2.1 Vowel inventories

Table 1 lists in alphabetical order a sample of 110 languages and dialects which have been reported to have full schwa and the language families those languages belong to. The third and fourth columns report the corresponding vowel inventories and stress patterns. The bibliographical references where the vowel systems have been taken from are reported in Appendix A. These languages may be grouped genetically (Austroasiatic, Austronesian, Indo-Aryan, Koreanic, Mongolic, Niger-Congo, Sino-Tibetan, Tai-Kadai, Turkic, Uralic) or areally (Amazonian/Tupí, Australian, N. American Indian, Papuan). In some cases subdivisions of language families are mentioned instead: Albanian, Caucasian, Celtic, Germanic, Iranian, Romance, Slavic (Indo-European); Bantu (Niger-Congo); Chadic, Semitic, Cushitic (Afro-Asiatic).

Table 1 Vowel systems and stress patterns for languages and dialects with full schwa. In specific languages vowels may be specified as long and/or short (Vː, V), as +ATR and —ATR (ATR = Advanced Tongue Root), and according to whether they occur in open or closed syllables. The stress patterns are coded (1) for fixed stress, (2) for unpredictable stress, (3) for movable and predictable stress, and (4a), (4b) and (4c) for the tonal, unknown stress pattern and rhythmic stress options. N. = northern, C. = central, S. = southern; op. syll. = open syllable; cl. syll. = closed syllable. In Paicĩ, Australian English, British English and Bulgarian, the symbol ɜ stands for schwa. See the body of the text and Appendix A for more information.

Vowel inventories include information about whether vowels: are phonologically short and long (V/V:), long (V:) or short (V) in languages exhibiting distinctive vowel length; are specified for a positive or negative ATR (Advanced Tongue Root) value in languages where schwa alternates with the low vowel harmonically (Nen, Degema, Ikposo); occur in open syllables with thus no consonant in coda position or else in syllables checked by at least one coda consonant (Chinese). In the vowel inventories, schwa may be transcribed occasionally as [⁜] (Paicĩ, English, Bulgarian) and nasal vowels are not included. Vowels enclosed within parentheses are infrequent (e.g. [ɔ] in Sanuma), appear mostly in loanwards (e.g. [쀆] occurs in English loanwards in Hindi and [o] in Persian loanwards in Taleshi), or else have allophonic status (e.g. [ɛ] and [ɔ] in Javanese, [i], [e], [o] and [u] in Sharwa).

In a few languages appearing in Table 1 the vowel of interest, which is represented as [ə], has also been transcribed as [ɤ] or, less so, as [Ʌ] in studies from the literature. This indicates that full schwa and mid back unrounded vowels may turn out to be very close acoustically presumably since they differ essentially regarding postdorsal constriction size and degree of tongue predorsum height, i.e. in principle, the front dorsum should occupy a lower position for a mid back vowel than for schwa. In some of these languages one or the other phonetic symbols have been used depending on scholar and even by the same scholar: Wa ([ə], Svantesson Reference Svantesson1993; [ɤ], Watkins Reference Watkins2002); Scottish Gaelic ([ə], Gillies Reference Gillies, Ball and Nicole2009: 236; [ɤ], Ladefoged et al. Reference Ladefoged, Ladefoged, Turk, Hind and Skilton1998); Romanian ([ə], Chitoran Reference Chitoran2001; [Ʌ], Renwick Reference Renwick2014); Chantyal ([ə], [Ʌ], Noonan Reference Noonan, Graham and Randy2003); Bulgarian ([ə], Sabev Reference Sabev2015 or [⁜], Andreeva et al. Reference Andreeva, Barry and Koreman2013; [ɤ], Ternes & Vladimirova-Buhtz Reference Ternes and Vladimirova-Buhtz1999); Thai ([ə], Abramson Reference Abramson1962, Apiluk 1996; [ɤ], Tingsabadh & Abramson Reference Tingsabadh and Abramson1999). In some other languages and dialects, i.e. Iaai, Nias, Paicĩ, Ngwe and Vöru and Mulgi Estonian, the vowel of interest, which is also often transcribed as [ɤ], has been or may be qualified as central or centralized based on its spectral characteristics (see references in Appendix A and formant frequency values in Table 2).

Table 2 F1 and F2 frequencies in Hz for stressed schwa in a subset of the languages listed in Table 1. When available, information is provided about whether formant frequencies belong to long or short vowels and the contextual environments where vowels occur. See the body of the text and Appendix B for more information.

