As mentioned above, most previous acoustic work on guttural fricatives has been conducted on Arabic, which has a set of five relevant phonemes – voiceless and voiced uvulars /χ, ʁ/ (realized as velars or post-velars in some dialects), voiceless and voiced pharyngeals /ħ, ʕ/, and the voiceless glottal /h/ (Watson, 2007). In terms of their articulation, the Arabic uvular fricatives have been described as produced with the tongue dorsum making a narrow constriction against the uvula, with a secondary constriction at the upper pharyngeal wall (Zawaydeh, 1999). The pharyngeals typically involve a retraction of the tongue root and/or the epiglottis towards the lower pharyngeal wall and some raising of the larynx. While the voiced pharyngeal /ʕ/ is typically classified as fricative, it is often produced with an approximant-like constriction (and occasionally as an epiglottal stop; see McCarthy, 1994; Khattab, Al-Tamimi, & Alsiraih, 2018; Shosted, Fu, & Hermes, 2018 for reviews). The articulation of the laryngeal /h/ is more controversial, as some studies have described it as lacking a supralaryngeal constriction (Zawaydeh, 1999), while others inferring a pharyngeal constriction (Shahin, 2003; Bin-Muqbil, 2006), thus corroborating the sound’s phonological behavior (cf. McCarthy, 1991, 1994).

Acoustic studies of Arabic sounds (Obrecht, 1961; Ghazeli, 1977; Nartey, 1982; Norlin, 1983; Alwan, 1986) found that voiceless guttural fricatives are characterized by a strong non-periodic noise, yet with some formant structure. The latter property is more evident in the voiced sounds and, particularly, in the pharyngeal /ʕ/ (given its frequent approximant-like realization). Voiceless and voiced uvulars were observed to have relatively compact spectra with prominent narrow peaks at mid frequencies, as well as secondary peaks further up. Spectra of pharyngeals were less compact, showing broad peaks at somewhat higher frequencies and a more abrupt decrease in energy at higher frequencies (Obrecht, 1961; Nartey, 1982; Norlin, 1983; Alwan, 1986; Bin-Muqbil, 2006). The laryngeal /h/ showed the flattest shape, with relatively low energy spread at low and mid frequencies (Obrecht, 1961; Norlin, 1983).

To quantitatively capture spectral differences in fricatives, many researchers employed measurements of spectral moments, such as center of gravity (COG, or centroid, in Hz), standard deviation (SD), skewness, and kurtosis (see Forrest, Weismer, Milenkovic, & Dougall, 1988). COG, which reflects the mean concentration of frication noise across frequencies, is known to correlate with place of articulation: the further back the constriction is in the mouth (and thus the larger the front cavity is), the lower is the COG of the sound (Gordon, Barthmaier, & Sands, 2002). Not surprisingly, studies of Arabic fricatives found COG to be lower for gutturals as a class (as opposed to e.g. coronals), as well as lower for pharyngeals than uvulars (Obrecht, 1961; Norlin, 1983; Al-Khairy, 2005; Abu-Al-Makarem, 2005; Bin-Muqbil, 2006; Al-Tamimi & Khattab, 2015), lower for the laryngeal /h/ than the uvular /χ/ (Obrecht, 1961; Abu-Al-Makarem, 2005; Al-Tamimi, & Khattab, 2015) and – in some cases – the pharyngeal /ħ/ (Norlin, 1983). Similar place differences were found for voiced uvulars and pharyngeals. In addition, COG was found to be lower for voiced than voiceless fricatives (Obrecht, 1961; Norlin, 1983; Al-Khairy, 2005; Abu-Al-Makarem, 2005; Al-Tamimi & Khattab, 2015), likely reflecting the prominence of formants in the former sounds.

Other spectral moments were also found useful in distinguishing the contrast. SD, which reflects the range of energy distribution around the mean (COG), was observed to be higher for uvulars than pharyngeals, at least for the voiceless set (Norlin, 1983; Al-Khairy, 2005; Abu-Al-Makarem, 2005; Al-Tamimi, & Khattab, 2015; but see Bin-Muqbil, 2006 on the reverse difference). The laryngeal /h/ was similar to either voiceless uvulars (Norlin, 1983; Al-Khairy, 2005) or pharyngeals (Abu-Al-Makarem, 2005; Al-Tamimi & Khattab, 2015). Skewness, which compares the energy distribution below and above the mean, was found to be higher for pharyngeals than uvulars, while yielding inconsistent results for /h/ (Al-Khairy, 2005; Abu-Al-Makarem, 2005; Bin-Muqbil, 2006). These findings reflect the earlier mentioned abrupt fall-off of energy at higher frequencies for pharyngeals and a relatively flat spectrum for /h/. Finally, place and voicing differences were also observed for kurtosis, a measure of the relative peakedness of the spectrum. Specifically, pharyngeals showed a higher kurtosis than uvulars, and so did voiced fricatives compared to voiceless ones (Al-Khairy, 2005; Abu-Al-Makarem, 2005; Bin-Muqbil, 2006).

