The first terminological issue concerns the definition of liquid and rhotic consonants. These terms are known to be difficult to define, even though there seems to be a commonly shared intuition among phonologists over which type of segment belongs to these categories and which does not. This issue has been discussed by a number of authors (most notably Lindau 1985; Labrune 1993; 2014; 2017; Ladefoged & Maddieson 1996; Walsh Dickey 1997; Proctor 2009; Wiese 2011; etc. See also the collection of papers in this volume).

Building on these previous studies, this paper adopts a traditional and conventional definition of the term liquid. The class of liquids will be defined as including all the traditionally so-called “l”-like and “r”-like sounds. They correspond to the API symbols shown in Table 1 (adapted from Walsh Dickey 1997: 11, 14, but note that lateral affricates and the uvular voiceless fricative χ have been excluded because they pattern more like obstruents than like sonorants. Note also that ʁ is categorized as an approximant).

If we remove all lateral segments from the Table 1, we obtain the list of rhotics. In this paper, a rhotic will be defined as a non-lateral liquid segment which is proto-typically a tap, a flap, or a trill articulated in the dental, alveolar or pre-palatal region (i.e. coronal), including the retroflex place of articulation. The palatalized, velarized, glottalized, lateralized and voiceless versions of these segments (all rare in the language sample) will also be labeled as rhotics. Given this definition, any segment transcribed by means of the symbols r̪ ɾ̪ r ɾ ɺ ɺ̣ ɽ in the language descriptions used for this survey (as well as their counterparts involving a secondary articulation) will be automatically categorized as rhotic.

The two approximants ɹ and ɻ will also be included in the class. This choice may be disputable, but note that only three languages of the sample (Diola-Fogny, English and Shipiho-Konibo) have ɹ as the sole phonological rhotic of their system, and none has ɻ without having any other rhotic.

The case of the uvulars /ʀ/ and /ʁ/ deserves special attention. The uvular trill ʀ and the voiced approximant (or fricative) ʁ are reported to exist as phonemes in 12 languages of the sample. In this paper, the inclusion of these two segments in the rhotic class will be system-dependent, or conditional, which means that their categorization as rhotic or non-rhotic will depend on their status within the phonological system that hosts them and shall be established on a case-by-case examination. The general principle is that when a language has a phonemic uvular trill and/or a voiced uvular approximant in addition to one or several phonemic coronal rhotic(s), these uvulars have not been included in the sub-inventory of rhotics. This is the case of Abipon, Egyptian Arabic, Armenian, Ingush, Kayah Li and Nivkh. When a language has one or more than one uvular rhotic – /ʀ/ or /ʁ/ –, and no phonemic coronal rhotic, this/these uvular/s has/have been counted as (a) rhotic(s), except when there exists explicit and convincing evidence in the phonology of the language that /ʀ/ or /ʁ/ do not behave as sonorants but rather pattern with some other non-liquid segment. Following this criterium, the uvulars of French, German, Greenlandic (West), Hebrew and Yup’ik (Central) have been counted as rhotics. Armenian, on the other hand, is classified as having three liquid consonants, /l/, /r/ and /ɻ/, and two rhotics rather than three or even four (thus excluding /ʁ/ and /χ/ from the list of rhotics). This contrasts with French, which has a /ʁ/ and a /l/ but no phonemic coronal rhotic distinct from /ʁ/, so French is counted has having two liquids among which one is a rhotic (modern French /ʁ/ also has [r] as one of its allophones, and [r] is known to have been the original realization of what has become /ʁ/ – [ʁ], [ʀ], [χ], [r] etc. – in contemporary French. A similar development has occurred in German). Along the same reasoning, Lakhota is analysed as having only one liquid, /l/, because its /ʁ/, /χ/ and /χ’/ are not considered sonorants by Lakhota specialists and they pattern with the fricatives.1 Lakhota is thus considered as having only one liquid phoneme, /l/ and no rhotic. It should be noted that ʁ and ʀ are relatively rare as phonemes in the languages of the sample, be they considered as rhotics or not (12 languages2 have one of them or the two of them) so a different choice would have had no major impact on the overall results of this study. The exaggerated importance that ʀ and ʁ have been granted in the class of rhotics undoubtedly comes from their salience in the dominant languages French and German, but from the point of view of the present research, they appear as marginal elements.

