📄 Phonetics of Speech Acts: A Pilot Study

Summary

This pilot study asks whether prosodic rhythm in Mandarin Chinese depends on the speech act being performed and on lexical tone. Prior speech-act research has focused on semantics and syntax and has said little about phonetics. Ten native speakers, all students at National Sun Yat-sen University, produced four fixed three-syllable phrases. Each phrase used a single tone throughout and the same final vowel. Speakers said each phrase as a command, warning, invitation, refusal and request. They could add their own words around the phrase but could not change it. The author measured duration, pitch and intensity at every vowel and expressed each measure as a pairwise variability index (PVI). One-way ANOVA tested the effects, followed by Tukey HSD. Speech act had no significant effect on any measure. Tone affected intensity and pitch PVI but not duration. The author suggests two reasons for the null speech-act result: the role-play elicitation felt artificial, and the cues of illocutionary force probably fall on the added framing words rather than the controlled phrase.

Key points

• Participants: 10 speakers (3 male, 7 female), aged 19–30. Each produced 20 utterances: 4 phrases × 5 speech acts, giving 200 observations. The phrases were paobaobao, miaochaohao, paobaodao and qiaobaogao.
• Speech act showed no significant effect on PVI:
  • intensity F(4,195)=.23, p=.92
  • pitch F=1.61, p=.173
  • duration F=.55, p=.702
• Tone significantly affected intensity PVI (F(3,196)=6.256). Tukey HSD found significant pairs tone1–tone3 (p=.015), tone2–tone4 (p=.024) and tone3–tone4 (p=.001).
• Tone also affected pitch PVI (F=4.513, p=.004), with significant pairs tone1–tone3 (p=.004) and tone1–tone4 (p=.022). Duration PVI was unaffected (F=1.264, p=.288).
• The lack of a tone2–tone3 difference is attributed to Mandarin third-tone sandhi, in which a tone 3 followed by another tone 3 surfaces like tone 2.

Interesting ideas

• I1. Illocutionary force tends to sit in framing material such as "do not want to", not in the content verb phrase, so a classifier that inspects only the core predicate will miss the act.
• I2. Fixing a core phrase while letting speakers freely improvise the frame around it isolates whether an effect lives in the core or in its surroundings.
• I3. Role-play without a given situation produced unconvincing data, and the author recommends near-natural sources like dramas or films instead.
• I4. A null difference between two categories was explained by a known neutralization rule (tone sandhi) rather than taken as evidence that the categories are the same.

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

• I1 (frames): speech-act detection in prag should read modal, negation and matrix-clause wrappers around the main verb as primary evidence, and explain each act label by pointing to those words.
• I2 (learn): build test sets where a fixed clause is wrapped in varied frames (command, refusal, invitation) to check that frame outputs change with the wrapper and not with the unchanged core, serving as a controlled negative check.
• I3 (learn): prefer silver speech-act data drawn from real dialogue (e.g., fiction and play scripts) over context-free invented prompts when training or evaluating pragmatics.
• I4 (tree): when two labels show no measurable difference after UD dialect conversion, first check whether a conversion rule collapses them before concluding the distinction is useless.