Studies that examined relative intensity of fricative noise found that the measure was generally higher for voiceless than voiced fricatives, except for /h/ which had the lowest intensity (Norlin, 1983; Al-Khairy, 2005). Most studies have examined duration of fricative noise, reporting it to be much shorter for /h/ than the other fricatives (Obrecht, 1961; Al-Khairy, 2005; Abu-Al-Makarem, 2005; but see Al-Ani, 1970 and Alwan, 1986 for no difference), and – in some cases – shorter for voiced than voiceless uvulars and pharyngeals (Al-Khairy, 2005).

Considerably less phonetic work has been done on gutturals in other languages. Most of the studies we are aware of examined uvular fricatives in comparison to either more anterior fricatives (velars, if contrastive, or coronals) or the laryngeal /h/. Gordon et al. (2002) studied the acoustics of the uvulars /χ, χʷ/ in Montana Salish and Western Aleut (where these sounds contrast with the velars /x, xʷ/). The authors reported that the uvulars in Montana Salish were characterized with an acute noise peak at a low frequency and additional peaks at higher frequencies. The velar fricatives showed a similar low-frequency peak (although slightly higher in frequency), while lacking any prominent higher frequency peaks. Despite these differences, COG failed to distinguish the fricatives in both Montana Salish and Western Aleut (while showing significant differences in formant transitions). Hargus, Levow, and Wright (2021) compared /χ/ to /h/ in Deg Xinag (Athapaskan), finding that the uvular was characterized by higher COG, SD, and intensity, and by lower kurtosis. Skewness did not distinguish the two fricatives. Nartey (1982) examined the voiceless uvular fricative /χ/ (among other fricatives) in Hebrew, characterizing its spectrum as having a low-frequency primary peak and a medium-frequency secondary peak. Frequencies of these peaks appeared similar to those of the Egyptian Arabic /χ/ and Navajo /x/, also examined in the study. No quantitative comparisons of within or between languages, however, were performed. While some – albeit relatively limited – literature exists on uvular fricatives, we are not aware of any spectral investigations of pharyngeals in languages other than Arabic. A few studies that have looked at the acoustics of pharyngeals in Interior Salish languages (Bessell, 1993; Shahin, 2003; Flemming et al., 2008) focused on these consonants’ formant patterns – likely due to these consonants’ typically approximant-like realizations. Similarly, there is a lack of acoustic data (apart from a few spectrograms) on pharyngeals in languages of the Caucasus, where these sounds have also been reported and noted to contrast with uvulars and laryngeals (cf. Ladefoged & Maddieson [1996] on Abkhaz, Agul, and Dargi).

The previous studies of guttural fricatives in Arabic reviewed above have established that spectral measurements of fricative noise, and primarily COG and SD, can relatively reliably differentiate the relevant place contrasts, while intensity and duration are particularly relevant to distinguishing the fricative voicing. To some extent, these findings were corroborated by studies of similar fricatives (mainly uvulars) in other languages. It should be mentioned that most studies reviewed above typically examined fricatives in either word-initial (e.g., Norlin, 1983; Al-Khairy, 2005; Abu-Al-Makarem, 2005) or word-medial positions (e.g., Alwan, 1986). Hardly any work has been done on fricatives in word-final position, or across positions (with Al-Ani [1970] being an exception). Therefore, we chose to examine possible acoustic variation of fricatives as a function of position.1 It has been observed that fricatives are generally more resistant to positional effects than other consonants, likely due to the higher articulatory precision required to generate noise turbulence (Byrd, 1996). Some positional effects, however, have been observed as, for example, a longer duration and higher COG for fricatives in word-initial position, compared to word-medial and word-final positions (Kochetov, 2017 on Russian).

The studies reviewed above have examined specific varieties of Modern Standard Arabic, as spoken primarily in Egypt (Norlin, 1983), Iraq (Alwan, 1986), Lebanon (Obrecht, 1961; Al-Tamimi & Khattab, 2015), and Saudi Arabia (Al-Khairy, 2005). Less is known about the acoustic realization of gutturals in Emirati Arabic, a relatively phonetically understudied variety of Persian Gulf Arabic (Leung, Ntelitheos, & Al-Kaabi, 2020). One exception to this is Abu-Al-Makarem’s (2005) study of fricatives in the Al-Khat colloquial variety of Persian Gulf Arabic spoken in Saudi Arabia. Even less is known about guttural fricatives in other languages of the broader Middle East region. Therefore, the research gap in the phonetic study of gutturals in this region was a major motivating factor for our choice of Iraqi Central Kurdish and Lebanese Western Armenian. Both languages are phonetically understudied, particularly with regards to fricatives. In addition, Western Armenian is of interest due to its endangered status in Anatolia and the Middle East (UNESCO World Atlas of Languages, n.d.). Importantly, a comparison of guttural fricatives of these two languages with the much more studied sounds of Arabic would allow us to identify cross-linguistically common patterns or differences in the phonetic realization of these sound categories.

