Overlap of Phonemes

An important lesson here is that the phonetic categories do not simply jump off the page when examining measurements of individual tokens. The data found here show tremendous overlap between different phonemes. In the most extreme example I found, the distribution of the // phoneme for Rita is nearly coextensive with the entire vowel space, as shown in Figure .

Figure: Rita's Vowels; Rita's /E/

52#52

Lisker (1949) found some irreducible overlap between F1 and F2 measurements for // and /æ/ in a particular context (p_p). (A replication of his thesis was conducted, and is described on pages ff.) Here we find that // overlaps not just /æ/ but all the vowels in the entire space. The dispersion of // is the most extreme case in Rita's data. (It is also, perhaps not coincidentally, the most recent vowel to join the chain shift.) However, many vowels in Rita's system spread over nearly as much of the space, and few cover less that half of the entire area. For similar results in another dialect, see the charts of Juba's (Jamaican Creole) long and short vowels, on page . The striking amount of dispersion shown in Figure  is not spurious or mistaken. Because each of the tokens was examined individually, and formant tracks were corrected,^7.18 the number of erroneous measurements are few: the dispersion of // tokens in Figure  cannot be attributed to gross errors of measurement. Thus we may conclude that when a single speaker's natural conversational speech is analysed, including both stressed and unstressed tokens and all the coarticulation and reduction to be found in such contexts, the amount of overlapping between vowels is extreme. The problem Lisker (1949) points out becomes much worse under these conditions.

The fact that the nucleus formant frequencies are a very reduced representation of the vowel tokens studied leads us to speculate that overlapping phonemes may be distinguished by other factors which are not present in these charts, such as glides in one direction or another, differences in vowel duration, F3 frequency differences, even formant bandwidth or amplitude differences. Information about the the stress level and phonetic context of the sounds measured, plus dialect-specific knowledge about the contexts and rules for natural-speech reduction, may be sufficient to disambiguate most tokens on the basis of phonetic and phonetically derived information alone. Overlap is not necessarily an insoluble problem for classification. It is quite possible, even likely, however, that a residue of unclassifiable tokens will remain, even after applying all the phonetic knowledge possible. Some sounds are simply ambiguous, and listeners must bring higher-level linguistic knowledge to bear on the decoding problem, or fail in their efforts.^7.19

In fact, since our primary concern here is with phoneme-internal variation, this is actually, for present purposes, a good situation. The more variation to be found in the pronunciations of a single phoneme, the better we will be able to substantiate the sub-phonemic alternations or the allophonic effects which are the focus of this work. This is an important reason that natural speech is superior to laboratory speech for studying effects of coarticulation and reduction. More processes apply, more frequently, in natural speech.

Contextual variation is explored in Chapter 10. This source of variation as well as time-varying spectral information are essential to recognizing the phonological representations corresponding to natural speech. The next section of this chapter discusses the mean locations of vowels in F1-F2 space, the locations from which vowels deviate when affected by particular influences of stress and phonetic context. The subsequent section investigates the influence of stress, and the last section examines the effects of the consonants in the adjacent phonological context.