UCLA

This dissertation investigates how the Taiwanese Tone Sandhi Groups are perceived, and the acoustic/phonetics correlates of listeners' judgments. A series of perception experiments have been conducted to scrutinize the following topics - Taiwanese tone neutralization, Tone Sandhi Group (TSG) as a prosodic domain, perceived boundary strength in Taiwanese and Taiwanese sentence disambiguation.

Taiwanese tone neutralization was examined with a checked tone recognition experiment and a corpus study. In the tone recognition gating experiment, the difference between the accuracy rate on citation and sandhi tones suggested that the two tones maybe be incompletely neutralized but as members of the same tonal category. The corpus study found that duration, f0 range and voice quality in domain-final position preserve distinctions between all the tone pairs. More specifically, the citation tones are longer in duration, wider in F0 range and creakier in voice quality.

The prosodic domain identification experiment revealed that participants were able to identify TSG as well as the other two prosodic levels, Word and Intonation Phrase even when the speech signal is low-pass filtered (i.e. only the prosodic cues - duration, f0, and voice quality - were retained).

The perceived boundary strength rating experiment further showed that participants assigned different ratings for the three prosodic levels in spontaneous speech stimuli, and the acoustic analyses revealed that duration, f0 and voice quality cues in the last syllable in each stimulus were useful to the listeners. However, multiple linear regression indicated that these acoustic cues were far from sufficient to account for listeners' perception. In other words, to discriminate one prosodic domain from the others, listeners apparently need acoustic/prosodic cues not only from the syllables at the prosodic boundary, but also the syllables before the boundary within the same prosodic domain.

In the sentence disambiguation experiment, participants were able to interpret ambiguous sentences accurately. The results suggest that prosodic cues such as final lengthening and pitch reset, found at the disambiguation points, provide a strong basis for sentence identification.