Neutralization in Taiwanese Tone Sandhi

Taiwanese tone sandhi, although a phrase-level pattern, has been argued to be essentially lexicalized. We thus expected it to be phonetically categorical and insensitive to pragmatic context. To test these predictions, we had native speakers read aloud pairs of sentences identical except for morphemes expected to be neutralized by tone sandhi, both across and within prosodic positions. Speakers were recorded in two pragmatic conditions: alone and in the presence of an attentive listener. The across-position comparisons showed no difference in f0, but the within-position comparisons did find a small but significant difference. However, this latter result is not consistent with true incomplete neutralization, since the slopes did not preserve the putatively underlying tone contours, and there was no effect of pragmatic condition. Moreover, in a second experiment, the same participants were unable to discriminate or identify the tones in the within-position comparisons. Taiwanese tone sandhi thus seems to involve the manipulation of categorical phonological units, though gradient adjustments can appear in reading pronunciations.