Based on results of previous studies (mainly on Arabic), we can make several predictions about the differentiation of fricatives in place and voicing. Specifically, uvulars should be distinguished from pharyngeals by a higher COG and SD, and a lower skewness and kurtosis. Compared to the other voiceless fricatives, the laryngeal /h/ would be expected to show a lower COG (at least compared to /χ/), and lower intensity and duration, with possible other spectral differences. Voiced uvular and pharyngeal fricatives, in their turn, would be expected to show lower COG, intensity, and duration, and higher kurtosis values than their voiceless counterparts. Based on results of Tahtadjian’s (2023) ultrasound study on Armenian dorsal consonants, the laryngeal /h/ in Lebanese Western Armenian should have a lower COG than the uvulars. No specific predictions are made about between-language differences for specific fricative types, given the near-lack of previous phonetic research on these sounds in Central Kurdish.

While many previous studies of fricatives made use of quality recorders and acoustic chambers, in this study we are evaluating the segment classification method based on data collected online, the acoustic quality of which may be variable due to different recording conditions (see Section 2). An advantage of this method, on the other hand, is the ability to record multiple speakers residing in their home countries. This is in contrast to much of the previous works where fricatives were recorded from a handful of speakers, typically residing in the United States. It remains to be seen, however, whether results obtained from online recordings are comparable to previous studies. This is one of the questions our study will address.

Easier access to technological devices (such as smartphones) with the ability of audio recording on the one hand, and the COVID-19 pandemic on the other, motivated many researchers to adopt online methods to collect phonetic data remotely and analyze compressed audio files. While the use of such devices facilitates the process of data collection and shows to be reliable (Freeman & Decker, 2021), there are limitations to this method. For instance, Ge, Xiong, and Mok (2021) and Sanker, Babinski, Burns, Evans, Johns, Kim, Smith, Weber, and Bowern (2021) suggest that the reliability of remote data collection might be affected due to the variation in values of some acoustic parameters including vowel formants (cf. Freeman & Decker, 2021) and the COG of fricatives with energies concentrated at higher frequencies (sibilants). This variation, which could result in confusion for listeners and in general degrade the quality of recordings, is greater than what is observed in recordings conducted in a sound-attenuated booth. Another important factor affecting the quality of remotely collected data is the variation in the use of devices participants used for audio recording (phones or laptops). Additionally, audio recordings conducted via popular video conference apps (e.g., Skype, Zoom, Microsoft Teams) have been shown to affect certain acoustic parameters, such as formants values of vowels as well as degrees of vowels nasalization (Freeman & Decker, 2021). However, despite these challenges and limitations of the remote data collection, the method, with some considerations, is concluded to be a worthwhile medium to obtain data for the acoustic analysis of various phonetic variables (particularly F0 and duration; Bulgin, De Decker, & Nycz, 2010; Freeman & Decker, 2021). As our focus in this paper is on guttural fricatives that are typically characterized by energy at lower frequencies, we expect the limitations of the online method to be relatively moderate.

It is important to note that although the three selected languages have a set of what seems to be phonetically similar guttural phonemes (cf. Table 1), the phonological patterning and historical sources of these sounds are different in each language. In Arabic, for example, gutturals are part of a larger class of segments including uvular stops, the glottal stop, and pharyngealized consonants (emphatics) (Watson, 2007). These sounds share the common behavior of lowering adjacent vowels (McCarthy, 1991); they also pattern together in lexical co-occurrence restrictions (exhibiting the Obligatory Contour Principle as a similarity avoidance effect; Frisch, Pierrehumbert, & Broe, 2004). In contrast, there is no evidence, as far as we know, for the common phonological patterning of pharyngeals, uvulars, and laryngeals in Kurdish (Kahn, 1976; Barry, 2019). Similarly, no evidence exists for the uvulars and /h/ forming a natural class in Armenian (Vaux, 1998). Moreover, the set of gutturals in Armenian is much smaller compared to the other two languages. If language-specific sets of contrasts and phonological patterning have any say in how sounds are realized phonetically, we should expect considerable differences in the realizations of guttural contrasts across the three languages. There may also be some phonologically-based similarities. For example, in Arabic, uvulars have been observed to lower or back vowels in varieties of Armenian (Vaux, 1998). Somewhat similar height effects were also reported for (Northen) Kurdish pharyngeals: Barry (2019) observed that these sounds have a strong tendency to occur next to low vowels,2 while Kahn (1976) reported that pharyngeals actively lower and back following vowels. Thus, the three selected languages exhibit robust phonological differences, as well as some similarities with respect to their guttural phonemes. Whether these factors play a role in the phonetic realization of sounds is an interesting question, to which we will return in the Discussion.