Table 1 includes a few more languages in which the vowel under study may be assigned a phonetic symbol other than [ə]: the presence of [ɤ] in Baarin and Daghur is in contrast with the use of the symbol [ə] for the close Mongolic languages Bonan, Chakhar and Buriad (see Svantesson et al. Reference Svantesson, Tsendina, Karlsson and Franzén2005: 146 and the corresponding references in Appendix A); in Awngi the only central vowel available may be realized as [ɨ] or [ə] depending mostly on syllabic factors (Joswig Reference Joswig2010: 19), while in Margi and Makaa [ə] may correspond to a high central vowel too (Heath & Heath Reference Heath and Heath1982, Maddieson Reference Maddieson1987); in European Portuguese the two central vowels are often represented as [ɨ] and [ɐ] in spite of the fact that the formant frequency values for the latter may be more appropriate for a mid central vowel (see Table 2), and a lowered variant of schwa is also found in Kurdish (McCarus Reference McCarus, Kaye and Daniels1997) and Digor Ossetic (Testen Reference Testen, Kaye and Daniels1997).

We have not included in Table 1 those languages where the vowel of interest is transcribed regularly as [ɤ] or [Ʌ] and shows formant frequencies which are most appropriate for a mid unrounded vowel. This is so for Vietnamese ([ɤ]; Kirby Reference Kirby2011, Emerich Reference Emerich2012), Standard Estonian ([ɤ]; Eek & Meister Reference Eek and Meister1998, Asu & Teras Reference Asu and Teras2009) and Nepali ([Ʌ]; Lohagun Reference Lohagun2018) even though in all these languages the vowel may show eventually a central realization in running speech (see Khatiwada Reference Khatiwada2009 regarding Nepali). We have also not taken into consideration those languages in which the only central vowel is [ɨ] since this vowel may not only be central but also fairly front, i.e. its F2 may rise up to 1859 Hz in Anong and 1950 Hz in Sakurabiat, and is somewhat higher than expected, i.e. its F1 ranges between 350 Hz and 450 Hz across languages (Becker-Kristal’s corpus; see Appendix B). A similar lack of homogeneity regarding F2 also applies to other central vowels such as [ɵ]. In particular, we have not included Cantonese where, among the short vowels, [ɐ] has formant values similar to those for [a], and [ɵ] exhibits an F1 about 550 Hz and an F2 about 1200 Hz (Zee Reference Zee, Maria Josep, Daniel and Joaquín2003), and Chuvash due to the uncertain spectral nature of its schwa-like lax vowels commonly represented as ĕ and ă (Clark Reference Clark, Lars and Csató1998: 434–435).

The vowel inventories presented in Table 1 reveal that, in addition to schwa, languages may show other central vowels in a stressed or prominent position. Schwa may co-occur mostly with high [ɨ] (Hup, Sanuma, Battambang Khmer, Car Nicobarese, Madurese, Sundanese, N. Welsh, Sharwa, Kashmiri, Cheju, Mono, Sepik, Yessan-Mayo, European Portuguese, Romanian, Amharic, Tangut, Thai, Selkup, Udmurt), and much less often with [ʉ] (Ao), [ɵ] (Bonan) and [ɐ] (Kabardian, Ngwe, Cangnan Min Chinese, Shangai Chinese). Occasionally there are two central vowels in addition to schwa: [ɨ ɘ] (Apinajé, Paicĩ), [ʉ ɐ] (Australian English, Pumi), [ʉ ɵ] (Tatar), [ɨ ɐ] (Nganasan). The co-occurrence of schwa with back unrounded vowels is also worth mentioning in view of the articulatory and acoustic similarity involved when the back unrounded vowels are also mid. Schwa may co-occur with one or more back unrounded vowels, and either no other central vowel ([ɯ], Wa, Scottish Gaelic, Kurdish, Akha; [ɤ], Standard Chinese; [Ʌ], British English RP; [ɯ Ʌ], Kensiu, Korean; [ɯ ɤ Ʌ], N. Khmer) or one or more central vowels ([ɨ ɤ], Car Nicobarese, Madurese; [ɨ Ʌ], Cheju; [ɯ ɐ], Cangnan Min Chinese; [ɤ ɐ], Shangai Chinese; [ɨ ɘ Ʌ], Apinajé; [ɯ Ʌ ɐ], Ngwe).

Overall, without considering the vowels appearing within parentheses in Table 1, full schwa is found most frequently in languages with 7 vowels (19) and with 6, 8 and 9 vowels (17 languages for each condition), followed by languages with 4 and 10 vowels (11 languages for each condition). This appears to be in line with the frequency of occurrence of optimal vowel systems and with the prediction that the mid central vowel should appear in vowel inventories with at least six vowels (Schwartz et al. Reference Schwartz, Boë, Vallée and Abry1997a).

