Vowel variation in Southern Sotho: an acoustic investigation

For the development of Human Language Technologies such as automatic speech recognition systems or text-to-speech systems, exact acoustic and phonological information of the language in question is essential. In the case of Southern Sotho, the current study is a first step in the direction of providing such information. Measurements of the frequencies of the first two formants of the vowels represented by the orthographic symbols i e a o u are made in several contexts in order to motivate particular phonetic and phonemic representations of these vowels, regardless of the conventional orthography. Our measurements corroborate some of the earlier impressionistic research into Southern Sotho. In particular, the four variants of the vowels orthographically represented by ‘e’ and ‘o’ respectively are clearly present, indicating that there are at least seven distinct (phonemic) vowels. These variants are distinguished mainly by vowel height, that is, by F1 differences. On the other hand, our results oppose the established knowledge in some important ways. Most notably, we find that harmonic vowel raising produces a larger change in vowel height than the height difference between the pairs of phonemically distinct middle vowels (/ɔ/ and /o/, /ε/ and /e/), respectively. The finding is seemingly reliable as this tendency is constant across all words investigated. An interesting finding is a significant modification of the vowels /u/ and /a/ in certain contexts. These are vowels previously believed not to have any allophones at all. Lastly, a number of important unresolved issues are highlighted. Introduction The development of Human Language Technologies (HLTs) depends on the availability of several forms of linguistic information. A state-of-the-art spellchecker, for example, requires not only a sufficiently diverse and accurate lexicon, but also accurate descriptions of morphological processes, syntactic structures and the like. For world languages such as English and Mandarin Chinese, such information is widely available, and the principal challenge of the HLT developer is to represent such knowledge in a way that meets the need of users. For the vast majority of languages, however, the necessary information is much harder to obtain. The relevant expertise is often held by a few (if any) linguists, and is generally not codified in a format that directly supports the development of HLT systems. Since the population of linguists working on a particular language may be small, there is often insufficient independent verification of those findings that have been published, which forces the developers of HLT systems to select from sometimes contradictory accounts of basic linguistic facts (while somehow filling in those portions that are necessary for their systems but not available in the published literature). For spoken language technologies, such as automatic speech recognition (ASR) systems or text-tospeech (TTS) systems, an exact definition of the phonemic inventory of a language is often the first linguistic foundation required. A complete phoneme set is employed to express all lexical items in terms of a productive set of units; this ‘pronunciation dictionary’ is then supplemented with relevant information such as the presence of major phonetic variants or suprasegmental phenomena, and used as the basis for the development of the statistical models that are the heart of HLT systems. A reliable description of the phonology (along with major phonetic variants) is therefore of great importance. Barnard and Wissing 256 In the current contribution, we describe our efforts to establish an exact and reliable description of the vowel system in a less well-studied language, namely Southern Sotho. Although Southern Sotho (along with the closely related languages Setswana and Northern Sotho) has been the subject of active linguistic study for almost a century, there is still considerable debate on the precise nature of its vowel system. In the second section we review this debate, and contrast the issues that remain contentious with those that are considered uncontroversial. The third section proceeds with a description of the experimental methods that were employed in order to address some of the issues raised in the second section. The fourth section contains the results of our investigations, and in the fifth section we conclude with a summary of our main findings as well as a discussion of unresolved issues that require further study. Background The 7 phonemes1 /i e ε a ɔ o u/ are generally accepted for Southern Sotho (Tucker, 1929; Doke & Mofokeng, 1957; Guma, 1971; Roux, 1983; Khabanyane, 1991; and Selebeleng 1997) as well as for Setswana (Krüger & Snyman, s.a.) and Northern Sotho (Poulos & Louwrens, 1994). These phonemes are also generally used in pronunciation dictionaries for these languages. They are described in various articulatory terms. Some authors (for example, Tucker, 1929; and Doke & Mofokeng, 1957) make use of the cardinal vowel system introduced by Daniel Jones and also employed by the International Phonetic Association (see Figure 1 for a slightly modified version, where vowel rounding is suppressed since it is not relevant to our discussion). Ignoring rounding, this system indicates 8 positions, 1 cardinal vowel per position. Front as well as back vowels can be close (high),2 that is, cardinal vowels 1 and 8; and open (or low), that is, cardinal vowels 4 and 5. Cardinal vowels 2, 3 and 6, 7 are all mid, specified by the addition of close or open, yielding the combinations close-mid (cardinal vowels 2 and 7), and open-mid (cardinal vowels 3 and 6). Six of the 7 Southern Sotho phonemes mentioned previously, /i e ε ɔ o u/, then correspond to cardinal vowels 1, 2, 3, 6, 7, 8; /a/ takes up a position between cardinal vowels 4 and 5.3 In the earlier literature (cf. notably Tucker, 1929; and Doke & Mofokeng, 1957) on the Southern Sotho vowels, two allophones of the open-mid phonemes are acknowledged (that is, front [ε] and back [ɔ]).4 Generally, such allophones are explained by a process of vowel raising,5 sometimes also referred to as assimilation or vowel harmony. This process of vowel raising is indicated in Figure 1 by solid arrows, and is predictable on the basis of specific phonetic contexts. A variety of such phonetic contexts is responsible for such vowel raising. The close vowels /i/ and /u/ in following syllables (cited by all of the authors on Southern Sotho quoted previously) are the most important triggers for the raising of vowels in previous syllables. Sometimes close-mid vowels are also identified as having such a raising effect on preceding open-mid vowels. Furthermore, some consonants are mentioned as having similar raising effects too, namely, syllabic nasal consonants /n/, /ɲ/, /ŋ/ (including the locative suffix /-ŋ/) and the alveolar consonants /tsʼ/, /s/. Vowels in word-final positions are said to be raised as well (especially mentioned by Tucker, 1929; and Doke & Mofokeng, 1957). In addition, based on the existence of minimal pairs such as /p’εp’a/ (‘a piece of chuff’) and /p’εp’a/ (‘carry on the back’), Khabanyane (1991) concludes that these vowels are not allophones but phonemes of Southern Sotho, thus to be indicated as /ε/ and /ɔ/, and not as [ε] and [ɔ].6 Selebeleng (1997) follows her in this assumption. In more recent work, two more vowels have come to be accepted, in addition to the abovementioned nine (phonetic) vowels, namely the raised close-mid [e] and [o] (Guma, 1971; Khabanyane, 1991; Selebeleng 1997; and Roux, 1983, an exception as he feels that a nine-vowel system is adequate). In summary, then, the existence of seven phonemic vowels and two additional open-mid allophones is currently not contentious. However, the possibility of an additional pair of phonemic vowels (which are collocated with these two allophones) and yet another pair of (close-mid) allophones, is somewhat more controversial. Note that we began our discussion at the orthographic level because this allows us to indicate unambiguously the tokens under discussion – not because we believe that the five orthographic forms have any special status (as pointed out above, the existence of at least seven vowels in Southern African Linguistics and Applied Language Studies 2008, 26(2): 255–265 257 Southern Sotho is uncontroversial). It is also true that the orthographic form (without diacritics) is usually taken as input by systems for speech recognition and speech synthesis, since these are the forms commonly used in writing. Methods Participants Twelve speakers of Southern Sotho, 6 of each sex, volunteered to read the speech materials (speech stimuli). Of these, 6 were either students or employees of the Qwaqwa campus of the University of the Free State (3 males and 3 females, aged 19–45 years), and 6 were Grade 11 learners at the Lekota Secondary School (also 3 males and 3 females, aged 17–20 years). Both these institutions are situated in the rural region of Phuthaditshaba, Eastern Free State, South Africa. Speech stimuli We focused on the mid vowels of Southern Sotho, which include the open-mid and close-mid vowels. In light of the uncertain status of some of these (in particular, the uncertainty surrounding the number of distinct phones as described in the second section), we take no a priori position as to the precise quality of the vowels of the words in the specific constructions we chose. For this reason we simply refer to these vowels as either ‘e’ or ‘o’. By careful measurement of the formant frequencies of these vowels, we hope to motivate particular phonetic and phonemic representations of these vowels, regardless of the conventional orthography. In this way, we hope to provide an empirical basis for the transcriptions provided in dictionaries and standard texts. Figure 1: The Southern Sotho phonemes and allophones imposed on the IPA Vowel Chart Phonemes appear outside the figure, allophones inside (they are underlined) ε