The study involved 59 participants: 18 speakers of Emirati Arabic (10 females, 8 males), 20 speakers of Iraqi Kurdish (10 females, 10 males), and 21 speakers of Lebanese Armenian (12 females, 9 males), residing predominantly in Abu Dhabi and Dubai (UAE), Sulaymaniyah and Kirkuk (Iraqi Kurdistan), and Beirut (Lebanon), respectively. The participants were mainly in their 20s, with the mean age being 22.2 for Emirati Arabic, 22.5 for Iraqi Kurdish, and 28.6 for Lebanese Armenian. All 18 EA participants identified Arabic as their first language (L1) and English as their second language (L2). Among them, three individuals stated that they spoke fluent French, while one participant mentioned proficiency in Korean, and another in Spanish. Regarding educational backgrounds, the language of instruction varied during elementary and secondary schooling, with some receiving education in Arabic and others in English. However, at the tertiary level, all participants were educated in English, except for one individual who pursued studies in French. All Iraqi Kurdish speakers reported this language to be their L1. Nine of them mentioned to be able to speak Arabic as well. All the other participants indicated Kurdish as their preferred language of speaking, reading, and writing. This is not surprising as the language of education in this autonomous region is Kurdish, and the use of Arabic has declined there over the last two decades (Procházka, 2019). Four participants were educated in Arabic, besides Kurdish. Finally, for Lebanese Armenian, 19 of 21 speakers reported Arabic as their L2, while the two of them as L1 alongside Armenian.

The participants were recruited through the authors’ personal networks and local contacts in respective countries, and all participants were paid an equivalent of 15 CAD for their participation.

The materials were real words with the fricatives /χ, ʁ, (ħ, ʕ,) h/ appearing in initial, medial and (when phonotactically possible) final positions. The target consonants were immediately adjacent to a low vowel. The full list of the stimuli is presented in Table 2. The word lists for each language were prepared by the first three authors who are native speakers of the respective languages (albeit the first and the second authors being speakers of different dialects of Kurdish and Arabic). For Iraqi Kurdish, this was done in consultation with a native language informant. Recall that Lebanese Armenian lacks pharyngeal fricatives, while Iraqi Kurdish does not permit /h/ in final position. This gave us 15 words for Emirati Arabic, 14 for Iraqi Kurdish, and 9 for Lebanese Armenian.3

The target words were embedded in a carrier phrase, provided for each language in Table 3. In all cases care was taken to select meaningful phrases where the target word receives primary focus. It was not possible, however, to control for the length of the phrases and some variation in prosodic patterns across the languages.4 In terms of phonetic contexts, target words in each case were preceded and followed by a bilabial stop (/b/, /p/, or /pʰ/) in order to minimize lingual coarticulation with fricatives. The words were presented in the respective language script.5

The sentences were randomized, and each sentence was presented three times to each participant. Audio recordings were performed using an online experiment platform, Gorilla (Anwyl-Irvine, Massonnié, Flitton, Kirkham, & Evershed, 2020). A sample Gorilla display with a stimulus is presented in Figure 2. The total number of collected tokens was 24246 (1032 for Emirati Arabic, 828 for Iraqi Kurdish, and 564 for Lebanese Armenian).

Participants were allowed to use a recording device of their choice. The experiment records showed that 24 participants used a computer (Emirati Arabic: 13 participants; Iraqi Kurdish: 3 participants; Lebanese Armenian: 8 participants) and 35 used a cell phone (Emirati Arabic: 5 participants; Iraqi Kurdish: 17 participants; Lebanese Armenian: 13 participants). The resulting sound files were saved as mp3 files, separately for each utterance.

The data were manually annotated in Praat (Boersma & Weenink, 2023), with boundaries set to indicate onsets and offsets of fricatives and adjacent vowels. A fricative onset was taken to be a clear appearance of random frication noise above 1000 Hz based on a spectrogram, with reference to the waveform; a fricative offset was taken to be the cessation of such noise. The voiced pharyngeal produced by Emirati Arabic and Iraqi Kurdish speakers, however, frequently lacked fricative noise, being realized as an approximant (91% of tokens for Emirati Arabic and 88% for Iraqi Kurdish; cf. Khattab et al., 2018). In this case, the consonant boundaries were estimated based on dips in intensity observed in the waveform. Sample annotations for the voiceless uvular in each language are shown in Figure 3. (Examples of annotations for all fricatives are given in the Appendix.)

As mentioned above, one downside of collecting data online and having no control over speakers’ choice of recording devices is that data would vary with respect to the frequency range of individual recordings. Our examination of all recordings revealed that the upper cut-off of frequency was between 7500 Hz and 15000 Hz (7500–15000 Hz for Emirati Arabic, 7850–15000 Hz for Iraqi Kurdish, and 7600–15000 Hz for Lebanese Armenian), with the average being 11411 Hz (12007 Hz for Emirati Arabic, 12099 Hz for Iraqi Kurdish, and 10126 Hz for Lebanese Armenian). This is in contrast to conventional in-person recordings (in the lab or in the field) using the upper limit of at least 22100 Hz (given the sampling rate of 44200 Hz). However, since the focus of this study is on fricatives characterized by concentrations of energy in low and mid frequencies, having less reliable higher frequency information was not considered to be a major drawback. Nevertheless, we decided to make the frequency range uniform for all speakers by filtering out the signal above 7500 Hz. (A comparison of filtered and unfiltered high frequency fricative noise data is further discussed in Section 4.3). We also filtered out the signal below 1000 Hz for spectral moments measurements (see below) in order to reduce their dependence on formants (cf. Lorenc, Żygis, Mik, Pape, & Sóskuthy, 2023), which can be relatively strong for voiced fricatives and /h/.

