive power of the voice quality metrics on their own, in order to evaluate whether any observed statistical difference are important enough to allow for discriminating the two consonantal categories. In each case, we used the above procedure to estimate the optimal number of trees needed to obtain the highest predictive accuracy, and adapted mtry to each case. The next section presents the results of this study.

GLMM
We started by running two GLMM analyses for each acoustic correlate: One including the acoustic correlate as a predictor (i.e., a Full model) and a second without the acoustic correlate (i.e., a Null model). Then we proceeded with model comparisons using likelihood ratio tests to derive all p values (Barr et al., 2013) and to test whether a particular acoustic correlate is significantly contributing to the difference between the two consonants. Out of the 26 different acoustic correlates used in this study, 16 made a significant contribution to the difference in JA, compared to 20 in MA. These are highlighted in bold in Table 3 (as a reference, complete descriptive statistics of the results are presented in the Table A3 in Appendix 3). Figures 1, 2, 5, and 6 were generated using the predicted probabilities of the individual acoustic correlates as obtained from the Full GLMM. These predicted probabilities were obtained from the predict function in blme based on a new dataset with the range between -3 and 3 z-score with a 0.1 increment (using a modified version of an R script developed by Winter, 2016). These were then plotted the following packages: lattice (Sarkar, 2008), latticeExtra (Sarkar & Andrews, 2016) and gridExtra (Auguie, 2016). Using this range allows for a meaningful comparison between all the measures. When looking at the predicted probability curves, we will refer to a sigmoidal shape or a flat shape; the former represents the logistic function curve and the latter a linear shape. A sigmoidal curve indicates a high level of separation between the groups and hence will have a high predictive function. To interpret the effect size associated with each acoustic correlate, the absolute β value will be used to report the standardized effect size; the negative or positive signs are indicative of falling or raising probability curves from /dˤ/ to /d/. As an example, Figure 1 shows the predicted probabilities for Z1, Z2, and Z3 at both onset and midpoint of the vowel in JA (blue solid) and MA (red dashed). In both dialects, Z1 at the onset of the vowel shows raising curves from /d/ (values starting from -3 z-score) to /dˤ/ (values ending with +3 z-score); Z2 shows the reverse pattern, i.e., falling curves from /dˤ/ (values starting from -3 z-score) to /d/ (values ending with +3 z-score); and Z3 shows an overall flat to raising curves from /d/ (values starting from -3 z-score) to /dˤ/ (values ending with +3 z-score), although the shape is not as sigmoidal as that seen for Z1 nor Z2. We have also provided the percent correct classification of each GLMM based on the β value by using the formula: % = exp|β|/1 + exp|β|.

Absolute Formants
From model comparison results (see Table 3), it can be seen that F1 and F2 at the onset and midpoint are, as expected, reliably associated with the differences between pharyngealized and non-pharyngealized consonants. In MA, F3 at the midpoint only showed a tendency towards being a correlate to the difference. GLMM results confirm the outcome as observed in the literature; F1 shows an increased β values whereas F2 shows a lowering in /dˤ/. F3 showed an overall higher values in /dˤ/, albeit non-significant. The predicted probabilities as displayed in Figure 1 show a clear pattern of rise, fall, rise patterns for F1, F2, and F3; with F3 showing a slight rise rather than a flat line. Given that formant frequencies were z-scored, it is possible to compare the strength of the effect as these are standardized effect size measures (Schielzeth, 2010). In JA, F2 at the onset has the highest absolute β value and shows more of a sigmoidal curve (see Z2 Onset, Figure 1). In MA, however, F1 mid and F2 at both onset and mid have the highest absolute β values and show a near complete sigmoidal curve. When converting the β values to percent correct, all four acoustic correlates (in JA  Table A3 in Appendix 3 and Section 1.1.1).
Following perturbation theory and the acoustic theory of speech production (Carré & Mrayati, 1992;Chiba & Kajiyama, 1941;Fant, 1960Fant, /1971Howard & Angus, 2009;Johnson, 2012;Mrayati et al., 1988;Stevens, 1989), the location of the constriction responsible for producing pharyngealization is predicted to be located in the pharyngeal area, potentially between the upper to mid pharynx, as is generally reported in the literature. F3 frequencies, albeit non-significant, show a tendency for an overall raising pattern from /d/ to /dˤ/; however, our results are consistent with a lowered F3 in front vowels (by around 70 Hz) and a raised F3 in back vowels (by around 200 Hz) from /d/ to /dˤ/; patterns that are observed in our previous studies (Al-Tamimi, J., 2007a, b, 2009) (see also Section 2.2.1).