2.2 Stress patterns

In order to verify that schwa occurs in stressed or prominent syllables, the languages listed in Table 1 have been grouped according to the following four stress patterns (see fourth column): languages with fixed stress (identification number 1), languages with lexically conditioned stress (2), languages with predictable but movable stress (3), and tone or register-tone languages (4a), languages with an unclear stress pattern (4b) and rhythmic languages (4c). The stress patterns for a considerable number of languages are based on the information provided by van der Hulst, Goedemans and van Zanten (Reference Hulst, Goedemans and van Zanten2010) and van der Hulst, Hendriks and van de Weijer (2008) and the references listed in Appendix A, and occasionally by other references not mentioned in the appendix. It needs to be understood that these patterns may reflect reality in a simplified way; indeed, stress placement often depends on several factors, and scholars may hold different views on the stress patterns available in the individual languages.

Identifying stressed schwa in languages with fixed or lexical stress (scenarios 1 and 2) is not problematic. In the case of scenario 1, stress always falls on the same syllable, most typically on the initial or final syllable of the word or the stem, as in Chechen and Taleshi Persian, respectively. Regarding scenario 2, stress placement varies with lexical item, and its identification may be retrieved from descriptive studies and dictionaries. Scenarios 3 and 4 are harder to account for.

As to the languages with predictable but movable stress (3), the stress position can be predicted though it is not fixed but dynamic (and thus falls on different word syllables) and conditioned by syllable weight (Gordon Reference Gordon2006). According to the syllable weight criterion, syllables may be heavy or light depending on vowel quality (i.e. on whether the vowel nucleus is lower or higher or else peripheral or non-peripheral), on vowel length (i.e. on whether the vowel is long or short, as in some Modern South Arabian languages; Lonnet & Simeone-Semelle Reference Lonnet, Simeone-Semelle, Kaye and Daniels1997) and on syllable type (i.e. on whether the vowel occurs in closed or open syllables, as in Kabardian; Gordon & Applebaum Reference Gordon and Applebaum2010). As explained in some detail in Section 5, in a considerable number of languages of class (3) schwa counts as short and takes stress only in special circumstances. A word needs to be said about Slovene which, in spite of being a pitch-accent language, is classified as a stress language exhibiting pattern (3) in Table 1. The reason for assigning pattern (3) to Slovene is that its vowels are considered to be stressed or unstressed (Petek et al. Reference Petek, Šuštaršič and Komar1996, Jurgec Reference Jurgec2005). Only when stressed, syllables are assigned an acute or a circumflex tone; more specifically, since vowel length is also distinctive in the language, stressed vowels are classified as long, with either the circumflex or acute tone, or short (Jurgec Reference Jurgec2005). Other pitch-accent languages such as Serbo-Croatian and Swedish are not included because they have no full schwa in their vowel system inventories.

Table 1 also reports tone languages (4a) and languages which have rhythmic stress or stress in alternating syllables (4c). Group (4b) includes Upper Bal Svan, Cheju and Upper Chehalis, whose stress status is largely unknown mainly because of the absence of relevant research on the subject. In languages of these groups, the availability of schwa in a prominent syllable may be inferred from its presence in monosyllables endowed with full semantic content or in longer words whose syllables have schwa as the only vowel nucleus, or else from its participation in the short/long vowel contrast (i.e. /ə/–/ə;/). Overall, the languages with patterns 4a, 4b and 4c represent about a 30% of the total number of languages of Table 1, and justify the use of the term ‘full schwa’ instead of ‘stressed schwa’ throughout the present study.

A note needs to be added about languages whose stress pattern has been assigned a question mark or else exhibit more than one stress pattern in Table 1. Some scholars believe that languages of the Indo-Aryan family have stress rules, albeit complex, which often depend on syllable weight (see for example Misra Reference Misra1967: 17–18 for Hindi; and Shackle Reference Shackle1972 for Punjabi), while others think that, if available, stress position cannot be determined with certainty (Ohala Reference Ohala1999); in particular, some experimental data for Gujarati have shown modest acoustic evidence in favor of fixed word initial stress, while the subjects put to test turned out not to have consistent stress intuitions (Bowers, Reference Bowers2019). As to the Mongolic languages, most but not all scholars consider that stress is found systematically on the first syllable of the word (see Svantesson et al. Reference Svantesson, Tsendina, Karlsson and Franzén2005: 96–97 for a summary). Another relevant case is that of Mandarin Chinese which, differently from other tone languages of the Sino-Tibetan family, appears to be endowed with inherent lexical stress, though the exact nature of this prosodic characteristic is unclear (Peng et al. Reference Peng, Chan, Tseng, Huang, Lee, Beckman and Sun-Ah2005). Finally, in the case of the Sepik language family, the stress pattern appears to vary between fixed (1) and variable (3) depending on the language taken into consideration (van der Hulst et al. Reference Hulst, Goedemans and van Zanten2010: 735–736).