Measurements were automatically extracted from the middle 20% of each annotated fricative and included four spectral moments (using a pass Hann band 1000–7500 Hz; see above):

Two other measurements were performed: relative intensity (in dB) and fricative duration (in sec). The former measurement was calculated as a difference between the intensity of adjacent vowels (preceding, following, or both) and the intensity of the fricative, with both taken at the mid 20% of the segment duration. Overall, all six measurements were performed on all annotated fricative tokens (n = 2424; see above), giving a total of 14544 data points for the analysis.

Data were analyzed using two kinds of linear mixed effects regression (LMER) models, performed using the lme4 package (Bates, Mächler, Bolker, & Walker, 2015) for R (R Core Team, 2022). The first set of models examined within-language differences in fricatives and included the fixed factors Consonant (five levels for Emirati Arabic and Iraqi Kurdish: /χ/, /ʁ/, /ħ/, /ʕ/, /h/; three levels for Lebanese Armenian: /χ/, /ʁ/, /h/), Position (initial, medial, final), and Gender (female, male). No interactions of Consonant with other factors were included, given the above-mentioned phonotactic gaps (in the case of Position) and in order to keep the outputs reasonably interpretable (in the case of Gender). The factors Consonant and Position were crucially related to our research questions. This is not the case for Gender, however, given the lack of clarity about its effects on guttural fricative noise (Gordon et al., 2002). We nevertheless decided to include this factor for exploratory reasons. Speaker was set as a random effect in these models, with a random intercept and a random slope by Consonant. Each model was performed separately for each variable (and separately by language), with a sample model being lmer(COG ~ Consonant + Position + Gender + (1 | Speaker), data_ea).7 In each case, likelihood ratio tests were used to compare a full model to a nested model excluding the factor of interest, employing the Anova function of the lmerTest package (Kuznetsova, Brockhoff, & Christensen, 2017). Posthoc tests with a Bonferroni correction for multiple comparisons were performed for Consonant and Position using the phia package (De Rosario-Martinez, 2015). In pairwise Consonant comparisons, our focus will be on place contrasts for fricatives of the same voicing and voicing contrasts for fricatives of the same place (e.g. /χ/ vs. /ħ/ and /χ/ vs. /ʁ/, leaving out contrasts like /χ/ vs. /ʕ/).

The second set of models was performed separately for each place category (uvular, pharyngeal, and laryngeal /h/), comparing their realizations across the languages. The fixed factor was Language (three levels for uvulars and /h/: Emirati Arabic, Iraqi Kurdish, Lebanese Armenian; two levels for pharyngeals: Emirati Arabic, Iraqi Kurdish). Random factors were Voicing, Position, and Speaker –– Voicing and Position were included as random factors given our primary focus at this stage on language differences. Again, each model was performed separately for each variable (and separately by place category), with a sample model being lmer (COG ~ language + (1 | voicing) + (1 | position) + (1 | speaker), data_uvular). Results of all LMER models were visualized using the ggplot2 package (Wickham, 2011).

We further used random forest analyses (Breiman, 2001) to determine the relative importance of acoustic variables for each language. This was done using the package party (Strobl, Malley, & Tutz, 2009) in R (R Core Team, 2022). The random forests method produces large sets (‘forests’) of conditional inference trees seeking to predict which variables (adjusted for collinearity) are most probable in accounting for the data (see Tagliamonte & Baayen, 2012 for a detailed description). Each conditional inference tree provides estimates of the likelihood of the value of the response variable (e.g., uvular vs. laryngeal) based on a series of binary questions about the values of predictor variables (e.g., COG, SD, duration). Predictions for each tree are made based on a random subset of the data and are further generalized to a final tree. An example of such a tree is shown in Figure 4, based on a subset of our data (Lebanese Armenian /χ/ vs. /h/). It shows that the algorithm recursively splits the data into clusters finding SD, duration, and – to a lesser extent – kurtosis (based on p-values) as useful predictors for distinguishing the contrast. A sample full model for a random forest analysis is la.cforest = cforest(C ~ COG + SD + skewness + kurtosis + rel_intensity + duration, data = data_la). The random forests analyses were run on the same data as for LMER models (a total of 14544 tokens).