Many languages have some sort of r and l sounds in free variation, or use the two articulations as allophones of a single phoneme. A close examination of the phonological and phonetic structure of the 200 languages under scrutiny has revealed that when a language has only one phonemic liquid (conditional or unconditional) segment – and this is the second most common pattern found in the sample (54 languages, see annex 1) – it sometimes happens that this unique liquid is transcribed as a lateral (we will delve into the caveats of transcription below in 3.3), even though it may have a genuinely rhotic allophone as in Korean, Luvale, Meithei, Nahuatl, Sanuma, for instance. In such cases, this unique lateral liquid has been labeled as rhotic. Similarly, a rhotic phoneme noted as /r/ for instance may frequently have a lateral allophone, especially in systems which have only one alveolar liquid.

A difficult case is when one or several lateral approximant symbols represent the only phonemic liquids within their system and are not explicitly reported as contrasting with a prototypical rhotic as defined above, nor as having a rhotic allophone.3 In such cases, the language has been considered as lacking a rhotic phoneme, although the possibility that the lateral phoneme also has an unreported rhotic allophone cannot be dismissed, especially in the case of under-described languages (which is what are almost all the languages which fall into this category). 15 languages in the sample correspond to this case: Araona, Dani (Lower Grand Valley), Imonda, Kiowa, Kongo, Koromfe, Lakhota, Mandarin, Miwok, Ndyuka, Oneida, Passamaquoddy-Maliseet, Pomo (Southeastern), Supyire, Vietnamese.

This being said, one should keep in mind that the most represented liquid system in the languages in the sample (72 languages, 36%) is quite expectedly a system that combines a coronal lateral /l/ and a coronal rhotic /r/, which is phonetically almost always a dental or an alveolar tap or trill.

The set of segments which will be recognized as rhotics for the present study are given in Table 2. Core, or unconditional rhotics, appear in bold.

Another term which requires clarification is that of “word” as used for instance in the recurring key expression of this paper “word-initial rhotic avoidance”. This term should be understood in its broadest acceptance. “Word” denotes a morpho-lexical unit, whose nature and status is obviously likely to vary across languages, and it is used in the context of the present study as a cover term for lexeme, wordform, morph, stem, root or base, depending on the languages and on the theoretical stands of the linguists who have provided the description and analyses on which I have relied.

A not uncommon case in the database is when a language accepts affixes or clitics beginning with a rhotic, but not full lexical words, as, for instance, Guarani (Dooley 2006), Japanese (Labrune 2014) or Kayardild (Round 2009). In such cases, “word” will mean “full lexeme”. In a language where no morphological unit, either autonomous or non-autonomous, accepts a rhotic in the initial position, the term “word” will have a broader acceptance and will have to be understood as “morph” or sometimes “stem”. Here again, it is impossible to achieve a perfectly satisfactory terminology, given the heterogeneity of the sources as well as specific properties of the various languages.

A related issue to that of “word” is what is meant by “initial position”. I take the term “initial” to mean the first phonological segment of a word, and the one which is most likely to be denoted in the orthography of the language – when the script is phonographic – or in the phonological transcriptions by linguists. But one should be cautious to note that this is not necessarily the form under which words are uttered at the phonetic level, either because the initial phonological rhotic may have a non-rhotic allophone word-initially, as in Awa Pit, for instance, or because it is preceded by a prothetic segment which is not consistently denoted in the orthography, as in Beja. This will lead us to distinguish two types of WIRA: phonological WIRA and phonetic WIRA. This distinction will be defined in Section 4 below.

Unfortunately, precise descriptions of both the phonological and phonetic nature of initial segments are not always available in the descriptions, especially in the case of under-described languages.