3.1 General spectral characteristics

In order to account for the spectral implementation of full schwa, we present in Table 2 data on the first (F1) and second (F2) formant frequency values for this vowel. Appendix B reports the bibliographical references where the formant values presented in Table 2 have been taken from, as well as the number and gender of the speakers who took part in the acoustic recordings. Data correspond to a single speaker or to averages across speakers of a subset of the languages listed in Table 1, and are coded with the same language number. All language groups are represented except for Amazonian, Australian, Chadic and Papuan. Data for a few languages are lacking for specific reasons; thus, for example, no formant frequency data have been found for /ə;/ in Korean, and /ə/ in Arrernte is endowed with a large spectral variability (Tabain & Breen Reference Tabain and Breen2011). Whenever possible, Table 2 provides information about whether formant frequencies correspond to long or short schwa in vowel systems where vowel length is distinctive, and about other contextual and lexical details. If preceded by ‘ca.’, the vowel formant frequencies have been inferred from F1 × F2 plots and therefore are approximate (Taleshi Persian, Cheju, Cangnan Min Chinese). The formant frequency values were obtained from words produced in isolation or in a carrier sentence and, less often, from stretches of continuous speech. They belong to male subjects since, for the most part, the monographs listed in Appendix B report data for male speakers only; in the case of Taleshi Persian and Bulgarian, however, the formant frequency values have been averaged across genders given that, as indicated in the appendix, the number of male and female speakers was relatively large in this case.

Figure 1 displays all F1 × F2 frequency values appearing in Table 2. The mean formant frequencies occur where the two discontinuous lines overlaid on the data points cross, and the corresponding standard deviation values equal the length of these lines. According to Figure 1, the F1 frequency ranges conform to the initial expectation in that they take place between about 400 Hz and 600 Hz, and thus happen to be centered around a frequency of 500 Hz (see Section 1). The mean F1 value across the data sample is 511.6 Hz and the standard deviation 81.5 Hz. As to F2, most values lie between 1300 Hz and 1500 Hz and are therefore found somewhat below the expected frequency of 1500 Hz, though still far apart from the 1200–1300 Hz range for mid back unrounded [ɤ] and [Ʌ]. The mean F2 amounts to 1428.2 Hz and the corresponding standard deviation to 128.5 Hz.

Figure 1 F1 × F2 frequencies in Hz for full schwa presented in Table 2. The mean F1 and F2 values and the corresponding standard deviations are indicated by the discontinuous lines.

3.2 Spectral change as a function of duration-related factors

It was hypothesized in the Section 1 that there could be a relationship between the spectral characteristics of full schwa and vowel duration in languages where vowel length is distinctive. If so, in comparison to long schwa, short schwa should exhibit a higher F2 and a lower F1, and thus a somewhat more centralized and thus less posterior schwa realization. This hypothesis was not confirmed by the formant frequency data for languages with short and long vowels (see Table 2). Only Battambang Khmer shows, as expected, a higher F2 for short vs. long schwa, though F1 is not higher for the long vs. short cognate, but the reverse. Essentially, small or no differences in F1 or F2 were found to occur between long and short schwa in N. Khmer, Chakhar, Thai and Vöru Estonian, and, contrary to the initial hypothesis, F2 happened to be higher for long schwa than for short schwa in Ndut.

