Overlap of Phonemes

As Lisker (1949) showed, distinct pronunciations of distinct phonemes overlap in formant space (in that case, /æ, /, in the context, p_p.) As I showed above in the Methods chapter, my pronunciations of the ``same'' sounds do not overlap to the same extent as Lisker's pronunciations, reflecting the likely fact that we did not pronounce them in the same way, and thus that these ``same'' sounds are not really the same at all, for the different speakers involved. Nonetheless, Lisker's point is important. Vowels which occupy the same location on a standard formant chart can sound quite different. A good example of this is the pair /I/ and /ey/. /I/ glides inward and down, and is an ``intrinsically short'' vowel, while /ey/ glides frontward and up, and is ``intrinsically long''. These vowels glide in different directions, for different durations. However, these vowels frequently glide through the same area of formant space. Thus the formant frequency measurements at the ``nucleus'' may be identical for the two classes despite the fact that they don't sound identical at all.

Overlap in formant space does not necessarily imply auditory identity of the overlapping tokens. On the other hand, small differences in formant space can correspond to very clear and audible phonetic differences, as shown in the study of /pæp/ and /pp/ discussed above (page ). The classical studies of Flanagan (1955) show how small an audible difference can be: in the range of 50Hz for F2 and of 30Hz for F1.

The data presented in this work contains vast amounts of overlap between phonemes. Other acoustic features besides the instantaneous measurements of the first two formant frequencies may still distinguish the overlapping sounds. While this overlap will in fact cause considerable trouble for the most simple-minded of speech processing algorithms, it need not alarm us. Overlap does not imply neutralization!

On the other hand, when considering different allophones of the same phoneme that are measured in a consistent way, one may infer, from measurable and statistically significant differences between distributions in formant space of two allophones, that there is some real phonetic difference made between them. A significant difference in formant space between linguistically characterized sets of sounds amounts to a proof that an allophonic difference exists.

We will proceed in this way, using this rule of inference to argue from the documented distributions on F1-F2 charts to inferred phonetic or phonological processes.