Bark-difference formants
Bark-difference formant frequencies are used as a psychoacoustic representation of the patterns observed in formants and is claimed to reflect the openness, retraction, and spectral divergence of vowels (Fahey et al., 1996;Hoemeke & Diehl, 1994;Sylak, 2011;Syrdal & Gopal, 1986;Traunmüller, 1981;Wood, 1986). Model comparison results summarized in Table 3 show that all Bark-difference metrics are significantly associated with the difference between the two categories. Specifically, in the /dˤ/ context, Z1-Z0 and Z3-Z2 show increased β values, whereas Z2-Z1 shows lowered ones. The predicted probabilities presented in Figure 2 show a near perfect sigmoidal curve for all Bark-difference metrics, with variable degrees of separation. Generally, MA shows higher β values for Z1-Z0 and an increased sigmoidal curve than in JA, reflecting a more open production for the vowels (see absolute formant frequencies results in Figure 1 above).
Z2-Z1 reflects compaction of the spectrum with more back and open vowel productions being associated with a lowered Z2-Z1. A low pharyngeal constriction will show Z2-Z1 values below 3 Bark values whereas a mid to upper pharynegal constriction will show higher Z2-Z1 values around 4.6 Bark (Al-Tamimi, F. . Our results show a lowering of Z2-Z1 values that is close to the latter figure, with JA being close to 5.5 Bark and MA around 6.2 Bark in the /dˤ/ context (see Table A3 in Appendix 3). And finally, Z3-Z2 values show large β values at the Onset for JA and at both positions in MA. They show a more retracted vowel quality that is close to that of the /o ɔ/ vowels (when compared with the results in Syrdal & Gopal, 1986, p. 1090). This large difference in the pharyngealized contexts is mostly due to the differences observed in the F2; an increased distance between Z3-Z2 is caused by the lowered F2 in the /dˤ/ rather than that of F3. This increased distance can also be correlated with spectral-flatness or divergence (Wood, 1986), and hence serves as an enhancing correlate to the already lowered Z2-Z1. The predictive strength of these correlates is variable with rates between 78-98% (see Table 3 for more detail).

Vowel spaces
We used relational formant frequencies to display vowel spaces; using the traditional F2*F1 (in Bark) vowel spaces and that of Z3-Z2*Z2-Z1 at both onset and midpoint of the vowels in JA and MA for both of /d/ (in red solid) and /dˤ/ (in blue dashed) contexts. Figures 3 and 4 show both Z2*Z1 (a) and Z3-Z2*Z2-Z1 (b) at onset (top) and midpoint (bottom) in both JA and MA respectively (charts generated with the package phonR, McCloy (2016)). Both figures are an attempt to provide a better understanding of how Bark-difference representation is well suited to account for the 'retracted' vowel quality as defined in the Laryngeal Articulator Model (Esling, 2005;Moisik et al., 2012;Sylak-Glassman, 2014b). Vowel spaces at the onset provide the clearest separation between the two categories compared with that at the midpoint, although in MA, the midpoint results lend support to a more open configuration (see Figure 4). At the onset, JA shows a clear back vowel quality as represented with absolute F2, whereas MA shows a clear open and back vowel articulation through absolute F1 and F2 respectively (see Figures 3a and 4a). Moving on to the Bark-difference results (see Figures 3b and 4b), and particularly at the onset (top), both JA and MA show a clear separation between the two categories, with vowels in the pharyngealized context showing a 'compact' (with lower Z2-Z1) and 'backed' (with higher Z3-Z2) production; they are 'retracted' in the sense of the Laryngeal Articulator Model (Esling, 2005;Moisik et al., 2012;Sylak-Glassman, 2014b). By using absolute formants only (Figures 3a and 4a), vowels in JA are not represented with an open configuration as all vowels seem to range around 4 Bark. Our claim is then if one is to acoustically evaluate

Voice quality
To our knowledge, this is the first attempt to evaluate any voice quality differences as associated with pharyngealization in Arabic (although see Alwan, 1986Alwan, , 1989, on the estimated bandwidth in pharyngeals). Our voice quality results are separated into spectral slope and high frequency energy (Figures 5 and 6 respectively). Starting with spectral slope, model comparison results suggest that H1*-A1*, H1*-A2*, and H1*-A3*, in either both or one position are significantly contributing to the difference between the two consonants in both dialects (Table 3). GLMM results show an overall decrease in β values in all significant metrics (except for a tendency to observe an increase in H1*-A3* Onset in JA). MA results show a stronger contribution of these metrics to the distinction between the consonants compared to JA. This is shown further in the curve shape in Figure 5 where a stronger declination from /dˤ/ to /d/ can be seen in MA for all significant correlates; H1*-A1*, H1*-A2* at both onset and midpoint show a near sigmoidal curve, although this is not comparable to those seen for absolute and Bark-difference formants (see above). In JA, the curves seem more linear in significant correlates, with lower β and a predictive strength varying between 56-82% (Table 3). These spectral slope measures seem to act as secondary correlates to pharyngealization as they do not show the same predictive power as formant frequencies.