Spectral differences as a function of vowel duration may also be associated with word and syllable position: vowels are expected to be shorter word medially than word initially and finally, and in closed vs. open syllables. Word position may account for why there is schwa instead of [ɤ] in non-initial syllables in the Inner Mongolic languages Baarin, Buriad, Chakhar and Daghur (Svantesson et al. Reference Svantesson, Tsendina, Karlsson and Franzén2005: 157), and why final lengthening prevents stressed schwa from occurring word finally in several languages: Welsh (Awbery Reference Awbery, Ball, Nicole and Jones1984); Hindi (Misra Reference Misra1967: 17); Malay (Clynes & Deterding Reference Clynes and Deterding2011); Balinese and Indonesian, where word syllable-final schwa is replaced by [ɨ] in open syllables (Ward Reference Ward1973: 28); Besemah and Moloko, where word syllable-final schwa becomes [a] in closed syllables and across the board, respectively (Gordon et al. Reference Gordon, Ghushchyan, McDonnell, Rosenblum, Shaw and Steve2012, Gravina Reference Gravina2014: 101–102). Differences in phonetic realization as a function of syllable type occur in Mandarin or Standard Chinese where the longer and more constricted vowel [ɤ], with a lower F1 and F2 frequency, is found in open syllables, and the shorter and less constricted vowel [ə], with higher formant frequencies more appropriate for a mid central vowel, occurs in syllables checked mostly by a nasal consonant (see Table 2, and Svantesson Reference Svantesson1984, Lee & Zee Reference Lee and Zee2003). As to Shangainese, F2 data in Table 2 indicate, however, that both /ɤ/ and /ə/ may be quite central in open syllables, and F1 data for /ə/ that it could be transcribed as [ɐ] as acknowledged in the literature (Chen Reference Chen2008).

In agreement with our initial predictions (see Section 1), a trend has been observed for stressed schwa to be not only longer and more prominent but also more posterior than unstressed schwa, and therefore to exhibit a higher F1 and a lower F2. In other words, if the back of the tongue approaches the mid or low pharyngeal wall to a greater extent for stressed schwa than for unstressed schwa, then one would expect the former to resemble /ɔ/ or /a/ more than the latter. Indeed, this difference holds in Cheremis and Romanian according to phrase-final data from Lehiste et al. (Reference Lehiste2005) and also to data from Renwick (Reference Renwick2014) plotted in Figure 2 (top graph). Also in Gheg Albanian, unstressed schwa has been reported to have an F2 frequency at about 1500 Hz and the more [ɒ]-like stressed cognate a lower F2 at about 1200–1400 Hz (6M; Granser & Moosmüller Reference Granser and Moosmüller2001). As revealed by Figure 2 (bottom graph), in Bulgarian, stressed schwa has a lower F2 but not necessarily a higher F1 than unstressed schwa according to data taken from Sabev (8F; 2015) and Andreeva et al. (8F, 12M; 2013). Available descriptions agree with these formant frequency values in pointing to the possible realization of schwa as [Ʌ] under stress and as [ə] otherwise in Nooksack (Galloway Reference Galloway, Eugene and Bill1983: 87), and as [a] in the stressed position in Arrernte (Breen & Dobson Reference Breen and Dobson2005).

Figure 2 F1 × F2 frequencies for unstressed schwa (empty symbols) and stressed schwa (filled symbols) in several languages. Top panel: Languages showing a higher F1 and a lower F2 for stressed vs. unstressed schwa. Bottom panel: Data for Bulgarian where stressed and unstressed schwa differ in F2 but not in F1. M = male speakers, F = female speakers.

3.3 Spectral variability

Another salient characteristic of stressed schwa is its considerable spectral variability. Acoustic variability for vowels may occur as a function of factors as diverse as contextual segment, position, stress, speech rate and speaker (see Chapters 3 and 4 of Hardcastle et al. Reference Hardcastle, John and Gibbon2010 for supporting evidence).

The effect of contextual variability is apparent in a study on Majorcan Catalan vowels produced by 5 male speakers. Stressed schwa was found to exhibit a high F2 frequency range with values between 1250 Hz next to labials and dark /l/ and the trill /r/ and 1500–1600 Hz next to other alveolars and to palatals, which turned out to be comparable to the F2 variability degree for /u/ and greater than that for the other stressed vowels produced in the same contextual conditions (Recasens & Espinosa Reference Recasens and Espinosa2006a). F2 frequency ranges for other languages have been plotted in Figure 3, and the corresponding bibliographical references where these data have been taken from are given in Appendix C, which also reports whether the vowel productions belong to male and/or female subjects and the number of subjects. The effect of context should be apparent in cases where the vowel was produced in different consonantal contexts (Paiwan, Cheju), in several words (Romanian, Mundurukú) or in long speech samples (Albanian, Javanese, Indonesian, Mono), while in cases where different tokens of a single word were recorded the vowel’s F2 variability should probably be attributed to speaker-dependent differences for the most part (Enganno, as well as Khonoma Angami, where the F2 data corresponds to a weighted average of F2 and F3 minus F1).

