PaperlayerView exact revision on arXiv

Source-linked AI summary

Child-Adapted Structured Phonological Representations for Interpretable Speech Sound Analysis

Abner Hernandez, Tomás Arias Vergara, Andreas Maier, Paula Andrea Pérez-Toro

arXiv:2610.00852v1cs.CLeess.AS

TL;DR

Adult-focused speech models leave open how structured phonological representations transfer to children, whose speech differs acoustically and developmentally. The study adapts PhonoQ-2.0 under varied alignment supervision and initialization, then evaluates against manual child-speech annotations. Adaptation improves voicing, child-informed supervision benefits manner recognition, and the velar-fronting contrast persists across model variants.

  • Problem

    Most pretrained models and downstream systems are developed primarily from adult speech, although child speech differs in vocal-tract characteristics, acoustic variability, and ongoing phonological development.

    Source: §1 · p. 1

  • Method

    The study adapts PhonoQ-2.0 with Adult, Adult+Child, or Child-only alignment supervision and Adult PhonoQ or scratch initialization, evaluating generalization against manual annotations.

    Source: §1 · p. 1, §2.2 · p. 2

  • Results

    0.804 macro-F1 for Adult+Child MFA (Adult-init); adaptation raises voicing to 0.972–0.987, place remains 0.871–0.902, and velar-fronting is preserved across variants.

    Source: §3.2 · p. 3, §3.3 · p. 3

  • Takeaways & Limitations

    PhonoQ posteriors are structured acoustic–phonological measures, not direct estimates of treatment outcome; longitudinal trajectories broadly follow clinical reports but vary by speaker.

    Source: §4 · p. 4

    The study lacks a common manual alignment reference for CHILDES, and its small clinical cohorts cover only two place-related processes.

    Source: §4 · p. 4

Abstract

from arXiv · show

Structured phonological representations provide an interpretable alternative to generic speech embeddings, but existing models are largely trained on adult speech. We adapt PhonoQ-2.0 to child speech using CHILDES-Aligned data and compare three alignment-supervision conditions (Adult, Adult+Child, and Child-only) across two initialization strategies (Adult PhonoQ and scratch). Generalization is evaluated against manual child-speech annotations. On 1,352 consonant targets from 58 typically developing children, child-speech adaptation improves voicing recognition across all supervision conditions, from 0.922 macro-F1 for Adult PhonoQ to 0.972--0.987 after adaptation. Manner is more sensitive to alignment supervision: Adult+Child MFA reaches 0.804 and 0.796, compared to approximately 0.70 under Adult MFA supervision. Place remains comparatively strong across systems (0.871--0.902), although per-class performance varies substantially. The velar-fronting contrast is preserved across all seven model variants. Longitudinal UltraPhonix analysis further reveals speaker-specific velar and post-alveolar changes that are largely preserved across models and broadly consistent with reported clinical progress.

1. INTRODUCTION

Child speech differs from adult speech, yet most pretrained models and downstream systems are developed primarily on adult data. This work adapts structured phonological recognition to children and evaluates its generalization and preservation of speech-sound contrasts.

  • Adult-focused systems face a transfer challenge because child speech differs in vocal-tract characteristics, acoustic variability, and ongoing phonological development.Source: §1 · p. 1
  • Many child-speech approaches use phoneme recognition, transcription, global assessment scores, or disorder classification rather than representing how individual sounds are realized.Source: §1 · p. 1
  • PhonoQ-2.0 directly predicts structured phonological features, but its English model was trained entirely on adult speech.Source: §1 · p. 1
  • The study adapts PhonoQ-2.0 using CHILDES-Aligned data and compares Adult, Adult+Child, and Child-only alignment supervision with Adult PhonoQ or scratch initialization.Source: §1 · p. 1
  • Manual child-speech annotations test generalization, while place posteriors are examined for velar fronting and longitudinal velar and post-alveolar fronting.Source: §1 · p. 1

2. METHODS

The methods retain PhonoQ-2.0’s structured prediction design while varying alignment supervision and initialization. Independent manual annotations and intervention data assess recognition and place-related contrasts.

2.1. PhonoQ-2.0 Child Adaptation

  • PhonoQ-2.0 uses a wav2vec 2.0 XLSR encoder, a two-layer Conformer, and four heads that produce 22 frame-level phonological outputs.Child adaptation retains the architecture and inventory and uses a speaker-disjoint CHILDES-Aligned subset.Source: §2.1 · p. 2

2.2. Alignment Supervision and Initialization

  • A 3 × 2 design crosses Adult, Adult+Child, and Child-only MFA supervision with Adult PhonoQ or scratch initialization; each alignment source supplies labels for the same adaptation subset.The scratch condition retains the pretrained XLSR encoder but randomly initializes PhonoQ-specific components.Source: §2.2 · p. 2, Figure 1 · p. 2

2.3. Independent Phonological Recognition on UXTD

  • Generalization is tested on 1,352 manually annotated consonant landmarks from 58 UXTD speakers, with manner, place, and voicing evaluated using macro-F1.Posteriors are averaged within ±25 ms of each landmark, and uncertainty uses speaker-cluster bootstrap resampling.Source: §2.3 · p. 2

2.4. Manual Velar-Fronting Analysis

  • Manual /k/ productions from four UXSSD children with velar fronting are compared with those from 10 age-matched UXTD controls using place posteriors averaged within ±50 ms.Speaker median token scores are compared using the between-group median difference, speaker-bootstrap confidence intervals, AUC, and an exact two-sided permutation test.Source: §2.4 · p. 2

2.5. Longitudinal UltraPhonix Analysis

  • Longitudinal analysis uses SLT-annotated velar- and post-alveolar-fronting targets from six and three UltraPhonix participants, tracking target-place separation across four sessions for all seven model variants.Source: §2.5 · pp. 2–3, Table 1 · p. 3

3. RESULTS

Child-informed alignment supervision particularly benefits manner recognition, while adaptation improves voicing and place remains comparatively strong. The velar-fronting contrast persists across model variants, and longitudinal trajectories show speaker-specific changes that vary across children and processes.

