Learning Sound Patterns I

Sound Pattern Acquisition and the Challenges of Early Language Perception

• The Problem of Continuous Acoustic Input: Infants face significant perceptual hurdles when acquiring language from acoustic signals:

  • Absence of Acoustic Gaps: Spoken language contains no physical pauses or gaps between words, unlike spaces in written text.

  • Lack of Abstract Concepts: Infants possess no initial conceptual understanding of what a "word" is.

  • Phonetic Variability: Speech sounds vary considerably depending on speaker voice, pitch, context, and time of production.

  • Contrastive Identification: Infants must determine which phonetic differences carry meaningful semantic distinction (e.g., $/p/$ versus $/b/$) and which variations are non-contrastive.

  • Sensory Overload: William James (1890) famously described early perceptual experience: "The baby, assailed by eyes, ears, nose, skin, and entrails at once, feels it all as one great blooming, buzzing confusion."

• Timeline of Early Perceptual and Acoustic Achievements:

  • In Utero Perception: Fetal auditory experience captures the rhythmic properties, prosody, and melodic contours of native speech.

  • At Birth (Within $4$ days): Newborns discriminate their native language from unfamiliar languages. For example, French newborns exhibit higher non-nutritive sucking rates when listening to French compared to Russian (Mehler et al., 1988).

  • Pre-Verbal Auditory Segmentation: Long before productive word speech emerges (around $12$\,months), infants learn to parse continuous speech into distinct sound units, categorize contrastive phonemes, and learn native phonotactic probabilities.

Infant Speech Segmentation Strategies

• The Necessity of Words (Lexical Units):

  • Without segmentation into discrete words, every distinct utterance would require memorization as a holistic sound block (e.g., mapping a unique unanalyzed complex sound to "the dog bit the dog").

  • Words enable infinite recombination and linguistic productivity (Hockett).

• Acoustic Realities of Word Boundaries:

  • Physical silence does not mark word boundaries in continuous speech streams.

  • In acoustic spectrograms (e.g., for the phrase "nineteenth century"), the primary silent gap occurs internally during the stop closure of the $t$ sound in "nineteenth", whereas zero acoustic gap exists between the words "nineteenth" and "century".

• Head-Turn Preference Paradigm (HTPP):

  • Experimental Design: Used with infants aged $4.5$\,to $18$\,months (older toddlers fail to sit still).

  • Familiarization Phase: Infants listen to specific auditory passages or artificial streams while looking at central/side lights.

  • Test Phase: Familiar versus novel auditory stimuli are played from directional speakers. Looking time toward the speaker is measured.

  • Interpretation: Statistically significant differences in orientation duration (looking time) indicate successful stimulus discrimination and auditory memory categorization.

• Empirical Demonstration of Word Extraction:

  • Jusczyk & Aslin (1995): Tested whether infants can isolate individual words from fluent passages.

  • Method: Exposed infants to passages containing target words (e.g., "His bike had big black wheels. The girl rode her big bike…"). Subsequently tested isolated target words ("bike") versus novel words ("dog").

  • Findings: $7.5$-month-old infants listened significantly longer to isolated familiar words, demonstrating successful extraction from continuous speech streams. $6$-month-old infants failed this task.

• Segmentation Mechanisms and Environmental Cues:

  • Familiar Word Anchors:

    • Approximately $10\%$ of infant-directed speech consists of isolated single words (e.g., "Mommy", "Ella").

    • Infants leverage highly familiar anchor words to segment adjacent novel material (e.g., breaking bankiritubendudifin into ban-kiri-tuben-dudi-fin).

    • Bortfeld et al. (2005): Showed $6$-month-old infants segment unfamiliar target words (e.g., "feet" or "bike") when placed adjacent to familiar name anchors (e.g., "Maggie's bike…", "Mommy's feet").

  • Phonotactic Constraints:

    • Definition: Language-specific rules defining permissible sequence combinations of sound segments at word beginnings, endings, or internal transitions.

    • English Example: Consonant cluster $nd$ occurs internally (bandage) or word-finally (hand), but cannot occur word-initially. Thus, a continuous sequence like Elgandoku is parsed by native English speakers as Elgan doku rather than Elga ndoku.

    • Cross-Linguistic Variations: Word-initial $skrs$ is legal in Czech; word-initial $nd$ is legal in Swahili (ndela); English word-initial $sp$ clusters are illegal in Spanish (requiring prosthetic vowel additions like espanish).

    • Developmental Timeline: Jusczyk et al. (1993) demonstrated that by $9$\,months, American infants prefer legal English nonwords (cubeb, dudgeon) over legal Dutch nonwords (zampljes, vlatke), while Dutch infants show inverse preferences. Infants utilize these constraints for word segmentation by $9$\,months (Mattys & Jusczyk, 2001).

  • Metrical Stress Patterns (Prosody):

    • English stress is predominantly trochaic (strong-weak/stressed-unstressed pattern, e.g., DOC-tor) rather than iambic (weak-strong/unstressed-stressed pattern, e.g., gui-TAR), appearing at a ratio of approximately $9\text{ to }1$.

    • Jusczyk et al. (1999): $7.5$-month-old English-learning infants correctly segment trochaic words (DOC-tor) from continuous speech, but missegment iambic target phrases like "guiTAR is" as TAR-is.

