Assessing a Mathematical Model of Syllable Production via CTW Alignment with EMA Data

• The paper distinguishes four transparency levels, motivated by questions about interpretability and transparency in articulatory modeling amid progress in neural speech synthesis.Source: §1 · p. 1
• The Level 3 model explicitly links gestures to linguistic content and uses mathematical syllabic planning with Maeda parameters; CTW aligns model trajectories to the EMA temporal grid.The approach prioritizes linguistic transparency and access to higher-level and low-dimensional control variables over perceptual synthesis quality, while CTW may introduce substantial local temporal distortions.Source: §1 · p. 1
• Level 4 would require physiological transparency, but the authors say the model’s purely mathematical formulation appears to conflict with that level of physiological fidelity.They nevertheless suggest it may help address substantial gaps in the field.Source: §1 · p. 1
• The study tests whether a model designed for prospective syllabic planning captures experimentally observed organization in regularly structured sequences, where anticipation was reported at least 200 ms before segment onset.Only minor adaptations were needed to reproduce the syllabic configurations used in.Source: §1 · p. 1
• The evaluation combines parametric generation, asymmetric temporal alignment, and statistical validation of articulatory-parameter and formant correspondence.Its objective is to assess whether the model captures experimentally observed syllabic organization within a shared articulatory parameter space.Source: §1 · p. 1

2. Methodological details

• The comparison asks whether a deterministic model driven by phoneme and syllable structure produces articulatory modulations that align with EMA recordings of regularly organized phrases.Source: §2 · p. 2

2.1. Data and Parameter Space

• The empirical dataset contains nine EMA repetitions of seven phrases from one speaker; a linear mapping projects 13 of 14 sensor coordinates into six Maeda-like parameters.The trajectories are decimated from 500 to 100 Hz, low-pass filtered at 8 Hz, and z-normalized; they are then rescaled using each file’s model statistics, though LipH discrepancies can remain because of model rectification.Source: §2.1 · p. 2

2.2. Model Generation

• The model represents syllables as directed graphs of vowel or consonant targets and transitions, using coarticulation and anticipatory vowels to connect adjacent syllables.The phrase analogs use a limited phonetic inventory and are built from six words per file; the model also generates Maeda parameters and formant trajectories.Source: §2.2 · pp. 2–3

3.1. Temporal Realignment

• Across 63 paths, Condition A warping paths stayed near a straight line, whereas Condition B paths deviated more; one path warped all six parameters together.The EMA-derived sequence retained 1000 samples, with the first 800 representing active speech; aligned model values were averaged and low-pass filtered at 8 Hz.Source: §3.1 · p. 3

3.2. Correlations between CS and Modeled Trajectories

• Condition A produced significantly higher mean correlations than Condition B across all six articulatory parameters (Holm-corrected p < 0.05).The largest differences were LipP (∆= 0.2195), Body (∆= 0.1506), and Dorsum (∆= 0.1192).Source: §3.2 · p. 3
• F1, F2, and F3 correlations were also significantly higher in Condition A than B (all Holm-corrected p < 0.05).The largest difference was F1 (∆= 0.2375), followed by F2 (∆= 0.1747) and F3 (∆= 0.0998).Source: §3.2 · pp. 3–4
• Condition B retained substantial correlations despite phonetic incongruence, consistent with shared syllabic structure and lexical material across phrases.The passage suggests that CTW can preserve higher-level syllabic organization even when segmental targets differ.Source: §3.2 · p. 4
• Phonetic congruence was associated with better articulatory and acoustic correspondence, while shared syllabic structure sustained baseline alignment in both conditions.Jaw and LipH correlations were comparatively lower in Condition A, but acoustic reconstruction, particularly F2, remained robust; renormalizing Condition B with incongruent-pair statistics yielded a similar contrast.Source: §3.2 · p. 4

3.3. Detailed Analysis of CTW Alignment

• In the detailed example, CTW compensated for many structural differences in Condition B, but the missing /a/ in F1 and deletion of /di/ remained notable mismatches.The /a/ mismatch corresponded to absent lip opening and jaw lowering, while many-to-one temporal mapping deleted /di/. Figure 3 compares F1–F3 trajectories for congruent and incongruent conditions, with CS and speech spectrograms for reference.Source: §3.3 · p. 4, Figure 3 · p. 4

• The experiment demonstrates that a mathematical model can generate articulatory parameters and align them with natural articulatory data in regular, controlled linguistic content.The model reproduces temporal modulations at the articulatory-parameter level, and CTW is presented as a promising tool for alignment and probing phonetic congruence parametrically.Source: §4 · p. 4