In both dialects, this decrease from /dˤ/ to /d/ is a clear indication of a tense/pressed voice quality that is related to a greater glottal constriction (Keating et al., 2015). This leads to variation in the bandwidths of F1 and F2 (through H1*-A1* and H1*-A2* respectively) that accompanies tense/pressed voice quality; an increased bandwidth of F1 and F2 causes a decrease in the amplitude of the harmonic closest to F1 and F2, respectively (Garellek, 2012;Hanson et al., 2001). And finally, a decrease in H1*-A3* is indicative of tense phonation with a flatter spectrum (Garellek, 2012;Klatt & Klatt, 1990;Kuang & Keating, 2012). This seems to be the case with MA as a more abruptly constricted glottis can lead to an increase in energy around F3 and hence a flatter spectrum and a lowered H1*-A3* (Hanson & Chuang, 1999;Hanson et al., 2001). For JA however, the increased H1*-A3* is indicative of a lowered energy around F3 in the pharyngealized context, which seems to correlate well with the predictions of a constricted glottis with a simultaneous front to back closure along the length of the vocal folds (for more detail, see e.g., Hanson & Chuang, 1999;Hanson et al., 2001;Klatt & Klatt, 1990;Stevens, 1998). Both dialects display tense voice quality that is associated with pharyngealization though spectral slope measures, but seem to display different patterns in terms of how constricted the glottis is; abrupt in MA and with a simultaneous front to back closure of the vocal folds in JA.
Moving on to high frequency components, and through model comparison (Table 3), not all acoustic correlates show a significant contribution to the difference between the two consonants. In JA, only A1*-A3* and A2*-A3* at the onset are significant, whereas in MA, it is A1*-A2* and A2*-A3* at the onset and midpoint. This can also be seen from the predicted probabilities as shown in Figure 6 where some curves are near sigmoidal, while others are linear. The observed differences (through β and percent correct values) are small compared to formant based measures, and these high frequency components seem to act as secondary correlates. More specifically, a lower A1*-A2* is correlated with a constriction closer to the tongue root (i.e., mid to lower pharynx) as is the case in languages with [-atr] vowel specification (see e.g., Aralova et al., 2011;Fulop et al., 1998;Guion et al., 2004;Kang & Ko, 2012, among others). This suggests that in MA, a lower constriction location for pharyngealization may be in operation, although a nonsignificantly lowered A1*-A2* at the onset in JA is obtained. With respect to the two other metrics, an increase in A1*-A3* and A2*-A3* is observed in both JA and MA with raised predicted probabilities (see Figure 6). This increase is indicative of a relatively decreased energy around F3 with respect to that of F1 or F2. In fact, A1 and A2 are already high in energy due to the pharyngeal constriction, and the change observed in A3 energy would be indicative of a change due to the epilaryngeal constriction that leads to a less/flatter spectral tilt (Halle & Stevens, 1969). The combination of spectral slope and the high frequency components are indicative of a constricted epilarynx; the former shows acoustic correlates of a tense voice quality, and the latter a constricted epilarynx with a relative increase in A1*-A3* and A2*-A3* due to decreased energy around F3. This combination suggests that pharyngealization in Arabic is associated with constricted ventricular folds (Moisik et al., 2012;Moisik & Esling, 2011;Sylak-Glassman, 2014a). This constriction of the glottis leads to a tense voice quality that is potentially associated with Figure 6: Predicted probability curves spanning -3 to +3 z-score for the High frequency components in both JA (blue solid) and MA (red dashed). Falling curves indicate a lowering in the predicted probabilities from /dˤ/ to /d/ (as in A1*-A2*) and vice versa. The reference line (in gray) indicates the non difference line. a constricted and/or raised larynx posture (Klatt & Klatt, 1990;Laver, 1994;Sundberg & Askenfelt, 1981). These novel acoustic consequences require further investigation from an articulatory point of view to shed light into this secondary correlate of pharyngealization in Arabic.

Summary of results
The results presented above showed the vowels in the vicinity of pharyngealized consonants in Arabic to be 'retracted' (Esling, 2005;Sylak-Glassman, 2014a, b). They are produced as more open (higher F1, Z1-Z0) and more back (lower F2, higher Z3-Z2), with a constriction in the pharyngeal area that causes compaction of the spectrum (lower Z2-Z1), and spectral divergence as an enhancing correlate to the compacted spectrum (higher Z3-Z2). Z2-Z1 seems to provide a better combined correlate to the compaction of the spectrum rather than the separate F1 and F2, although both are related to each other (see Section 3.1.3).
