Notes on Prosody outweighs statistics in 6-month-old German-learning infants' speech segmentation

Introduction

• Topic: Weighting of prosodic cues versus statistical cues (transitional probabilities, TPs) in early speech segmentation

• Age and language context: 6‑month‑old German‑learning infants

• Core idea: Infants use both prosodic information (e.g., lexical/stress cues) and distributional information (TPs) to segment fluent speech. The study tests which cue is weighted more heavily when cues conflict, and whether German‑speaking infants rely on prosody earlier than English‑speaking peers

• Key background: Prior work shows mixed findings on cue dominance and developmental trajectories. English‑learning infants often show stronger reliance on TP early, shifting toward prosody with age (Thiessen & Saffran, 2003; Johnson & Jusczyk, 2001; Thiessen & Erickson, 2013; Johnson & Seidl, 2009). German and Dutch studies suggest strong trochaic prosody and language‑specific prosodic biases from infancy (Friederici et al., 2007; Höhle et al., 2009, 2014; Delattre, 1963).

• Main hypothesis of the current study: German‑learning 6‑month‑olds weight prosodic cues more heavily than TP information when cues conflict, due to language‑specific prosodic structure (trochaic dominance) in German.

• Design overview: Three experiments with the same test phase but different familiarization/test configurations to manipulate TP and prosody cues.

• Implication: Findings inform whether cue weighting is universal or language‑specific, and whether early segmentation relies on prosody or statistics as bootstrapping mechanisms.

Key concepts and definitions

• Prosodic cues: Lexical stress, intonation, rhythm; in German, trochaic (strong‑weak) pattern is dominant for disyllabic words.

• Statistical cues: Transitional probabilities (TPs) between syllables; high TP typically inside a word, low TP at word boundaries.

• Disyllabic words: Units used for testing segmentation, with stress patterns either aligned with or against the TP boundaries.

• Trochaic vs iambic: Trochaic pattern stressed on the first syllable (STRONG‑weak); iambic pattern stressed on the second syllable (weak‑STRONG).

• Headturn Preference Procedure (HPP): Infants’ listening times to test stimuli indicate discrimination or segmentation; longer looking times can indicate novelty or familiarity depending on context.

• Test conditions (Experiment 1): Statistical words (TPs 1.0 inside words), Prosodic words (stress‑aligned but lower within‑word TP), Non‑words (syllables never co‑occurred; TP = 0.0).

• Statistical properties (Table references in the paper): See Table 3 for potential segmentations, Table 4 for test word TP, frequency, and stress in the test phase.

TP values and prosodic properties used

• Within‑word TP: $TP_{ ext{within}} = 1.0$

• Across‑word TP (in familiarization string): $TP_{ ext{across}} ext{ ranges between } [0.2, 0.4]$

• Non‑word TP: $TP_{ ext{non-word}} = 0.0$

• Prosody: Second syllable of four statistical words is consistently stressed (iambic pattern) in the familiarization string

• Test word properties (Experiment 1):

  • Statistical words: puda, bido (TP = 1.0; stress on final syllable in string)

  • Prosodic words: buta, dego (TP ≈ 0.51 and 0.49; stress on initial syllable in string)

  • Non-words: dabi, bide (TP = 0.0; no co‑occurrence in the string; stress patterns not aligned with test words)

• Familiarization string duration and structure: $2 ext{ min }11 ext{ s}$, continuous (no pauses), starter dummy syllable /ke/ to prevent immediate segmentation

Background connections to prior literature

• Saffran, Aslin, & Newport (1996): infants can compute TP across syllables to segment artificial language strings (8‑month English‑learners)

• Johnson & Jusczyk (2001); Thiessen & Saffran (2003): cue collision studies showing prosody can override TP, or TP can dominate depending on age and language; development often shows a shift from statistical to prosodic weighting

• Thiessen & Erickson (2013); Pelucchi et al. (2009): TP-based segmentation in younger infants and cross‑linguistic generalization