Figure 5 presents Fast Fourier Transform (FFT) spectra of representative tokens for each fricative by language (measured around the midpoint). Beginning with voiceless fricatives, we can see that the uvular /χ/ in all three languages shows a very prominent low-frequency peak, as well as lesser defined concentrations of energy at higher frequencies. The voiceless pharyngeal /ħ/ in Emirati Arabic and Iraqi Kurdish is characterized by two peaks of similar intensity towards lower frequency, with some energy at medium frequency and a drop-off at higher frequency. The spectra for the laryngeal /h/ in Emirati Arabic and Lebanese Armenian are generally flat, apart from the intensity boost at the lowest frequency. The /h/ in Iraqi Kurdish, however, shows more prominence in low-to-medium frequencies. Voiced fricatives show overall similar patterns to their voiceless counterparts. Of note is a more abrupt decrease in energy at higher frequencies for the pharyngeal /ʕ/ (compared to /ħ/). How these spectral differences are manifested in terms of spectral moments across the full dataset will be examined in the next sections.

LMER models with the fixed factor Language and random factors Voicing, Position, and Speaker (with random slopes by Consonant) were performed separately for each consonant place, separately for each of the six variables. The analyses for uvulars (combined /χ/ and /ʁ/) and /h/ were based on three languages, while the analyses for pharyngeals were based on two languages (Emirati Arabic and Iraqi Kurdish).

A summary of the results is provided in Table 7. The results for uvulars showed no significant Language differences apart from relative intensity (p < 0.05): fricatives in Lebanese Armenian were less loud (compared to the adjacent vowels) than their counterparts in the other two languages (both p < 0.05). It should be noted that although COG differences were not significant, values for voiceless uvulars in Emirati Arabic were, on average, higher than for the other languages (by about 150–200 Hz). The results for pharyngeals showed significant effects of Language for COG (p < 0.05) and Duration (p < 0.001): Emirati Arabic speakers produced these fricatives with a higher COG and longer duration than Iraqi Kurdish speakers. Finally, the results for the laryngeal /h/ showed significant effects of COG (p < 0.05), skewness (p < 0.001), kurtosis (p < 0.01), and relative intensity (p < 0.001). This fricative produced by Lebanese Armenian speakers showed lower COG (both p < 0.001) and relative intensity (p < 0.05–0.001), as well a higher skewness (both p < 0.001) and kurtosis (both p < 0.001) than /h/ produced by Emirati Arabic and Iraqi Kurdish speakers. There were no other significant effects.

In summary, most language differences were found in the realization of the laryngeal /h/ and juxtaposed Lebanese Armenian speakers with the Emirati Arabic and Iraqi Kurdish speakers. There were hardly any language differences in the realization of uvulars, apart from the lower intensity of these consonants for Lebanese Armenian speakers. Pharyngeals in Emirati Arabic and Iraqi Kurdish were also relatively similar, yet with some differences in COG and duration.11

Random forests analyses were performed separately for each language, as well as across the three languages. The results of variable importance are illustrated in Figure 9. We can see that the most important (i.e., most informative, capable of distinguishing most contrasts) variable for Emirati Arabic was Standard Deviation (SD), followed by duration (sec) and then COG (Hz). The other variables – relative intensity, skewness, and kurtosis – were considerably less important. For Iraqi Kurdish, duration was most important, followed by SD, and at a considerable distance by skewness and the other variables. For Lebanese Armenian, duration and SD were also the top two variables, followed at a distance by COG and the other two variables. For the entire dataset (where Emirati Arabic tokens were most numerous), SD and duration were the top two predictors of the fricative contrasts, followed – in reducing importance – by COG, relative intensity, skewness, and kurtosis.

It should be noted that these analyses were performed to predict consonant categories, some of which involved place and some voicing. Running a random forests analysis for place categories (on the full data set) showed a greater importance of COG and relative intensity (while still distant in importance from SD) compared to duration. A random forests analysis for voicing, on the other hand, showed a dominance of duration, followed distantly by COG, SD, and then the other variables. Thus, overall, we can conclude that SD, duration, and COG were most crucial to distinguishing fricatives across three languages, with SD and COG being more important for place, and duration being more important for voicing.

The main goal of this study was to provide a cross-language investigation of guttural fricatives across different positions in three languages – Emirati Arabic, Iraqi Central Kurdish, and Lebanese Western Armenian – in order to contribute to a better understanding of the cross-linguistic typology and phonological patterning of these sounds. A secondary goal was to assess the validity of the online data collection for the acoustic analysis of fricatives. Our analysis of fricative contrasts involved measurements of four spectral moments, relative intensity, and duration, evaluated statistically using linear mixed effects regression models and random forests.

The results revealed that all place and voicing contrasts in fricatives across the languages were reliably differentiated by the selected acoustic variables. In terms of place, the uvular-pharyngeal contrast was consistently distinguished by COG, SD, and relative intensity, as well as partly by skewness (for Emirati Arabic and for voiced fricatives in Iraqi Kurdish), kurtosis (for voiced fricatives), and duration (for voiceless fricatives in Iraqi Kurdish). The uvular-laryngeal contrast was consistently distinguished by COG, SD, and duration, as well as partly by skewness (for Emirati Arabic and Lebanese Armenian), and kurtosis (for Lebanese Armenian). The pharyngeal-laryngeal contrast was consistently distinguished by SD and duration.