Our novel results with respect to spectral slope and the high frequency components seem to suggest that pharyngealization in Arabic is associated with a secondary constricted ventricular folds and hence a constricted epilarynx; this induces a tense voice quality with lower H1*-A1*, H1*-A2*, H1*-A3*, and A1*-A2*, and an increased A1*-A3* and A2*-A3* due to the decreased energy around F3. This is caused by a secondary constriction of the epilarynx, and more specifically the ventricular folds. Given that we used individual GLMM analyses due to the constraints of our data (high collinearity), the individual β and percent correct values provided some insights into the discriminatory power of each of these acoustic correlates. However, it is not clear how these behave together and which acoustic correlates are more informative than the others. The next section presents an exploratory classification technique, namely random forests, that will shed light into the discriminatory power of the combined acoustic correlates.

Random Forests
The results presented in the previous section showed that none of the acoustic correlates exhibited a completely flat line (i.e., β = 0 and less than 50% correct), thus we included all 26 acoustic correlates (the 13 acoustic correlates at both onset and midpoint) in the random forest analysis for each dialect separately (a separate analysis on the significant only results showed the same overall percent correct classification, and correlation with actual data). Overall, the random forest analysis correlated well with our actual data (JA: r 2 = 0.87, mean squared error: 0.065; MA: r 2 = 0.82, mean squared error: 0.088). After running the random forest analysis, we used the predict function from the party package to provide predictions based on the out-of-bag data for cross-classification by specifying OOB=TRUE. This allows the algorithm to train itself on two-thirds of the data (the in-bag set), and then to use the remaining third of the data (the out-of-bag set) for classification (see Hothorn et al., 2006;James et al., 2013, for more detail). The overall percent correct classification with out-of-bag cross-classification was 93.5% for JA and 91.1% for MA, indicating a high discriminatory power of all of our acoustic correlates. Following this, we ran conditional permutations variable importance to measure the strength of each of the acoustic correlates conditional on each other and by using varimpAUC and conditional=TRUE. Figure 7 provides the ranking of the different acoustic correlates using the mean decrease in accuracy in the overall classification when a particular acoustic correlate is removed.
The results show a clear separation between formant-based and voice quality-based measures with the former being highly predictive of the two consonant category (Figure 7). Looking at the results in detail, the ranking of the predictors in JA (Figure 7a) shows that F2 Onset, Z3-Z2 Onset, followed by Z2-Z1 Onset, are the most important correlates, whereas in MA (Figure 7b), Z2-Z1 Onset followed by F2 Onset, and Z3-Z2 Onset are the most important. All the other correlates have lower values and thus can be regarded as secondary. It can be claimed that these differences in order are indicative of a difference in the phonetic implementation of pharyngealization across these dialects; when compared with the GLMM results above, MA potentially shows more mid-lower pharyngeal constriction, while JA shows mid-upper pharyngeal constriction. More articulatory data is needed to shed light on this difference. It should be noted that the differences between the dialects in the range of mean decrease in accuracy is large, with JA showing a score of 0.027 of the conditional mean decrease in accuracy when F2 Onset is removed, whereas in MA it is a 0.015 decrease for F2 Onset. These results correlate well with the β values presented in the GLMM results (see Section 3.1).
Following the initial random forest classification results, we ran six additional ones that are only used for their predictive accuracy. Table 4 provides a summary of the predictive accuracy of each of these additional random forests and as a comparison, the results of the initial classification are provided (first column).
The results point toward formant-based metrics as the most predictive for the difference between the two contexts. This is to be expected as pharyngealization is primarily reported as affecting formant frequencies. When both absolute and Bark-difference metrics are used together (i.e., Form+BkDiff), they show a non-significant lower classification rate than that of the full model (by 0.1-0.3%). This indicates that formant-based metrics have the most explanatory power, and may be suggestive of a   3b and 4b) and the current classification results show that they correlate well with the 'retracted' quality of the vowels in the pharyngealized context, following LAM (Esling, 2005;Sylak-Glassman, 2014a, b). Overall, the ranking of the various metrics was comparable to those reported in the full model (see Figure 7). However when absolute formants were used on their own, F2 at the Onset was the main correlate in both dialects. As expected, these results point to the fact that formant-based metrics are the primary correlates to pharyngealization in JA and MA. Finally, the predictive accuracy of voice quality metrics on their own was assessed. The results presented in the last column of Table 4 suggest that in both JA and MA, voice quality metrics are relatively important as they provide a classification accuracy of 70.2% for JA and 75.8% for MA. These rates are lower than those obtained for formant-based metrics, but provide support for voice quality metrics to act as a secondary correlates to pharyngealization. The ranking of correlates (not presented here) matches what we already saw in Figure 7, in that A1*-A3*, A2*-A3*, H1*-A1*, H1*-A2* are the most predictive acoustic correlates in both dialects. However, in MA, spectral slope correlates are used more than in JA, and JA tends to use more the high-frequency components more. In both dialects, voice quality measures are indicative of an epilaryngeal constriction used as a secondary articulatory setting with formant-based metrics being the best predictors.