Figure 3 Approximate F2 frequency ranges in Hz for full schwa in languages and dialects where they exceed (top) or not (bottom) the F2 frequency ranges for all or most of the other system vowels. Top graph, from left to right: Albanian, Enganno, Iaai (F), Iaai (M), Javanese, Indonesian, Mono, Romanian (F), Romanian (M), Khonoma Angami (F), Khonomna Angami (M). Bottom graph, from left to right: Paiwan (F), Paiwan (M), N. Taleshi, C. Taleshi, S. Taleshi, Cheju, Mundurukú.

In the languages displayed in the top graph of Figure 3, the F2 frequency ranges vary between about 300 Hz and 1000 Hz, and exceed those for all or most of the other system vowels in the language. Large F2 ranges for Wa (Watkins Reference Watkins2002) and Madurese (Cohn & Lockwood Reference Cohn and Lockwood1994) have not been included because precise formant frequency data for these languages are not provided by the literature sources. In languages plotted in the bottom graph of Figure 3, on the other hand, a smaller F2 frequency spread generally around 300 Hz happens to be comparable to, and thus not larger than, that for the other vowels of the phonological system. This scenario is also at work for /ɤ/ in open syllables in Taiyuan Chinese (Xia & Hu Reference Xia and Hu2016) and Shangainese (Svantesson Reference Svantesson1989, Chen Reference Chen2008).

These spectral variability data are consistent with descriptions about the allophones and dialectal realizations of full schwa. Indeed, just to mention a few examples, this vowel has been reported to vary a great deal as a function of segmental context, mostly in anteriority degree, in Sharwa (Gravina Reference Gravina2014: 136); Gilaki (Rastorigueva et al. Reference Rastorigueva, Kerimova, Mamedzade, Pireiko and Edel’man2012: 8–9); Cheremis (Lehiste et al. Reference Lehiste2005: 87–89) and Moksha (Estill Reference Estill, Stefan and Tomi2010). Variability also involves the height dimension and thus F1 whenever the vowel realizations range from [Ʌ]/[ə] to [ɨ], as reported for Balinese (Ward Reference Ward1973: 28); Amharic (Iwatsuki Reference Iwatsuki2012) and Macedonian dialects (Friedman Reference Friedman2001: 10). Also in Makaa (Heath & Heath Reference Heath and Heath1982: 16); Margi (Maddieson Reference Maddieson1987) and Wadu Pumi (Daudey Reference Daudey2014: 50), [ɨ]-like realizations occur next to (alveolo)palatals and other consonants involving front tongue body raising, and [ə]-like realizations in other segmental contexts.

A high degree of articulatory and acoustic variability for the vowel under study is available in languages exhibiting a small number of vowel phonemes (see Table 1). This is the case for Arrernte, Kabardian and several Austronesian languages where the contextual allophones cover a large portion of the vowel space (Foley Reference Foley1986: 49–50; Colarusso Reference Colarusso1988, Choi Reference Choi1991, Breen & Dobson Reference Breen and Dobson2005), and also for several Salishan languages where the vowel has been reported to be realized as [ə], [ɪ], [ʊ], [Ʌ], [ɑ] and [ɔ] depending on consonantal context (Bessell Reference Bessell1998, Nolan Reference Nolan2017: 24).

The context-dependent articulatory and acoustic variability for full schwa may be restricted by the number of central vowels, as indicated by the relatively small degree of F2 variability for the vowel of interest in Paicĩ (Gordon & Maddieson Reference Gordon and Maddieson2004) and Cangnan Min Chinese (Hu & Ge Reference Hu and Ge2016). Moreover, as expected, F2 variability is less for full schwa than for reduced schwa in unstressed position in languages such as Lekwungen (Nolan Reference Nolan2017: 91) and Cheremis (Lehiste et al. Reference Lehiste2005: 71–81).

4.1 Position

This section investigates the extent to which, based on the formant frequency values listed in Table 2, the spectral characteristics of full schwa may change with vowel inventory size by looking at the placement of this vowel with respect to the mid peripheral vowels (excluding [쀆]) in the F1 × F2 space. This analysis has been carried out separately for those languages and dialects showing 2 mid vowels (one mid front unrounded vowel and one mid back rounded vowel), 3 mid vowels (two front unrounded vowels and one back rounded vowel, or vice versa) or 4 mid vowels (two mid front unrounded and two mid back rounded vowels). Overall, there were 16 languages and dialects with 2 mid vowels, 10 with 3 mid vowels, and 14 with 4 mid vowels. The languages with 2 mid vowels are coded with numbers 1, 16, 19, 21, 24, 29, 32, 36, 44, 51, 85, 86, 94, 98, 105 and 106 in Tables 1 and 2; they include Kabardian (29) which, in spite of having essentially 3 vowels, shows often [e;] and [o;] as surface realizations of underlying diphthongs (Gordon & Applebaum Reference Gordon and Applebaum2010). The languages with 3 mid vowels are coded 7, 15, 40, 43, 57, 62, 63, 65, 67 and 95; they include Babine (67) where [ɛ] occurs often after fortis consonants, and Wolof (65), for which no formant values are reported for short /o/ in Calvet (Reference Calvet1966). Finally, languages with 4 mid vowels are coded 8, 10, 22, 25, 31, 33, 47, 59, 61, 64, 79, 82, 101 and 102.