In this paper, the terms “avoidance”, or “restriction”, will be used in preference to other possible terms also found in the literature dealing with word-level phonotactics, such as “prohibition” or “absence”. Thus we shall speak of “word-initial avoidance” rather than of “word-initial prohibition” or “absence”. This is because, thorough progress into this research has revealed that the possibility for a rhotic segment to occur in the word-initial position cannot be easily reduced to a ‘yes’ or ‘no’ appreciation. In other words, we are dealing with a scalar phenomenon. It is thus more appropriate to consider things in terms of avoidance or restriction, rather than in categorical terms such as prohibition or absence. It is important to note that there are actually very few languages with literally zero (0, i.e. not even 1) words beginning with a rhotic in their entire lexicon, if periphery lexemes such as loanwords and onomatopoeic words are included (on the issue of loanwords and onomatopoeia, see Section 3.2 below). Another issue is that a given segment might be rare in the initial, but also in the non-initial (medial or final) position. Thus, a low number of words beginning with a rhotic should not necessarily be taken as revealing WIRA if the rhotic is also rare in other positions. An example can be found in Taba, in which /r/ beginning words are rare, but this phoneme is said altogether to be “relatively unfrequent” (Bowden 1997: 57). Because there does not seem to be any asymmetry in the occurrence of /r/ within Taba words, the apparent rarity of this consonant word-initially cannot be considered as revealing WIRA, and Taba has thus not been labelled as a language with rare initial rhotics in the database. An extreme case is that of languages containing no phonemic liquid at all (there are actually seven such languages in the database, 3.5%) and hence no phonemic rhotic. For obvious reasons, these languages do not possess word-initial phonemic rhotics (although they may, in theory, possess phonetic ones) but they are not considered to be WIRA languages.

It appears that the best indicator of a possible word-level distributional restriction is the ratio between the number of word-initial rhotics and that of non-initial rhotics, or between word-initial rhotics and rhotics in all word positions (including the initial). For instance, in Maninka (Rovenchak 2011), /r/ occurs 16 times in the initial position of words in a textual corpus of 28.338 tokens, but 1564 times in all positions (including the initial).4 So, even though a number of 16 words beginning with /r/ cannot be considered especially low in the absolute, and that Maninka cannot be considered as a language with no word-initial rhotics, one can nevertheless assess that /r/ is rare word-initially in Maninka because of the positional asymmetry revealed by the statistical data. So “avoidance” should not be understood as corresponding to an absolute 0 but to a remarkably and presumably significant low frequency. Unfortunately, frequency data of this sort are extremely rare for under-described languages, but the Maninka case provides a good illustration of how one can approach a better understanding and definition of what “avoidance” and “rarity” mean, especially in a cross-linguistic approach.5

The language sample used for this study is based on the set of 200 languages designed by the World Atlas of Language structures (Dryer & Haspelmath 2013) in order to serve as a representative sample of typological, genealogical and geographical diversity (see annex 2). For the purpose of this research, this set of 200 languages has been organized into a database using Excel. For each language, the following data have been collected: number of phonemic liquids in the overall inventory, sub-inventory of the phonemic laterals and rhotics and their phonetic realization, patterns of word-initial occurrence (i.e whether at least one of the liquids undergoes word-initial avoidance and the pattern according to which it does, including the allophonic processes), sources and references, and other relevant information when necessary.

This database has then been surveyed in order to reveal rhotic (and more generally liquid) phonotactic patterns in relation to word-initial avoidance, the results of which are presented in Section 4.

Relevant information concerning languages which are not included in the 200-language sample has also been collected, but all the statistics and general observations are based on the 200–language sample, except when otherwise specified.

The core lexicon of a language is made up of all its native words,6 excluding mimetic words and proper names. Loanwords have not been taken into account in order to validate the observation whether rhotics are allowed word-initially in case a language allows rhotics in the initial position of loanwords but not of native words. However, loanwords generally prove to be very useful data in order to observe the behavior of word-initial liquids in WIRA languages, since loanwords reveal the strategies that WIRA languages develop when confronted with a word-initial rhotic in the source language, an issue that will be examined in Section 5.

Let us consider a few concrete examples. Following this criterium, Imonda, for instance, has been considered as a one-liquid, rhotic-less language, because Imonda has two liquids, a /l/ and a /r/, but the /r/, an alveolar trill, occurs only in sound imitating words (Seiler 1985).

A more delicate case is that of Drehu, in which /r/ (/or /ʀ/) is said to occur in a few loans but also in local personal names (Unë & Ujicas 1984). In the present study, Drehu has been considered as containing no /r/, but as having two liquids phonemes which are /l/ and /l̥/.

