The Treatment of Vowels Preceding 'r' in a Keyword Lexicon of English

Work is progressing on a keyword lexicon aimed at enabling the synthesis of various regional accents of English. This paper focuses on a particular issue, that of vowels before orthographic 'r'. These vowels are discussed with respect to rhotic and nonrhotic accents, in terms of both keyword sets and phonetic realisation. Criteria for the use of keysymbols are discussed, and it is noted that these criteria result in inclusion of postvocalic |r| in the lexicon, with deletion by rule for non-rhotic accents. It is noted that some keyvowels in our original set have had to be split, while others may prove to be redundant. 1. THE KEYWORD LEXICON As described in a previous paper [1], the keyword lexicon, rather than using conventional phonetic or phonemic symbols or their ASCII equivalents, uses transcriptions based on keywords. This strategy enables one lexicon to represent numerous regional accents. Where an accent has a phonemic distinction, this will be represented in the dictionary; for instance, 'horse' and 'hoarse', although homophones in RP, are distinguished in Scottish English and so must be represented by different keysymbols in the lexicon. These can be described as the NORTH vowel and the FORCE vowel [2]. 2. DICTIONARY TREATMENT OF 'R' 2.1. Terminology In this paper I shall use the terms 'rhotic' and 'non-rhotic' to refer to accent types. Although somewhat unsatisfactory, the traditional term 'post-vocalic' is used to describe / / in both pre-consonantal and word-final environments: C V r /_ { # In these environments / / is consistently pronounced in rhotic accents; in non-rhotic accents / / does not exist in preconsonantal position, and is variable in word-final position depending on both the following word and the regional accent. 'Word-internal pre-vocalic' is used for the environment: r /_ V In this context / / is pronounced whatever the accent type. 2.2. Accent-independent Transcription of 'r' The keyword lexicon covers both rhotic and non-rhotic accents. Initially it was planned to use a special symbol for post-vocalic 'r' as opposed to word-internal pre-vocalic 'r' (c.f. [3], which uses |rr| versus |r|). However, as work has progressed it has proved undesirable to include all regional differences as separate keysymbols in the lexicon, as this results in numerous keysymbols to encode differences which are in many cases predictable. Given that post-lexical rules are in any case necessary for cross-word environmental conditioning, and are useful for within-word allophones such as glottal stops, which have different scopes of application in different accents, we need to decide on the balance between lexical encoding and rulebased derivation (see [1]). The criteria which have been drawn up are based on the traditional distinction between phonemes, which we need to include in the lexicon, and phones, which are predictable. These criteria are: Principle I. For segments, all phonemes in each accent, and only units which have phonemic status in at least one accent, should be encoded differently in the base lexicon. Principle II. If the phonetic realisation of a unit is predictable from the environment (which includes keysymbols, syllable and morpheme boundaries) then this will be derived by accent-dependent post-lexical processing. In the case of a conflict, Principle II overrides Principle I as it reduces redundancy in the transcriptions. Following these criteria, we must encode the vowel difference between Scottish 'horse' and 'hoarse' in the base lexica, although many accents of English do not make this distinction; on the other hand, dark and light / /, which are never contrastive, will be generated post-lexically. Principles I and II give us a logical framework for lexical encoding of linguistic information versus derivation by rule. Following the criteria, we must reject the use of separate keysymbols for post-vocalic and word-internal pre-vocalic 'r'. Given the keysymbol transcriptions farm f * ar r m safari s @ . f * ar . r iy we can predict that the |r| symbol in 'farm' will be realised as (or , , and so on) in rhotic accents, and as null in nonrhotic accents. On the other hand, the |r| in 'safari' will be realised as an in all accents. We no longer have the simple mapping of the earlier splitsymbol approach, in which |r| was realised as in all accents, and word-internal |rr| was realised as in rhotic accents and null in non-rhotic accents; the balance of description has shifted slightly from the accent-independent lexicon to accentdependent rules. However, it should be noted that in the splitsymbol approach post-lexical rules are still necessary to predict the realisation of word-final 'r' in non-rhotic accents which use linking . For the word far f * ar r we need to know what follows before we can predict the pronunciation of |r| in non-rhotic accents such as RP. (Note that in some accents of English, for instance South African English, linking 'r' is not pronounced, so for these accents this does not apply.) Intrusive 'r', although used to differing degrees by many people, is still regarded as erroneous (though see [4] and related references), and so it is not included in the keyword transcriptions or post-lexical rules; it could, though, be introduced by rule if desired (see [5]). Although Principles I and II above are based on the traditional phonological distinction between phonemes and allophones, the use of these criteria in a multi-accent lexicon means that the resulting output does not correspond to the traditional division between the two. Keysymbols in the lexicon are not directly equivalent to phonemes, and the output of the post-lexical rules cannot be defined in terms of allophones. RP, phonemic transcription RP, phonetic transcription far / /