• Langus, Nespor, & colleagues (2012–2016): role of prosody in discovering hierarchical structure; prosody can bootstrap word boundaries

• German language properties: strong trochaic dominance; higher proportion of initially stressed words in German; inflectional system promotes trochaic realizations (Delattre, 1963; Höhle et al., 2009, 2014)

• Marimon, Höhle, Langus (2022): individual differences in cue weighting in German‑learning infants at 9 months; context for cross‑linguistic comparisons

EXPERIMENT 1: PROSODY VERSUS STATISTICS

• Objective: Determine how much weight infants give to prosodic cues vs. TP cues when conflicting cues are present

• 2^i design in familiarization: TP words vs. prosodic cues conflict (TPs high inside words; second syllable stressed; iambic pattern)

• Test words (8 items) under three conditions: Statistical words, Prosodic words, Non‑words

2.1 Methods

• 2.1.1 Participants

  • Twenty‑four 6– to 7‑month‑old German monolingual infants (12 girls); mean age ≈ 6;19; range 6;12–7;00

  • 8 additional infants tested but excluded for fussiness, crying, experimenter error, technical problems

  • Ethics: Helsinki guidelines; informed consent; approved by University of Potsdam Ethics Committee

• 2.1.2 Stimuli

  • Familiarization string built from four disyllabic sequences: gobu, tade, bido, puda (statistical words)

  • No real German words formed by syllables; syllables recorded by a female native speaker in infant‑directed style

  • Prosodic manipulation: second syllable of statistical words stressed (iambic pattern); within‑word TP = 1.0; across‑word TP varied 0.2–0.4 to create boundaries

  • Frequency balancing (Aslin et al., 1998): two statistical words (tade, gobu) presented more often (≈ 90 repetitions) than the other two (45 reps each); prosodic words (buta, dego) appeared 46 and 45 times respectively

  • Non‑words (dabi, bide) constructed from syllables never appearing adjacent in string; TP = 0.0; stress not aligned with string context

  • Acoustic properties table (Tables 1–2): details on duration, intensity, F0 for stressed vs. unstressed syllables

• 2.1.3 Procedure

  • Headturn Preference Procedure (HPP) setup

  • Familiarization: continuous string played for 2:11; infant looked toward center light; red side lights flashed; familiarization ended after 2 consecutive looks away or 30s total per segment; string played regardless of infant behavior to ensure consistent exposure

  • Test phase: 12 repetitions per test word per trial; 18 seconds per trial or until look away >2s; trials separated by 800 ms pauses; same string across conditions

  • Test variables: structure to determine whether infants segment based on TP or prosody

• 2.1.4 Data analysis

  • Normality check: Shapiro test; results non‑normal (W ≈ 0.942, p<0.001)

  • Statistical tests: Wilcoxon Signed‑Rank tests (non‑parametric); effect size via Cliff’s delta δ

  • Significance threshold adjusted with Bonferroni (p values reported with correction)

  • Effect size guidelines: δ ≈ 0.147 (small), ≈ 0.33 (medium), ≈ 0.474 (large)

• Table references (for Experiment 1): Table 1 (acoustic properties of stressed/unstressed syllables), Table 3 (possible segmentations based on TP vs prosody), Table 4 (test word properties: TP, frequency, stress)

2.2 Results and discussion

• Looking times (average):

  • Prosodic words: $7.55 ext{ s} ext{ (SE=0.46, SD=4.56)}$

  • Statistical words: $8.52 ext{ s} ext{ (SE=0.47, SD=4.69)}$

  • Non‑words: $8.59 ext{ s} ext{ (SE=0.46, SD=4.54)}$

• Distribution of infants’ longest look by condition: 11 for non‑words, 2 for prosodic words, 11 for statistical words

• Pairwise comparisons (Bonferroni):

  • Statistical vs. Prosodic: $V=230, z=-2.3, p=0.021, δ=0.22$

  • Non‑words vs. Prosodic: $V=72, z=-2.24, p=0.024, δ=0.19$

  • Non‑words vs. Statistical: $V=147, z=-0.07, p=0.94, δ=-0.05$

• Interpretation:

  • Infants listened longer to non‑words and statistical words than to prosodic words, suggesting a novelty effect for the two conditions involving non‑prosodic alignment with the familiarization string.