The case of Japanese is interesting and deserves a detailed discussion because it is representative of a number of other comparable cases in the database. A superficial examination of a modern Japanese dictionary reveals that there exist thousands of /r/-beginning words in the language. At first sight, Japanese would thus appear as a non-WIRA language. However, a closer inspection shows that all these words fall into one of the following categories: loanword (mainly of Chinese and European origin), non-autonomous morpheme (suffix), mimetic word, word having undergone initial vowel deletion, or special slang word resulting from moraic inversion. In other words, it appears that /r/-initial words in Japanese are all of secondary development (Labrune 1993; 2014). In the end, only one Japanese noun which does not resort to the aforementioned origins can be found, the word risu (‘squirrel’). Given what we know of the language, we can suspect that this word is probably a borrowing from a dialect of Chinese or from another indeterminate language. From a panchronic point of view, Japanese can thus be labelled as a WIRA language. However, this analysis is possible only because Japanese is one of the best studied languages of the world, with a long and well-documented history and a rich philological tradition. If Japanese had been an under-studied and endangered language, for which only one general grammatical description was the only available documentation, no doubt that the secondary nature of most of its word-initial rhotics would have remained ignored, and Japanese would have been excluded from the set of WIRA languages.

The main difficulty of this study lays in the lack of comparability of the sources, because the level of phonetic detail provided by different descriptors varies considerably. Transcriptions and inventories thus differ depending on authors, language varieties, the coverage of the description (whether loanwords are included or not), the theoretical approach of the author, and a number of other factors. There therefore exist true and serious transcription and comparability problems among the data. Transcriptions, especially, are more or less precise. Some sources stay at a very superficial phonemic level and do not provide all the information needed for the present research concerning allophonic variation. Such heterogeneity in the sources undoubtedly represents the main difficulty in doing research on phonological typology.

The inventories of liquid consonants used in this study all come from direct primary sources. In a second step, these primary data have also been checked in the Lapsyd database (Lyon Albuquerque Phonological Systems Database, http://www.lapsyd.ddl.ish-lyon.cnrs.fr/lapsyd/, referenced as Maddieson et al. 2014–2020). In a few cases, when the Lapsyd data were in contradiction with some of the methodological choices explicitly made in the present paper (for instance, if the Lapsyd recognizes as a phoneme a liquid segment which occurs only in loanwords), the sources have been checked again in order to achieve a satisfying choice meeting the criteria of the present study. A total of 30 languages of the WALS 200-language set were not yet described in the Lapsyd database (as to September 2019).

The Lapsyd inventory data has been used to double check the phonemic inventories, because, as explained on the homepage of the Lapsyd website, all the inventories provided by the database have been checked and homogenized by one unique compiler, Ian Maddieson, in an effort to provide a uniform style of analysis, particularly as it relates to the inventories of consonants and vowels. Lapsyd “selects a preferred analysis for each language and attempts to harmonize the descriptions and transcriptions across all the languages” (Maddieson et al. 2014–2020). The same type of harmonizing approach as developed by Lapsyd has been pursued for the WIRA database. To the extent that the data provided by different sources have been “filtered” by a single phonologist – the author of this paper – it is hoped that a satisfying degree of uniformity and homogeneity has been achieved.

Dictionaries have also proven a very useful source of information a far as initial occurrence is concerned, for obvious reasons. When investigating the beginning of words, it is rather easy to check how many words begin with <r> or <l> (or whatever grapheme chosen to represent rhotics) in monolingual or bilingual dictionaries which use a phonographic system of transcription, even if one should be cautious about loans, the notion of word vs. root, the use of possible prothetic sounds which are usually not transcribed in standard orthographies, and the morphological structure of the language.

Finally, when available, I have more than often relied on statements such as the ones exemplified in (1) above to decide whether a given language should be categorized as a WIRA language or not.

Due to the high number of the languages under investigation, the heterogeneity of the sources, the differences in transcription according to different authors and the lack of documentation for a number of languages, the results of this study are inevitably incomplete – if not erroneous – for a number of languages. It is certain that some occurrences of WIRA have been overlooked, and that the WIRA figures are thus under-estimated. When only one short descriptive source, and no dictionary, exists for a language, the uncertainty is especially high; one cannot be sure that a possible WIRA has not been missed, or simply not mentioned, by the descriptor, because many descriptions, especially short ones, do not provide information concerning phonotactic patterns and restrictions. However, it is hoped that this first attempt to categorize and quantify the patterns of WIRA will offer a reliable picture of a phenomenon which has been largely ignored in phonological and typological research,7 and that it will help stimulate further research. Any comment or complementary information on any of the languages of the sample can be sent to the author.