In terms of voicing, this contrast in uvulars was consistently distinguished by duration, and partly by COG, SD, and skewness (all for Emirati Arabic and Lebanese Armenian). Voicing in pharyngeals was consistently distinguished by COG, kurtosis, skewness, relative intensity, and duration, as well as partly by SD (for Emirati Arabic). This can be at least partly attributed to the frequent approximant-like realization of /ʕ/.

Based on LMER models, the variables SD, COG, and duration were most informative for differentiating fricative contrasts, while relative intensity, kurtosis, and skewness were considerably less informative. This was consistent with our evaluation of acoustic variables using random forests: COG, duration, and particularly SD were most successful at predicting consonant categories. Interestingly, SD did not differentiate the languages from each other (while the other variables did, at least for some contrasts). This suggests that SD is the most suitable variable for characterizing the examined fricative contrasts.

With respect to differences among the languages, most of those involved /h/ in Lebanese Armenian, which unlike Emirati Arabic and Iraqi Kurdish, was produced with a lower COG and relative intensity and a higher skewness and kurtosis. Lower intensity was also observed for the Lebanese Armenian uvulars. The source of these differences is unclear, but perhaps they reflect a weaker, more vocalic realization of /h/ and a more posterior location of the uvulars in Lebanese Armenian. Remarkably, the way uvulars and pharyngeals were produced in Emirati Arabic and Iraqi Kurdish was very similar, at least based on our measurements. The only language differences were in COG and duration for pharyngeals, indicating a more posterior and longer constriction in Iraqi Kurdish.

Some positional differences were also found. Namely, word-medial fricatives in Emirati Arabic and Lebanese Armenian showed higher relative intensity, lower COG, and shorter duration than their counterparts in initial and/or final positions. Positional differences were minimal in Iraqi Kurdish, being limited to duration; they were most extensive in Lebanese Armenian, involving almost all phonetic variables. It should be noted that many of these differences are indicative of lenition of fricatives in medial or final positions, thus confirming the cross-linguistic tendencies in this process (e.g., Hualde, Lujanbio & Zubiri, 2010; Henriksen & Harper, 2016). Gender differences in the production of fricatives were rather inconsistent, being absent in Emirati Arabic, present for just two variables (skewness and duration) in Iraqi Kurdish, and manifested by almost all variables (except relative intensity and duration) in Lebanese Armenian. It appears that gender differences in the production of guttural fricatives are largely language-specific.

The results obtained in this study are remarkably similar to many previous findings for Arabic (see Section 1.1). Taking voiceless fricatives, for example, we see in Figure 10 that COG values reported in several previous studies of Arabic fricatives (Norlin, 1983; Al-Khairy, 2005; Abu-Al-Makarem, 2005; Al-Tamimi & Khattab, 2015) were lower for pharyngeals and laryngeals; the same was observed for a previous investigation of Eastern and Western Armenian by Tahtadjian (2023). The results obtained for Emirati Arabic, Iraqi Kurdish, and (in part) Lebanese Armenian in this study showed essentially the same effects. The lower COG for /h/ in our results is also consistent with the findings of Abu-Al-Makarem (2005). Note also that there is considerable agreement in mean values for the fricatives across the studies, despite the very different recording conditions and signal filtering procedures. Similarly, our results for general spectral patterns, SD, and duration, and to some extent those for skewness and kurtosis, are in line with previous studies of gutturals in Arabic and other languages (Nartey, 1982; Gordon et al., 2002; Hargus et al., 2021).

The current results, therefore, provide evidence for the overall similar acoustic properties of guttural consonants in three different languages – Emirati Arabic, Iraqi Central Kurdish, and Lebanese Western Armenian – thus contributing to the phonetic typology of fricative contrasts. Having said that, we cannot discount the possibility that the production of fricatives by our Iraqi Kurdish and Lebanese Armenian speakers had been influenced by local varieties of Arabic. This may be the case particularly for the latter group, given their extensive use of Lebanese Arabic. Further work is needed to tease apart possible language contact effects by examining the production of Iraqi Kurdish and Western Armenian fricatives vis-à-vis similar consonants in Iraqi and Lebanese Arabic, by both bilingual and monolingual speakers.

One outstanding question is whether the Emirati Arabic, Iraqi Kurdish, and Lebanese Armenian fricatives referred here as uvulars (/χ, ʁ/) are indeed produced at the uvula or the velum, as suggested in some previous works (see Section 1.4). Based on our spectral analyses, these fricatives (/χ, ʁ/) are comparable to uvulars in other languages (cf. Hargus et al., 2021); furthermore, uvular fricatives examined in this study have auditory properties similar to those in other languages. Yet, it is worth noting that differences between uvulars and velars were not always successfully established through spectral analysis (Gordon et al., 2002). As the next step, it would be useful to examine formant transitions for uvulars in the three languages and compare those to other uvular (the stop /q/) and velar sounds. Future work should also extend speech materials to other vowel contexts and multiple lexical items. In addition, more can be done for examining cross-language variation in guttural fricatives by combining acoustic analysis with ultrasound tongue imaging (cf. Tahtajian, 2023).

