📄 Simple Syllable Structure and English Pre-nucleus Glides

Summary

The paper examines if Duanmu's Simple Syllable Structure (SSS), originally developed for Chinese and later extended to Korean, can also account for English onsets. SSS allows one onset slot. Under Feature Geometry, a consonant and a glide made with different articulators may share one feature tree. If SSS holds, then onglides in words like twin or pure should be secondary articulations on the consonant and should be about as short as one segment. To test this, the author ran a small acoustic pilot. Two male speakers each read 18 words five times inside a fixed carrier sentence. Each set contrasted a consonant+glide onset with the bare consonant and the bare glide, for example twin, tin and win. On average, CG onsets were clearly longer than a single consonant but slightly shorter than C and G added together. The two glides also behaved differently: Cj came close to two full segments, while Cw was noticeably compressed. SSS and Feature Geometry predict the same behaviour for both glides, so the results argue against them. The /j/ versus /w/ split instead fits earlier phonological claims that place /j/ in the nucleus.

Key points

• Materials and setup:
  • Six triplets (C / G / CG), recorded in a soundproof room.
  • Each word sat in the frame "Say X again", which avoids lengthening before a pause.
  • Stop durations counted both closure and aspiration; glide transitions were left out.
• Mean durations:
  • Single C: 136.7 ms.
  • Glide alone: 59.0 ms.
  • CG: 165.6 ms.
  • C+G sum: 195.7 ms.
  • SSS predicted roughly 97.9 ms (half the sum) or 136.7 ms (the longer consonant).
• Glide asymmetry:
  • Cj was 179.8 ms against a C+j sum of 187.1 ms, a gap of only 7.3 ms.
  • Cw was 151.4 ms against a C+w sum of 204.3 ms, a gap of 52.9 ms.
  • With the two-segment sum scaled to 2, Cw comes to 1.48 and Cj to 1.92.
• The findings agree with Davis and Hammond (1995), who put /w/ in the onset and /j/ in the nucleus. They also agree with Duanmu (1997), who treats /ju/ as a long diphthong.
• The author calls this a pilot. Proposed follow-ups are more speakers and tokens, statistical tests, whole-word and vowel durations, coda clusters, and a comparison with Korean.

Interesting ideas

• I1. The paper tests a structural claim by measuring the composite against two baselines: the single part and the sum of the parts. This tells "one unit" apart from "two units" instead of just reading a raw number.
• I2. Each test item is placed inside a fixed carrier frame, so that edge effects do not distort what is measured.
• I3. The pooled average hid the real finding; only splitting the data by glide type showed that /j/ and /w/ behave in opposite ways.
• I4. Raw values are rescaled so that "two separate units" equals 2, which makes conditions with different absolute sizes directly comparable.
• I5. The paper stresses that a theory permitting one representation does not force it, so evidence has to decide whether a combination is one unit or two.

For Mova

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

• I1 (learn): Evaluations can check a merged analysis, such as a multiword expression treated as one node, against both the component-wise baseline and the single-unit baseline, as a built-in negative control.
• I2 (learn): When probing the parser or tagger on target words, embedding them in constant template sentences separates the word's own behaviour from boundary and position artifacts.
• I3 (tree): Breaking parser accuracy down by fine-grained construction subtype, instead of reporting one LAS, can expose subclasses that need different treatment in UD conversion or repair rules.
• I4 (world): Normalizing measures against a meaningful reference point gives the world model scale-free comparisons that stay explainable when quantities differ in magnitude.
• I5 (tokens): Decisions on merging or splitting tokens, such as contractions and compounds, should be driven by measurable evidence rather than by the bare fact that a single-token analysis is formally possible.