The level of analysis with which this study is concerned is primarily phonological. The most frequent and typical instance of WIRA in the database is phonemic in nature, and will be henceforth labeled as emic-WIRA. Emic-WIRA is represented by languages which possess one or several phonemic rhotics, yet at least one of them does not occur word-initially. In the most typical case, it means that the language has no native words beginning with at least one of its rhotic phonemes at the lexical level. This pattern is the most frequently recorded WIRA pattern in the database. It includes languages such as Basque, Japanese, Ju|’hoan,8 Khoekhoe, Kunama, Lak, Sango, Spanish, Turkish, Yukaghir, and many others. Note that among this category, some languages may possess several rhotics, but only a subset of the rhotics may be avoided word-initially, while others are licit, as for instance in Gooniyandi, Nunggubuya, Spanish, Trumai and many Australian languages.9 Further research is needed to investigate whether there exists a correlation between the number of rhotic phonemes of a given language and the manner in which a WIRA pattern occurs in that language.

There also exists another type of WIRA which can be labeled as phonetic, or etic-WIRA. This latter type is less easily detectable and presumably less often reported in sources but it cannot be ignored. Two different sub-types of etic-WIRA have been identified. In the first sub-type, a given rhotic occurs word-initially at the phonological and lexical levels, but not at the phonetic level, because in the initial position, it is either realized as a non-rhotic or it is preceded by a prothetic vowel. For instance, in Warao, the rhotic is a flap intervocalically but always a stop [d] in the initial (Romero-Figueroa 1997). So in Warao, there exist word-initial phonemic rhotics, but no phonetic ones. Another example is Wichita: in Wichita, /ɾ/ is nasalized in the initial position (Garvin 1950). An example of a language which adds a prothetic element to a word-initial rhotic is Tiwi. In Tiwi, a language with two rhotic phonemes, /r/ does not occur word-initially, while /ɹ/ is rare in that position. According to Osborne (1974), in the few words in which /ɹ/ occurs initially, it is often preceded by a slight introductory glide. Tiwi thus appears as a language which exhibits both emic- and etic-WIRA. In Armenian, too, a language with two rhotics, a prothetic schwa is optionally inserted before the handful of /ɻ/ beginning words (Vaux 1998: 122).10

The second sub-type of etic-WIRA pertains to languages which contain a rhotic segment which stands as the non-initial allophone of some other, non-liquid phoneme, (mainly /d/ or /t/), not as the allophone of a rhotic or of a lateral phoneme. For instance, in Koromfe, a language with no phonemic rhotic, the alveolar flap [ɾ] is an allophone of /d/ in native words, and /d/ occurs as [d] only word-initially and after a nasal stop consonant (Rennison 1997). Another example is Dani (Lower Grand Valley), which has only one liquid phoneme, /l/, and no rhotic phoneme, but a rhotic ([r]) occurs as an allophone of /t/ intervocalically (van der Stap 1966). American English, which has one phonemic rhotic /ɹ/ which appears word-initially, is not an emic-WIRA language, but it is an etic-WIRA language because an alveolar flap [ɾ] occurs as the inter-vocalic allophone of /t/ or /d/.

Deciding whether a language is an emic- or etic-language sometimes poses a methodological problem because in a number of cases, for instance Dani or Koromfe, the language could probably have been described as having a rhotic phoneme, say /r/, with a word-initial allophone [d] or [t], depending on the analytical choices made by the descriptor. Yet it should be noted that in languages such as Dani and Koromfe, even if the rhotic allophone had been granted the status of representing the phoneme in preference to the non-rhotic allophone, or if the language had been analyzed as containing two different phonemes standing in complementary distribution, the language would still be classified as an etic-WIRA or emic-WIRA language according to the approach11 followed in this study.

There are thus, strictly speaking, two types of WIRA that need to be distinguished: phonological, or phonemic WIRA (= emic-WIRA) and phonetic WIRA (=etic-WIRA). These two major types are synthesized in Table 3.