  • No significant difference between statistical words and non‑words implies TP information may be challenging to track for 6‑month‑old German learners when prosodic cues are present.

  • Prosodic words differed from both non‑words and statistical words, indicating successful segmentation based on prosodic cues.

  • Conclusion: 6‑month‑old German‑learning infants weight prosody more heavily than TP cues in this conflict situation.

• Post‑hoc check: dabi (first syllable stressed in familiarization) vs prosodic words; dabi > prosodic words (p=0.02); but dabi vs other non‑word bide not significantly different (p=0.16). Suggests the effect is not simply due to positional carry‑over of stress.

• Additional interpretation: The results underscore the importance of including a non‑word reference to disentangle the direction of preference and interpret cue weighting accurately.

2.3 Summary of Experiment 1 conclusions

• When prosody and TP cues conflict, 6‑month‑old German‑learning infants robustly rely on prosodic cues for segmentation, over TP cues.

• TP information alone does not reliably drive segmentation at this age in this language context, particularly when natural prosody is present.

• The inclusion of a non‑word test condition provides a crucial reference point for interpreting the direction of infants’ preferences.

EXPERIMENT 2: CONTROL CONDITION WITHOUT FAMILIARIZATION

• Objective: Test whether the preference found in Experiment 1 persists in the absence of familiarization, to rule out inherent preferences for the test words

3.1 Methods

• Participants: Twenty‑seven 6‑ to 7‑month‑old German monolingual infants (14 girls); mean age ≈ 6;20; range 6;15–7;00

• Stimuli and procedure: Identical to the test phase of Experiment 1, but without the familiarization phase

• Ethics and participant allocation as in Experiment 1

3.2 Results and discussion

• Data non‑normal (W ≈ 0.946, p<0.001)

• Looking times (average):

  • Prosodic words: $8.53 ext{ s} ext{ (SE=0.43, SD=4.51)}$

  • Statistical words: $8.83 ext{ s} ext{ (SE=0.48, SD=5.06)}$

  • Non‑words: $8.77 ext{ s} ext{ (SE=0.46, SD=4.79)}$

• Infants’ longest looks: 10 for non‑words, 8 for prosodic, 9 for statistical

• Bonferroni‑corrected pairwise comparisons: no significant differences among any pair (mostly p > .57; δ near 0)

• Conclusion: Without familiarization, infants show no systematic preference among the three test word types, supporting that Experiment 1’s effects arose from processing of the familiarization string rather than inherent preferences for the test items

EXPERIMENT 3: FAMILIARIZATION WITH ONLY STATISTICAL CUES

• Objective: Test whether German‑learning infants can segment based solely on TP information when prosodic cues are absent (synthesized speech used to remove natural prosody cues)

4.1 Methods

• Participants: Thirty‑one 6‑ to 7‑month‑old German monolingual infants (13 girls); mean age ≈ 6;21; range 6;15–7;08

• Stimuli and procedure

  • Syllables are synthesized using MBROLA (Dutoit et al., 1996) but with identical order as Experiment 1; prosodic cues removed

  • Test word naming: puda, bido (statistical words from Experiment 1), buta, dego (prosodic words from Experiment 1), dabi, bide (non‑words, same TP=0.0)

  • TP values: within words = 1.0; across boundaries 0.2–0.4; non‑words = 0.0

• 4.1.2 Procedure: Same as Experiment 1

4.2 Results and discussion

• Data non‑normal (W ≈ 0.918, p<0.001)

• Looking times (average):

  • Part-words (non‑words of Experiment 1 style, i.e., TP‑driven test items) ≈ 6.61 s (SE=0.37, SD=4.12)