Our finding that guttural fricatives in three languages are largely similar phonetically runs against the fact that these sounds do not pattern the same way in each of these languages. As mentioned in Section 1.5, the Arabic /χ, ʁ, ħ, ʕ, h/ are part of a larger segmental class including the stops /q, ʔ/ and coronal pharyngealized consonants (Watson, 2007). As McCarthy (1994) shows, gutturals in Arabic (and Semitic in general) pattern together in multiple processes, including the lowering adjacent vowels, avoiding coda position, and exhibiting strong co-occurrence restrictions in roots (OCP) (cf. Frisch et al., 2004). This shared behavior is attributed to these segments being specified for the same place feature, [pharyngeal] or [RTR] (McCarthy, 1994; Shahin, 2003; Watson, 2007). While somewhat similar consonant-to-vowel assimilatory effects have been reported for Kurdish varieties (as vowel lowering and backing: Kahn, 1976; as a phonotactic preference: Barry, 2019), these involve pharyngeals only. Uvulars (which also include /q/) do not pattern with pharyngeals as a class, showing no assimilatory effects on vowels. Pharyngeals do, however, exhibit some affinity with laryngeals (/h, ʔ/), appearing with those in free variation under certain conditions (Kahn, 1976). In contrast to Kurdish uvulars (but similarly to uvulars in Arabic), these sounds in Armenian can lower and/or back adjacent vowels (as been reported for some dialects). This prompted Vaux (1998) to specify these consonants for the features [+back, -high, -ATR]. These sounds do not pattern together with the laryngeal /h/, which acts as placeless. Furthermore, there have been no reports of lexical co-occurrence restrictions involving gutturals in either Kurdish12 or Armenian, thus additionally pointing to the relative phonological dissimilarity of the respective places of articulation. Altogether, what this shows is that there is a mismatch between the phonetics and phonology of guttural sounds: Their relatively similar phonetic realizations across three languages correspond to rather different phonological sets of contrasts and behaviors. This, in turn, underscores the relative independence of phonetics and phonology, and particularly the indirect mappings between distinctive features and acoustic properties of sounds. This state of affairs, however, may not be surprising if we assume that features represent emergent language-particular generalizations over acoustic/articulatory sensory maps and patterns in the lexicon (cf. Mielke, 2008).

One clear advantage of the online data collection adopted in this study was being able to target speakers of lesser studied languages residing in their home countries. It would have been clearly difficult, if not impossible, to find dozens of speakers of the examined languages (and reasonably control for age, gender, and dialect) anywhere in North America. One disadvantage of this method, however, is the reduced quality and variability of recordings obtained online, as has been previously observed in a number of studies (Ge, et al., 2021; Freeman & Decker, 2021; Sanker et al., 2021).

As mentioned in Section 2.5, our speakers used their own devices to make recordings, with over half of the speakers making recordings on their cell phones. Given the different settings for different device types, frequency ranges for the recorded data were inevitably different. For our analysis, we chose to adopt the lowest frequency cut-off in our sample to all speakers’ data (7500 Hz). One question arises is how this could have affected the results, specifically for spectral moments. To examine this, we compared the currently used values for COG and SD for each fricative (separately for each language) to the corresponding original values of these moments prior to the upper limit frequency filtering. These results are summarized in Table 8 in the Appendix. As one would expect, both COG and SD averages were reduced by the filtering procedure. The COG reduction was on average 253 Hz for Emirati Arabic, 250 Hz for Iraqi Kurdish, and 168 Hz for Lebanese Armenian (where the frequency range was originally the lowest). In all cases, the reduction was greater for uvulars, which are in general characterized by a higher COG. The SD reduction was on average 434 Hz for Emirati Arabic, 497 Hz for Iraqi Kurdish, and 329 Hz for Lebanese Armenian. This reduction affected voiceless fricatives to a greater extent than voiced ones. Overall, this shows that the use of devices with lower frequency cut-off has a tangible effect on spectral moments measurements, and this effect can be different depending on the fricative’s place and voicing. Nevertheless, it is important to note that even with a reduced frequency range, we obtained significant results for most fricative contrasts, and these effects were similar to those reported in previous studies. One would expect an even greater differentiation of these sounds if we had selected participants using devices with the largest frequency range (0–15000 Hz). Future work should aim to ensure the use of higher frequency devices, particularly if the focus is on more anterior fricatives, such as sibilants.

To conclude, the increasing availability of online platforms has a clear potential to facilitate data collection from participants residing in diverse geographical locations, allowing for a broader representation of speakers and linguistic varieties. This study serves as an initial contribution to a more extensive cross-linguistic phonetic documentation of guttural fricatives by covering new languages/varieties and using relatively large speaker samples. This study also serves to confirm the validity of the online audio recording method for acoustic analysis, which has been increasingly used in phonetics since the pandemic of COVID-19.