3.1. Child-Speech Adaptation

  • With Adult PhonoQ initialization, held-out macro-F1 rises from 0.685 under Adult MFA to 0.761 under Adult+Child MFA and 0.780 under Child-only MFA.Under Child-only supervision, Adult- and Scratch-init scores are nearly identical at 0.780/0.779.Source: §3.1 · p. 3

3.2. External Phonological Recognition on UXTD

  • Adult+Child MFA reaches manner macro-F1 0.804/0.796 and Child-only MFA 0.762/0.775 (Adult-/Scratch-init), versus 0.705 for Adult PhonoQ and 0.698/0.701 for Adult-MFA adaptations.The similar pattern across initializations suggests manner recognition is more associated with supervision source than reuse of Adult PhonoQ-specific parameters.Source: §3.2 · p. 3
  • Place macro-F1 spans 0.871–0.902 without monotonic supervision gains, while voicing rises from 0.922 for Adult PhonoQ to 0.972–0.987 after adaptation.Velar and postalveolar place are consistently strong, whereas labial place is weaker; rhotic and approximant remain difficult manner classes.Source: §3.2 · p. 3
  • Per-class manner gains are uneven, clearest for stops, fricatives, and laterals, while rhotics and approximants remain more challenging.Velar and postalveolar place are consistently strong, while labial place is lower.Source: §3.2 · p. 3

3.4. Longitudinal UltraPhonix Analysis

  • For post-alveolar fronting, 06M shows the clearest longitudinal change: baseline mean ∆ ranges from −1.33 to −0.31 across adapted conditions, increasing to 4.64–6.82 post-treatment and 6.39–8.24 at maintenance.Source: §3.4 · p. 3

3.3. Manual Velar-Fronting Analysis

  • Across all seven variants, UXSSD fronting exceeded UXTD (median difference 1.421–1.769; AUC=1.00; p=0.005), preserving the contrast.All 95% speaker-bootstrap confidence intervals excluded zero, and all four UXSSD scores exceeded those of all 10 UXTD speakers.Source: §3.3 · p. 3, Figure 2 · p. 4

3.4. Longitudinal UltraPhonix Analysis

  • In longitudinal velar-fronting trajectories, 03F and 15M show pronounced later increases in target-place separation, while 11M remains comparatively stable.Figure 3 shows representative trajectories, averaging the two initialization variants within each supervision condition.Source: §3.4 · p. 3, Figure 3 · p. 4
  • Among other longitudinal cases, 04M shows a pronounced later velar increase, whereas 01F and 17M show smaller or less consistent changes.For 19M, an initially negative contrast approaches zero post-treatment and remains comparatively small at maintenance.Source: §3.4 · p. 4

4. DISCUSSION AND CONCLUSION

Child-speech adaptation affects phonological dimensions differently: supervision source is most consequential for manner, while voicing improves across child-adaptation conditions. Longitudinal posteriors capture some speaker-specific changes, but should be read as acoustic–phonological measures rather than treatment-outcome estimates.

4. DISCUSSION AND CONCLUSION

  • Manner is most sensitive to alignment supervision: Adult+Child and Child-only supervision outperform Adult-MFA adaptation on manually annotated UXTD speech.Gains are uneven across classes, with clearer improvements for stops, fricatives, and laterals than for rhotics and approximants.Source: §4 · p. 4
  • Voicing is consistently higher after child-speech adaptation regardless of alignment source, while place remains comparatively strong but varies by feature class.Velar and postalveolar place F1 is consistently high, while labial targets are lower.Source: §4 · p. 4
  • Longitudinal posteriors broadly track reported clinical patterns for some speakers, but 07M’s clinically reported improvement is less clearly reflected in the PhonoQ trajectory.03F and 15M show strong later velar improvement, 11M remains largely stable, and 06M shows pronounced post-alveolar increases after treatment and at maintenance; interpret posteriors as structured acoustic–phonological measures, not direct treatment-outcome estimates.Source: §4 · p. 4
  • The evaluation lacks a common manual CHILDES alignment reference, uses MFA systems trained on different data, and includes small clinical cohorts covering only two place-related processes.The conclusion notes that interpretable place contrasts remain robust across model variants.Source: §4 · p. 4

5. COMPLIANCE WITH ETHICAL STANDARDS

The study reanalyzes previously collected child-speech datasets, with UltraSuite collected under ethics approval and guardian consent for research data sharing.

5. COMPLIANCE WITH ETHICAL STANDARDS

  • The study used previously collected child-speech datasets, with UltraSuite collected under NHS Research Ethics Service approval and guardian consent for research data sharing.No new participants were recruited or data collected; all datasets were used under their applicable access, data-sharing, and licensing conditions.Source: §5 · p. 5
Loading 2610.00852v1…