  • Statistical words ≈ 6.95 s (SE=0.41, SD=4.54)

  • Non-words ≈ 6.95 s (SE=0.42, SD=4.62)

• Infants’ longest looks: 9 non‑words, 11 part‑words, 11 statistical words

• No significant differences among any pair (all p > .61 after Bonferroni correction; δ ≈ 0)

• Conclusion: When prosodic cues are removed and speech is synthesized to emphasize TP patterns, German‑learning infants do not segment based on TP alone at this age, suggesting limited or no TP‑driven segmentation without natural prosody

GENERAL DISCUSSION

• Main finding: German‑learning 6‑month‑olds rely more on prosodic cues than on TP information when both cues are present and conflict. They also fail to segment when prosody is absent and TP cues are the only cues available

• Cross‑linguistic contrast: English‑learning infants often show a developmental shift from TP dominance to prosody with age (roughly around 8–11 months), whereas German‑learning infants show early reliance on prosodic structure and may not display a robust TP‑driven segmentation even at younger ages

• Potential explanations for cross‑linguistic differences:

  • Stimulus naturalness and acoustic cue strength: The study used natural prosody with stronger cues (e.g., higher average intensity, higher mean F0 for stressed syllables) compared to some TP‑dominant studies that used synthesized prosody; natural prosody may more effectively guide segmentation to prosodic cues in German

  • Prosodic strength in German: German trochaic bias and morphosyntactic inflectional patterns create a robust prosodic landscape that infants can exploit early, potentially reducing reliance on TP cues

  • Syllable duration differences: In German, unstressed syllables can be longer than stressed ones in some contexts, contrasting with English where stressed syllables tend to be longer; such differences may influence how strongly duration cues support lexical stress recognition

  • Language‑specific bootstrapping: The results align with a view that statistical segmentation strategies are not universal; they may be adapted to a language’s phonological and prosodic structure

• Experiment 3 implications: The lack of TP‑driven segmentation in synthesized speech suggests that 6‑month‑old German infants may require language‑specific prosodic knowledge or richer natural prosody to bootstrap segmentation. It also challenges the idea of a universal, language‑independent TP gateway for early segmentation, at least for German

• Developmental trajectory considerations:

  • English‑learning infants show a developmental shift from TP to prosody between 7–11 months (Thiessen & Saffran 2003; Johnson & Jusczyk 2001; Johnson & Seidl 2009)

  • German‑learning infants at 6 months show early prosody weighting; evidence for TP dominance may emerge later (e.g., by 9 months in some German studies) but not in the current 6‑month‑old sample

  • The authors propose that cue weighting may be language‑specific rather than following a universal timeline; future work could examine later ages and additional languages to map development more precisely

• Broader implications for theory and methodology:

  • The role of cue interaction: When cues conflict, prosody can override statistics, suggesting prosodic bootstrapping might be more robust when language prosody is strong

  • The importance of non‑word controls: Including a non‑word condition clarifies interpretation of novelty vs familiarity effects and helps identify true cue weighting

  • Cross‑linguistic corpus considerations: Some languages may rely more on forward TP vs. backward TP or different distributional cues; segmentation strategies may be language‑dependent rather than universal

• Limitations and future directions:

  • The naturalness of stimuli appears to influence results; future studies could compare natural vs. synthesized stimuli more systematically

  • Additional languages with varying prosodic systems could help determine the generalizability of the observed cross‑linguistic difference

  • Longitudinal designs could reveal when and how the shift, if any, occurs in German‑learning infants and whether an eventual TP advantage emerges with more exposure or altered task demands

Practical and ethical notes

• All experiments adhered to the Declaration of Helsinki; informed consent obtained; ethics approval documented

• Data and stimuli available at: https://osf.io/4g7yr/

• The study uses a mix of natural and synthesized stimuli to investigate cue weighting; acknowledges potential differences in cue salience between natural and synthesized speech