Several aspects about this vowel material need to be mentioned in the case of languages where vowel length is distinctive. The general criterion has been to compare schwa with peripheral vowels of equal length. Thus, in the case of languages where vowels, including schwa, may be short and long, the data subject to analysis correspond to the long vowels (Battabang Khmer, N. Khmer, Chakhar, Ndut, Estonian, Thai); however, due to restrictions imposed by the available data material, formant values were analyzed for the short vowels (Scots Gaelic) or for a mixture of long and short vowels (Iaai). As to those languages in which schwa was always short and the peripheral vowels could be long and/or short, schwa was compared with the short peripheral vowels (N. Welsh, Afrikaans, Wolof) with some exceptions: we took into account the data for long peripheral vowels whenever some relevant peripheral vowels happened to be only long (Slovene), and those for a mixture of long and short vowels whenever peripheral vowels differing in quantity also differed in quality (S. Welsh) or due to restrictions imposed by the literature reports (Eton). As to languages where schwa can occur only in closed syllables, the peripheral vowels correspond to the open syllable condition which is where mid vowels are found (Shangai Chinese).

Whenever possible, we have averaged data from different studies in order to achieve more representative formant frequency values after normalizing them according to the procedure described below. In some cases averaging has been carried out across male speakers’ productions, as reported by the studies enclosed within parentheses: Wa (Svantesson Reference Svantesson1993, Watkins Reference Watkins2002), Sundanese (van Zanten & van Heuven Reference Zanten and van Heuven1984, Yusuf Reference Yusuf1993), Afrikaans (Taylor & Uys Reference Taylor and Uys1988, Botha Reference Botha1996), Maithili (Yadav Reference Yadav1984, Jha Reference Jha2001), Romanian (Teodorescu Reference Teodorescu1985, Renwick Reference Renwick2014), Bulgarian (Lehiste & Popov Reference Lehiste and Popov1970, Tilkov Reference Tilkov1970, Andreeva et al. Reference Andreeva, Barry and Koreman2013), Vöru Estonian (Parve Reference Parve2000, Teras Reference Teras, Maria Josep, Daniel and Joaquín2003, Reference Teras2004). In four instances, we have averaged data for males (M) and those for females (F) not given in Table 2 and thus not plotted in Figure 1: Madurese (1M, 1F; Cohn & Lockwood Reference Cohn and Lockwood1994), Marathi (5M, 5F; Shinde et al. Reference Shinde, More, Sonawane, Pawade and Kulkarni2017), Shangainese (3M, 10F; Chen Reference Chen2008) and Cheremis (4M, 4F in Lehiste et al. Reference Lehiste2005, and 2F in Estill Reference Estill2012). Data for English RP male speakers have been taken from Wells (Reference Wells1963).

Formant frequencies values were normalized in order to be able to carry out the relevant comparisons, and given the fact that speakers differ in vocal tract size whether they belong to the same sex or, even more clearly, to different sexes. The normalization procedure was performed as follows: we first computed the F1 distances between each vowel and /a/ (i.e. the vowel with the highest F1) and the F2 distances between each vowel and /i/ (i.e. the vowel with the highest F2); then we calculated the percentages of each F1 distance over the maximal F1 distance available (that for /a/–/i/) and those of each F2 distance over the maximal possible F2 distance (that for /i/–/u/). Thus, for example, an F1 percentage of 75% for a given vowel means that F1 for this vowel is located at three quarters of the /a/–/i/ space below F1 for /a/, and thus at one quarter above F1 for /i/. On the other hand, an F2 percentage of 33% for a given vowel means that F2 for this vowel is closer to /i/ than to /u/, i.e. one third of the /i/–/u/ space away from /i/ and two thirds away from /u/.