Discussion and Conclusion
This exploratory study is aimed at investigating whether pharyngealization in Arabic is associated with an epilaryngeal constriction from an acoustic perspective. Traditionally, pharyngealization in Arabic is generally assumed to involve tongue body (dorsum) retraction towards the upper-pharyngeal areas that leads to a lowering of the second formant in the surrounding vowels (e.g., Bin-Muqbil, 2006;Ghazeli, 1977;Watson, 2007;Zawaydeh, 1999;Zawaydeh & de Jong, 2011). However, both classic and more recent articulatory evidence show this constriction to be located much lower in the pharynx, with epiglottis retraction, and raising of the larynx that leads to a pressed/tense voice quality (see e.g., Cantineau, 1960;Hess, 1998;Laufer & Baer, 1988;Lehn, 1963;Zeroual & Clements, 2015;Zeroual et al., 2011, among others). Following the Laryngeal Articulator Model (Esling, 2005), 'true' pharyngeals seem to show this type of configuration, with the whole epilarynx being constricted around the aryepiglottic folds. Pharyngealization on the other hand seems to show this type of constriction albeit as a secondary one with a primary pharyngeal constriction (Esling, 2005;Moisik, 2013a;Sylak-Glassman, 2014a, b). Acoustically, only the first three formants are usually investigated (see Section 1.1.1 above). However, these are only indicative of a pharyngeal constriction and do not directly explain any voice quality correlates that are a by-product of an epilaryngeal constriction. An epilaryngeal constriction is expected to show a combined effect of the back and down movement of the tongue as well as laryngeal muscles constriction (Esling, 2005;Moisik, 2013a;Sylak-Glassman, 2014a, b).
This study investigates the absolute formant frequencies, which is typical of such studies, but also complements this by looking at Bark-difference metrics and voice quality correlates. This is the first study that combines these acoustic correlates in investigating pharyngealization in Arabic. Our results with respect to absolute formants are in agreement with previous literature, with vowels in the pharyngealized context showing decrease in F2 and an increase in F1 regardless of the vowel quality. In addition, F3 shows an overall increase in both dialects, but seems to show lowering when the vowel is front, and rising when the vowel is back (see Section 3.1.1 and Figure 1). Using the Distinctive Regions Model (DRM) (Carré & Mrayati, 1992;Mrayati et al., 1988) that is based on the principles of Perturbation Theory (Chiba & Kajiyama, 1941), it is possible to predict the constriction location based on the acoustic output. When F1 is rising, F2 is falling, and F3 is rising as in back and central vowels, the location of the constriction is close to the DRM region R4 in the upper pharynx. A combination of rising F1, and falling F2 and F3 seems to be close to the DRM region R3 in the mid-lower pharynx (see e.g., Carré & Mrayati, 1992, Figure 8, p. 150 and Figure 12, p. 156). These acoustic results are compatible with the articulatory accounts presented in Ghazeli (1977, p. 174, as cited in Laufer & Baer, 1988 that 'emphatic' consonants have a secondary tongue retraction located midway between uvulars and pharyngeals. Following the predictions of the Laryngeal Articulator Model (Esling, 2005;Sylak-Glassman, 2014a, b) pharyngealization is associated with a retracted production that has a combined back and down gesture. When considering this articulatory configuration, it is expected to observe both a lowering of F2 and a rising of F1 as a combined acoustic consequence. Thus the distance between these two formants can be seen as an alternative metric to signal the effects of pharyngealization, i.e., Z2-Z1. We investigated other formant metrics, e.g., Z1-Z0 and Z3-Z2 as correlates of a more open and a more back articulation and these correlated well with the traditional F1 and F2 dimensions separately. These have provided a clearer picture to the difference between the two consonants as reflected by the acoustic vowels spaces (see Figures 3 and 4).
The figures obtained for Z2-Z1 (see Section 3.1.2) can also be correlated with an uppermid pharyngeal constriction. Compared to /d/, Z2-Z1 in the pharyngealized context was significantly lower and averaged around 5 to 6 Bark in both dialects (see Table A3 in Appendix 3), which is close to the range reported in other studies (see e.g., Al-Tamimi, F. . These figures are of course indicative of a much higher constriction location than 'true' pharyngeals that have a smaller distance of about 3 Bark. Z1-Z0 and F1 results in both dialects are indicative of a more open production of the vowels in the vicinity of /dˤ/ (Fahey et al., 1996;Hoemeke & Diehl, 1994;Syrdal & Gopal, 1986;Traunmüller, 1981). A 0.5 to 1 Bark difference (78 to 118 Hz) that correlates well with a one-degree change on the openness dimension agrees with the similarity scales between vowels and pharyngeals or uvulars as reported in Sylak-Glassman (2014b). When in contact with either category, vowels tend to be produced with one supplementary opening degree. In fact, the figures reported for Z2-Z1 are correlated with spectral integration (Fahey et al., 1996;Traunmüller, 1983Traunmüller, , 1984. When dealing with 'phonetically' similar vowels, i.e., allophones changing in quality due to (non-) pharyngealization, a Z2-Z1 distance close to or below 6 Bark signals spectral integration (Traunmüller, 1983, p. 5-6). This as a whole leads to the conclusions that Z2-Z1 seems to be the major correlate to distinguish the two categories as the closeness of formants leads to them being perceived as one. And finally, Z3-Z2 provided another important acoustic correlate to the /d/-/dˤ/ distinction: The pharyngealized category showed higher values with a difference of 1 to 2 Bark. In the non-pharyngealized context, all Z3-Z2 values were below the 3.5 Bark threshold for spectral integration (Chistovich & Lublinskaya, 1979;Syrdal & Gopal, 1986). We use the term 'spectral flatness' or divergence (Wood, 1986) to explain the results in the pharyngealized context; vowels in the /dˤ/ environment show a clear separation between the F2 and F3 due to the extremely lowered F2 and the relatively 'stable' F3. Thus the spectral divergence through the larger Z3-Z2 difference in the pharyngealized context is acting as an enhancement feature to the already spectrally integrated peak, i.e., Z2-Z1.