Connections to foundational ideas and real‑world relevance

• Demonstrates that early speech segmentation is sensitive to the language’s phonological and prosodic properties, not just abstract statistical regularities

• Supports the notion that language acquisition is shaped by language‑specific cues and experience, aligning with theories of bootstrapping that emphasize perceptual grouping biases and prosodic structuring

• Shows that infants’ segmentation strategies cannot be assumed to be language‑neutral or universally TP‑driven; educational tools and early language interventions should consider language prosody and stress patterns

Key numerical and methodological references (selected)

• TP values and word boundary design in Experiment 1: $TP<em>{ ext{within}} = 1.0,\quad TP</em>{ ext{across}} ext{ in } [0.2, 0.4],\quad TP_{ ext{non-word}} = 0.0$

• Test phase durations and trial structure: test trial length = $18 ext{ s}$, interval between repetitions = $800 ext{ ms}$

• Acoustic properties (stressed vs. unstressed) summarized in Tables 1 and 2

• Test word properties (Table 4): includes frequency of occurrence in the string and TP values for each test word

• Statistical tests: Wilcoxon Signed‑Rank tests; effect size via Cliff’s delta $ext{δ}$ with interpretation thresholds as provided (small ≈ 0.147, medium ≈ 0.33, large ≈ 0.474)

• Non‑parametric approach due to non‑normal listening time distributions (Shapiro test results reported in each experiment)

Summary at a glance

• Experiment 1: Prosody dominates over statistics in 6‑month‑old German infants when cues conflict; non‑word condition provides a reference that clarifies cue weighting; prosodic segmentation achieved; TP alone not robust

• Experiment 2: No inherent preference without familiarization; supports that Experiment 1 results are due to exposure to the familiarization string

• Experiment 3: No TP‑only segmentation with synthesized speech; pro‑prosody absence eliminates segmentation cue; supports language‑specific cue weighting and natural prosody’s role

• Overall conclusion: For German‑learning infants at 6 months, prosodic structure appears to guide early segmentation more than distributional statistics; cross‑linguistic differences suggest cue weighting is language‑specific and shaped by the ambient language’s prosody

The results of the study have several key implications:

1. Language-Specific Cue Weighting: The primary implication is that early speech segmentation is not universally driven by the same cues. German-learning 6-month-olds prioritize prosodic cues over statistical (Transitional Probability, TP) information when these cues conflict, suggesting that cue weighting is language-specific and shaped by the ambient language's phonological and prosodic structure.

2. Challenge to Universal TP Gateway: Experiment 3, where infants did not segment based on TP alone in synthesized speech, directly challenges the idea of a universal, language-independent TP mechanism for early segmentation. It suggests that for German-learning infants at this age, natural prosody or language-specific prosodic knowledge is crucial for bootstrapping segmentation.

3. Cross-Linguistic Differences in Development: Unlike English-learning infants who often show an initial reliance on TP shifting to prosody later, German-learning infants demonstrate early prosody weighting. This implies that developmental trajectories for acquiring speech segmentation can vary significantly across languages, without a single universal timeline.

4. Importance of Stimulus Naturalness: The findings highlight that the naturalness of speech stimuli (e.g., natural vs. synthesized prosody) can profoundly influence cue salience and, consequently, the segmentation strategies infants employ. Natural prosody with stronger acoustic cues may guide segmentation more effectively.

5. Prosody as a Robust Bootstrapping Mechanism: For languages with strong prosodic regularity, such as German's trochaic bias, prosodic cues appear to be a more robust bootstrapping mechanism than statistical cues, especially when cues collide.

6. Methodological Considerations: The study underscores the critical importance of including non-word control conditions in infant speech perception research. These controls help researchers accurately interpret novelty versus familiarity effects and disentangle true cue weighting biases.

7. Real-World Relevance: These findings support theories of language acquisition that emphasize the role of language-specific experience and perceptual grouping biases. For educational tools and early language interventions, this means considering the specific prosodic and stress patterns of a language is important, as infants' segmentation strategies are not language-neutral.