Figure 4 shows the normalized F1 × F2 data for groups of languages with 2, 3 and 4 mid vowels. In each graph, the F1 × F2 frequencies for /i/, /u/ and /a/ (represented in large filled squares) occur at the upper left corner, upper right corner and F1 baseline, respectively. The data points inside the graphs correspond to mid front unrounded and mid back rounded vowels, and to full schwa. Analogously to Figure 1, the mean F1 × F2 values for schwa are located where the two discontinuous lines cross and the standard deviation for each formant is indicated by the length of the lines. The actual mean formant frequencies for schwa and the mid front unrounded and mid back rounded vowels, together with the corresponding standard deviations, are given in Table 3.

Figure 4 Position of stressed schwa and mid vowels in F1 × F2 space in languages with 2 (left), 3 (middle) and 4 (right) mid vowels. The cross-language mean F1 and F2 values and the corresponding standard deviations are indicated by the discontinuous lines. Y-axis: Percentages over the F1 distance between /a/ and /i/; X-axis: percentages over the F2 distance between /i/ and /u/.

Table 3 F1 and F2 distance percentages for stressed schwa, and for the mid front unrounded and back rounded vowels, relative to the (maximal) F1 distance between /a/ and /i/ and the (maximal) F2 distance between /i/ and /u/. Percentages and standard deviations are provided across languages for vowel systems with 2, 3 and 4 mid vowels.

Two main trends can be observed in Figure 4 regarding the location of full schwa in vowel space. First, there appears to be a relationship between the position of schwa along the F1 dimension and the number of mid vowels. In systems with 2 mid vowels (leftmost graph), schwa occupies a mid high position at about 60% of the /a/–/i/ distance, respectively. Moreover, also in this case the F1 for full schwa tends to be aligned with F1 for the mid peripheral vowels, which include /ɛ/ and/or /ɔ/ in addition to /e/ and /o/. Given that F1 is positively associated with degree of vowel opening, what this means is that schwa is articulated with a more open oral cavity in language systems with mid high and mid low vowels than in systems endowed with only mid high vowels.

F2 for full schwa, on the other hand, happens to be fairly equidistant from F2 for mid front and mid back vowels. This fairly equidistant relationship becomes apparent through inspection of the three discontinuous vertical lines located at the mean F2 for the 2, 3 and 4 vowel systems in the three graphs of Figure 4. As indicated in Table 3, schwa is found at about 57% of the F2 distance between /i/ and /u/ irrespective of whether vowel systems have 2, 3 or 4 mid vowels. Moreover, this percentage indicates a position somewhat closer to the area occupied by the mid back unrounded vowels than to that for the mid front unrounded vowels, which is in accordance with the absolute F2 values appearing in Figure 1. Posterior realizations of schwa, which correspond to the rightmost data points in the graphs of Figure 4, belong mostly to Austronesian and Indo-Aryan languages (Iaai, Javanese, Madurese, Sundanese, Gujarati, Maithili, Marathi, Urdu).

4.2 Variability

Judging from the standard deviations reported in Table 3, both F1 and F2 for full schwa appear to be affected by the number of peripheral vowels also regarding formant frequency dispersion.

The F1 frequencies for full schwa are more disperse in larger vs. smaller vowel systems, i.e. they spread over 10%, 16% and 17% of the distance between /a/ and /i/, respectively. In particular, the F1 spread occurs towards higher F1 frequencies. Thus, taking the 50% /a/–/i/ distance as reference, F1 occurs above this distance in 14 out of 16 languages in 2 mid vowel systems, and below this frequency in 10 out of 14 languages in 4 mid vowel systems. Languages with 3 mid vowels show an intermediate scenario, i.e. 5 languages above and 5 languages below the 50% /a/–/i/ distance in question. It may thus be concluded that the F1 dispersion for schwa keeps a direct relationship with F1 dispersion for the mid peripheral vowels, which is certainly greater in vowel systems with more vs. less mid vowels.

Regarding F2 variability, the standard deviation values in Table 3 indicate a decrease in the spread of the F2 frequency values for full schwa as the number of mid vowels in the language system increases, i.e. over 11% (2 mid vowels), 9% (3 mid vowels) and 6% (4 mid vowels) of the distance between /i/ and /u/. Such a decrease may be associated with the fact that mid front and mid back vowels lie somewhat further apart in 2 mid vowel systems than in systems with 3 and 4 mid vowels; indeed, the separation between the mean F2 averaged across mid back and mid front vowels is 77.7%, 72.51% and 71.83% of the /i/–/u/ distance in languages with 2, 3 and 4 mid vowels, respectively.

In sum, there appears to be a trend for F1 but not F2 for full schwa to increase with the number of mid peripheral vowels, and for the F1 and F2 variability for full schwa to increase with the mid vowels spread along the F1 and F2 dimensions, respectively. We will return to these findings in the Section 6.