The combination of these three psychoacoustic metrics at the onset of the vowels in the pharyngealized context seems to show a direct psychoacoustic manifestation of the [+flat] feature as initially suggested by (Jakobson, 1957(Jakobson, /1962Jakobson et al., 1952Jakobson et al., /1976) as an acoustic correlate to pharyngealization. Z2-Z1 seems to be the main psychoacoustic correlate as it is close and below the threshold for spectral integration, Z3-Z2 seems to play a role as an enhancement feature and Z1-Z0 seems to correlate well with the more open articulation.
A second articulatory consequence of an epilaryngeal constriction is voice quality changes. Ventricular folds constriction leads to a tense, harsh, laryngealized voice quality, whereas an aryepiglottic constriction leads to 'trilling' as seen in 'true' pharyngeals and an 'enhanced' voice quality (Edmondson & Esling, 2006;Esling, 1996;Moisik, 2013a, b;Story, 2016). Our results with respect to voice quality correlates-spectral slope and tilt-provide a secondary description to pharyngealization in Arabic. Pharyngealization is associated with a tense voice quality that is caused by constricting the intrinsic muscles of the larynx-the ventricular folds (see e.g., Garellek, 2012;Hanson & Chuang, 1999;Hanson et al., 2001;Keating et al., 2015;Kuang & Keating, 2012). Our estimated F1 and F2 bandwidths correlate well with a tense/pressed voice quality through a decrease in H1*-A1* and H1*-A2* metrics (Garellek, 2012;Hanson et al., 2001). Our estimated bandwidth through harmonic amplitude difference shows a clear difference from the results obtained by Alwan (1986Alwan ( , 1989, as an increased F2 bandwidth was found in pharyngeals but an increased F1 bandwidth in uvulars. Our results suggest pharyngealized consonants to be different from both pharyngeals and uvulars through the combined increase in the bandwidths of F1 and F2. Finally, a lowered H1*-A3* is also suggestive of a tense voice quality through an abrupt constriction of the glottis, specifically in MA (Garellek, 2012;Hanson & Chuang, 1999;Hanson et al., 2001;Klatt & Klatt, 1990). In JA, an increase in H1*-A3* is indicative of an increased spectral tilt and partially reduced energy around F3; consequences of a constricted glottis with a simultaneous front to back closure of the vocal folds (Hanson & Chuang, 1999;Hanson et al., 2001). This also seems to correlate well with a tense/pressed voice quality (Garellek, 2012;Hanson et al., 2001;Keating et al., 2015;Kuang & Keating, 2012).
The high frequency components provide an additional correlate to the glottal constriction through a constricted epilarynx and this constriction provides additional energy to the upper formants when comparing the two categories (Halle & Stevens, 1969;Stevens, 1977;Story, 2016). Our results showed that in both JA and MA an increased A1*-A3* and A2*-A3* is found. This is caused by a decrease in the energy around F3 that is potentially caused by the type of constriction in the glottis. Constricting the epilarynx on its own would lead to a decrease in the upper harmonic differences as this would lead to a concentration of the harmonics closest to F3, F4, and F5. This is seen in the singer's formant-a primary technique used in signing to enhance the voice quality (Story, 2016;Titze & Story, 1997). In our case, constricting the epilarynx is secondary; pharyngeal constriction leads to a boost in the energy around F1 and F2, and the changes seen with respect to F3 will be minimal. This causes a decrease in the energy around F3 leading to the increased harmonic difference observed in this study. Liénard and Di Benedetto (1999) found that higher levels of vocal effort seem to increase the amplitudes of A2 and A3 more than what is seen in A1; the latter is relatively higher in amplitude that the former. Their harmonic differences, based on our estimations, seem to show a similar pattern to the results reported here; a relatively increased A1*-A3* and A2*-A3* in the high effort is observed compared to the normal setting (Liénard & Di Benedetto, 1999, Figure 3, p. 416). In this particular case, the high effort was measured in the vowels; our results are mostly seen at the onset of the vowel and are indicative of this secondary constriction that shows a high effort comparable to what is seen in tense voice. Finally, and according to Story (2016, Figure 7, p. 10), a constricted epilarynx would lead to an increase in the frequency and amplitude of F3 to make it move closer to F4. This closeness correlates with the singer's formant and changes the 'timbre' or quality of voice of a particular speaker. F1 and F2 are responsible for the phonetic quality of a vowel, whereas voice quality is manifested through F3 to F5 frequencies and amplitudes. Hence if we postulate that pharyngealization is associated with an epilaryngeal constriction, F3 to F5 amplitude and frequencies should be investigated as well. This future work is currently planned on a forthcoming dataset.