Statistical Learning and Transitional Probabilities

• The Bootstrap Dilemma: Relying exclusively on phonotactic rules or stress patterns creates a circular dependency: discovering structural rules requires a pre-existing vocabulary, yet constructing a vocabulary requires segmentation tools.

• Mathematical Definition of Transitional Probability (TP):

  • Transitional probability measures the statistical likelihood of syllable $Y$ following syllable $X$:

$TP = P(Y \mid X) = \frac{\text{frequency}(XY)}{\text{frequency}(X)}$

• TP Properties across Lexical Boundaries:

  • Within Words: Syllables internal to a word follow each other with high predictability ($TP \rightarrow 1.0$). For instance, inside the word pretty, the probability of ty following pre is high ($85\%$).

  • Across Word Boundaries: Syllables spanning word boundaries are significantly less predictable ($TP$ drops dramatically). The syllable ty at the end of pretty can precede any initial syllable of subsequent words (e.g., bay, bo, nay), yielding low individual transition probabilities ($5\%\text{ to }15\%$).

• Landmark Study: Saffran, Aslin, & Newport (1996):

  • Artificial Language Construction: Designed a controlled language comprising four 3-syllable nonsense words: bidaku, padoti, golabu, tutaba (or tupiro).

  • Auditory Familiarization Phase: $8$-month-old infants were presented with a continuous $2$-minute stream of synthesized computer speech containing each word repeated $45$\,times in random order (bidakugolabudutabagolabupadotibidakudutabapadoti…).

  • Experimental Controls: Flat monotonic voice, absence of pauses, uniform pitch, absence of stress variations, uniform consonant-vowel (CV) syllable structure. This eliminated all prosodic, acoustic, and phonotactic cues, isolating statistical patterns.

  • Statistical Parameters: Within-word $TP = 1.0$; across-word-boundary $TP = 0.33$.

  • Testing & Results: HTPP evaluated listening times for intact "words" (bidaku, golabu) versus cross-boundary "part-words" (dakugo, buduta).

  • $8$-month-olds demonstrated significantly longer mean looking times for part-words ($7.60$\,seconds) compared to intact words ($6.77$\,seconds).

  • Conclusion: Infants detect word boundaries based strictly on tracking conditional statistical probabilities between adjacent speech units.

Domain Generality and Evolutionary Implications of Statistical Learning

• Cross-Species Statistical Tracking:

  • Cotton-Top Tamarins (Saguinus oedipus): Hauser et al. (2001) exposed tamarins to $20$\,minutes of Saffran's artificial language stream. Tamarins exhibited significantly greater orienting head-turns toward speakers broadcasting part-word sequences.

  • Rats (Rattus norvegicus): Toro & Trobalón (2005) established that laboratory rats successfully discriminate words from part-words using transitional probabilities despite lacking language processing modules.

• Cross-Modal and Early Developmental Pervasiveness:

  • Non-Linguistic Modalities: Humans apply identical statistical tracking mechanisms to musical tone sequences (Saffran et al., 1999) and visual geometric shapes presented in temporal streams (Fiser & Aslin, 2001).

  • Neonatal Statistical Learning During Sleep: Teinonen et al. (2009) played $15$\,minutes of tri-syllabic artificial words to sleeping newborns under $48$\,hours old. Event-Related Potentials (ERPs) revealed enhanced neural responses to unpredictable word-initial syllables relative to predictable internal syllables.

• Theoretical Implications:

  • Statistical learning represents an ancient, domain-general cognitive computation mechanism rather than a human-unique, language-specific module.

  • Species Differences & Human Biases: Non-human species (e.g., rats) miss subtle prosodic/perceptual cues that human infants automatically integrate, and humans possess distinct perceptual biases regarding which structural patterns are easily learned.

Concept Verification and Empirical Applications

• Contextual Word Extraction Check:

  • Scenario: A $6$-month-old infant hears the string "Look at Mommy's book".

  • Inference: The infant is most likely to infer that "book" is a distinct word unit because it follows the highly familiar, pre-stored anchor word "Mommy's".

• Phonotactic Parsing Check:

  • Scenario: English speakers presented with the novel string Elgandoku parse it as Elgan doku instead of Elga ndoku.

  • Mechanism: Native English phonotactic rules disallow word-initial $nd$ clusters.

• Methodological Isolation Control in Saffran et al. (1996):

  • Scenario: The artificial speech stream was delivered in a monotonic synthetic voice without pauses or stress.

  • Rationale: Eliminates prosodic, acoustic, and metrical cues, ensuring that transitional probability is the sole variable driving segmentation performance.

• Transitional Probability Calculations in Continuous Speech:

  • Phrase: "the-ve-ry-red-car"

  • Highest TP Pair: The syllable transition ve $\rightarrow$ ry possesses the highest transitional probability because it occurs internally within the word very ($TP \rightarrow 1.0$), whereas ry $\rightarrow$ red spans a word boundary ($TP \ll 1.0$).

• Cross-Species Generalizability:

  • Finding: Laboratory rats successfully differentiate words from part-words based on statistical regularities.

  • Conclusion: Proves that transitional probability tracking is neither unique to human beings nor specific to linguistic domains.