Recall that there are also non-WIRA languages, which are of two different types, too: languages with one or several phonemic liquids which appear word-initially with no restriction, and languages which lack both phonemic and phonetic rhotics. Among the latter type there is a rather high proportion of languages for which we lack precise and detailed descriptions, especially concerning the possibility of an etic-WIRA feature.12

After defining the various degrees of emic-WIRA, this section provides the general statistics for this type of WIRA. Emic-WIRA has been categorized along a scale of six values which serve to identify and label the different word-level distributional patterns displayed by rhotic phonemes, as well as the level of information which could be gathered for each language. The status of each of the 200 languages of the sample with regards to these labels can be found in annex 2. The values are as follows:

Examples: Aymara, Basque, Burushaski, Japanese, Trumai, etc.

Examples: Daga, Kera, Maricopa, Selknam, Swahili, etc.

Examples: Arapesh, Chinantec (Lealao) (only two languages).

Examples: Diola Fogny, Huitoto (Minica), Krongo, Lango, Ngiti, etc.

Examples: Arabic, Cayuvava, English, Maori, Quechua (Imbabura), etc

Examples: Ekari, Ket, Supyire, Pirahã, Usan, etc.

Note that all the languages coming under one of the above labels may also exhibit etic-WIRA in addition to emic-WIRA.

The detailed figures of emic-WIRA in the 200-language sample appear in Table 4. The five first categories exclusively concern phonemic rhotics, thus providing data for emic-WIRA. See also annex 2 for the complete list of languages and their WIRA status.

The number of languages which exhibit some degree of emic-WIRA in the language sample amounts to 78 out of 200 (39%), vs. 81 (40.5%) which do not.

Furthermore, it should be noted that the remaining 81 languages are not all necessarily languages which can be considered as accepting rhotics word-initially. They also include languages about which no sufficient information on the status of rhotics word-initially was available (such languages are likely to be found in the “present?” category.). This is because, when no specific information about WIRA was found for a language, this language has been classified, by default, as a word-initial rhotic accepting language. So the number of word-initial rhotic avoiding languages might be higher than indicated.

Languages pertaining to the last category (“irrelevant”) contain no phonemic rhotic, but while they do not represent cases of emic-WIRA, they may qualify for etic-WIRA (just as emic-WIRA languages may, too). For this reason, Table 4 does not tell us the whole story about WIRA. We also have to survey the language sample for specific cases of etic-WIRA, because etic-WIRA languages may or may not have phonemic rhotics. This will be done in the next section.

Clearly, the number of languages which avoid rhotics word-initially at the phonemic level is strikingly much higher than expected on a purely random basis, a feature which has been overlooked, probably owing to the fact that the most studied Indo-European languages like English, French, German or Russian do allow rhotics at the beginning of words, while in Spanish, another dominant Indo-European language, WIRA is obscured by the orthography. The first finding of this study is thus that rhotic avoidance in the initial position of words constitutes a recurring structural property in the world’s languages.

Let us now examine etic-WIRA. As already mentioned, etic-WIRA occurs under two different sub-types: in the first sub-type (b. in Table 3), a phonemic initial rhotic undergoes mutation and is phonetically realized as a non-rhotic segment, generally a coronal stop, or is preceded by a prothetic element, always a vowel, generally a schwa. In the second sub-type (c. in Table 3), a phonetic rhotic stands as a positional allophone of a non-rhotic phoneme in the non-initial position. In both types, and putting apart the prothetic vowel cases, one observes a complementary distribution between a non-rhotic and a rhotic in, respectively, the word-initial and the non-initial position. The languages which exhibit etic-WIRA in the sample are presented in Table 5, with the phonetic details of the alternation involving the rhotic segment. Note also that etic-WIRA sometimes implies neutralization, whereby the rhotic is distinctive in medial position but neutralized with some other phoneme in word-initial position.

There are 30 languages which have been identified as etic-WIRA languages, representing 15% of the sample. Note that nine of them were also in the category of emic-WIRA. It is highly probable that there exist many other cases of etic-WIRA in the sample. This type of WIRA is probably under-estimated because the level of phonetic detail which allows its identification is not always achieved in descriptions. Moreover, dictionaries may not record the presence of a prothetic vowel in r-initial words. Or, on the contrary, the prothetic vowel has become fossilized and lexicalized, and it is now denoted in the orthography, which makes the language look like an emic-WIRA language, as in Yup’ik (Central).