The Random forest technique was used as a predictive model to test for classification rates and to provide a fine-tuned ordering of the various correlates used in this study. The results showed a clear separation between formant-based and voice quality-based measures. Voice quality results are secondary in the sense that they allow for a finer description of pharyngealization in Arabic as involving a constricted glottis through constriction of the epilarynx. The primary acoustic correlates are formant-based and reflect the changes in the vowel quality into a more open (high F1 and Z1-Z0), a more compact (low Z2-Z1), a more retracted (low F2, and high Z3-Z2) with spectral integration (low Z2-Z1) and divergence (high Z3-Z2). The primary acoustic correlate is different in the two dialects if taking into account the psychoacoustic correlates: In JA it is F2-based with retraction and spectral divergence, while in MA it is a combined F1 and F2-based through a more compact spectrum with retraction. When considering absolute formant frequencies alone, both dialects display the same pattern i.e., F2-based followed by F1-based; results that are largely reported in the literature (see Section 1.1.1). Bark-difference metrics highlight the fact that the supposed differences between Arabic dialects in how pharyngealization is implemented (with some dialects having a more 'guttural' quality) can be evaluated from an acoustic point of view: It is possible that MA has a more 'pharyngeal' quality than JA (see e.g., Bellem, 2007;Embarki et al., 2011), though it is not clear whether these differences are only based on formant patterns or voice quality ('timbre' of the voice). More detailed articulatory and acoustic description is required to shed light into these differences.

Implications for formal representations
Various features were proposed in the literature to describe the effects of pharyngealization, and these include: [Constricted Pharynx] (Hoberman, 1987); [Constricted Tongue Root] (Stevens, 1998); [Constricted] (Edmondson & Esling, 2006); [Constricted Epilaryngeal Tube] (Moisik, 2013a;Moisik et al., 2012;Moisik & Esling, 2011); [Constricted Epilarynx] (Sylak-Glassman, 2014a); [Expanded Pharynx] (Lindau, 1975(Lindau, , 1978; [flat] (Jakobson, 1957(Jakobson, /1962Jakobson et al., 1952Jakobson et al., /1976); [Guttural] (Hayward & Hayward, 1989;Watson, 2007); [Lower/Upper Pharynx] (Czaykowska-Higgins, 1987); [Pharyngeal] (McCarthy, 1994;Zeroual & Clements, 2015); [Retracted] (Esling, 2005;Sylak-Glassman, 2014b); [Retracted Tongue Root] (Shahin, 1996(Shahin, , 1997(Shahin, , 2011Watson, 2007). This list is not exhaustive, however, it highlights the need for a unified set of feature(s) to account for pharyngealization. Currently, [+rtr] is the widely used feature specification for pharyngealization at least in Arabic that reflects the activity of the retraction of the root of the tongue. It is not clear however whether this retraction causes only lowering of F2 frequencies (i.e., Z2 and Z3-Z2) or whether it accounts for the more open production with rising of F1 (i.e., Z1, Z1-Z0). In addition, the more compacted spectrum through the lower Z2-Z1 does not seem to be reflected by [+rtr]. Finally, it is not clear whether constricting the epilarynx can be accounted for by [+rtr]. It is of course possible to add an additional laryngeal specification, i.e., [+constricted glottis] that can account for our secondary correlates. However, given that the tongue retraction as seen here is a by-product of constricting the epilarynx (e.g., Esling, 2005;Moisik, 2013a;Sylak-Glassman, 2014a, b, and Section 1.2), we follow Moisik et al. (2012), Moisik and Esling (2011), and Sylak-Glassman (2014a) and postulate that pharyngealization in Arabic is produced by an epilaryngeal constriction that causes the tongue root and body to be retracted by a back and down gesture. This induces a laryngeal constriction leading to a raised larynx posture and a tense/pressed voice quality. Based on these descriptions and on our results, pharyngealization in Arabic will then be associated with a [+cet] that has 'retraction' and 'constricted glottis' as one component. The feature [+ce] (Sylak-Glassman, 2014a) that follows [+cet], can account for this combined effect, though it is not clear whether [+constricted glottis] is inherent to [+ce] or should be added as an additional feature to describe pharyngealization (see Sylak-Glassman, 2014a, p. 136-140). The inclusion of voice quality metrics allows for pharyngealization in Arabic to be described with [+cet]; if formant metrics only were used here, our conclusion would be that [+rtd] (for 'retraction,' following Moisik et al., 2012;Moisik & Esling, 2011) is the only correlate of pharyngealization. Of course this analysis is only based on the outcomes of this study and does not look at the patterns in the language. However, highlighting this laryngeal activity as associated with pharyngealization in Arabic can potentially shed light into what a 'guttural' quality is from a phonetic point of view and whether it is a 'voice quality' in the sense of 'timbre' of the vowel (following Story, 2016), a 'darker' or 'heavier' auditory quality from formant merging, or both.

Limitations
This exploratory study highlights some novel acoustic evidence in describing pharyngealization in Arabic. However, it should be noted that only coronal voiced stops were investigated. Our formant-based measures are congruent with those reported in the literature in the direction of the effect, however, there are some differences with the amount of distance between these consonants (on F1, F2, and Z2-Z1 values). It is not clear whether some of the results obtained are restricted to this category of sounds or whether they can also be extended to other coronal consonants. We expect the acoustic correlates of pharyngealization reported here to be reflected in all pharyngealized consonants to different degrees. Our voice quality results should not be seen as restricted to this category of sounds, as it is associated with a tense/pressed voice quality through an epilaryngeal constriction in general, thus we expect to obtain the same degree of high frequency components regardless of the consonant category. An additional limitation to our study is the use of stimuli varying in location of the pharyngealized consonants according to syllable structure and to the surrounding sounds, i.e., other gutturals. This was due to finding real words fitting this consonant/vowel combination. Thus some of the outcomes may have been different if the corpus was more balanced; some dialectal differences may be due to variable syllable structures, or voice quality results at the midpoint would have been different. Given the exploratory nature of this study, we wanted to evaluate the results on such a corpus to allow for testing this particular hypothesis. Our exploratory study did not investigate the acoustic correlates of 'true' pharyngeals or uvulars as our aim was to evaluate which acoustic correlates are mostly associated with pharyngealization to allow for an extension to other categories. Finally, our study is acoustic in nature, and any conclusions on exact articulatory consequences should be evaluated cautiously. A planned subsequent project will investigate further this idea of an epilaryngeal constriction as associated with back consonants and will specifically use a combined articulatory (e.g., ultrasound, and electoglottagraphy) and advanced acoustic data of all front and back consonants in Arabic to shed light into the laryngeal constriction and the exact constriction location of each consonant; this will allow for a meaningful articulatory-to-acoustic mapping.

Conclusion
This exploratory study of Jordanian and Moroccan Arabic looked at the combined effect of multiple acoustic correlates in describing pharyngealization. By following the predictions of the Laryngeal Articulator Model (Esling, 2005) and its subsequent developments (Moisik, 2013a;Sylak-Glassman, 2014a), we showed how pharyngealization in Arabic is associated with a 'retracted' production, with a combined back and down gesture through an epilaryngeal constriction leading to a constricted glottis. By using 13 different acoustic correlates obtained at each of onset and midpoint of the vowel, including formant-based and voice quality-based measures, it was possible to evaluate the 'retraction,' compaction, and epilaryngeal constriction in our data. Formant distance measures (through Bark-difference) allowed for a better evaluation of the spectral convergence (Z2-Z1) and divergence (Z3-Z2) and showed differences in how pharyngealization is implemented between the two dialects. In summary, the results suggest that MA is associated with a more mid-low pharyngeal constriction and JA with an upper-mid pharyngeal constriction. Voice quality measures from spectral slope and high frequency components correlated well with a tense/pressed voice quality and an epilaryngeal constriction respectively. These novel results were assessed through GLMM and exploratory random forest analyses. 'Retraction' (i.e., combined back and down gesture) is the primary acoustic correlate of pharyngealization in Arabic with an epilaryngeal constriction leading to a [+constricted glottis] as a secondary correlate through an epilaryngeal constriction. If formant-based measures were used alone, a simple 'retraction' would be the main and only correlate to pharyngealization. Voice quality correlates allowed for the epilaryngeal constriction to be included. A combined articulatory and acoustic investigation of the state of the epilarynx in Arabic pharyngeals and pharyngealization is worth pursuing in order to evaluate the role of the epilarynx as an active 'articulator.' A subsequent perceptual study will shed light into which acoustic correlates are the most prominent in identifying pharyngeals and pharyngealization in Arabic. All these can also be the basis for dialectal classification and discrimination as presented in the current study.