Learning Sound Patterns II

Overview and Logistics

• Assignment 1 submission has been completed, with grades anticipated to be returned in approximately $2\,\text{weeks}$.

• Required reading for the upcoming topic includes Chapter 5, Sections 5.1–5.3 (pp. 157–181) covering Learning Words I, focusing on:

  • The overall scale of word learning.

  • Quine's gavagai problem of inductive uncertainty in reference.

  • Words acting as an interface to acoustic representations, semantic reference, and conceptual structures.

• Midterm Exam 1 is scheduled for Tuesday, September 29, covering course material from Weeks 1 through 5.

Acoustic Foundations: Phonemes vs. Allophones

• The English alphabet contains $26\,\text{letters}$, but Standard American English utilizes a significantly larger set of distinct speech sounds.

  • The letter "G" produces completely different sounds in words like gap versus giraffe.

  • Acoustic physical differences exist even for the same perceived letter segment; for example, the physical production of the sound segment [t] in tan differs from the [t] in Stan.

  • Holding a palm directly in front of the mouth reveals a noticeable burst of expelled air (aspiration) when pronouncing tan, which is absent when pronouncing Stan.

• Scrambled speech swap manipulations demonstrate that swapping identical target letters across different words produces physical acoustic differences that listeners often normalize or fail to consciously identify.

• Phonemes:

  • Abstract, mental categories of speech sounds stored in the mind.

  • Sound alterations that produce a change in lexical meaning correspond to distinct phonemes.

  • Example: Replacing the initial sound of ten to create den alters the word's meaning.

  • Phonemic units are formally transcribed using forward slashes: /t/, /d/.

• Allophones:

  • Surface-level, physical variations of a single underlying phoneme.

  • Sound alterations that do not alter the lexical meaning of a word represent allophones of the same phoneme.

  • Example: Substituting the aspirated [tʰ] from tan into Stan produces an unusual pronunciation but does not transform the word into a different lexical item.

  • Allophonic variations can be just as physically and acoustically distinct as separate phonemes.

  • Allophonic variations are formally transcribed using square brackets: [t], [tʰ].

• Minimal Pairs:

  • A minimal pair consists of two words that differ by exactly one single sound segment in the same position.

  • Minimal pairs serve as the diagnostic test to determine whether a sound contrast is phonemic or allophonic:

  • If changing a single sound segment alters word meaning, the distinction is phonemic.

  • If changing a single sound segment does not alter word meaning, the distinction is allophonic.

  • Minimal pair examples:

  • tan vs. man (differing initial segment changes meaning: /t/ vs. /m/).

  • tan vs. ten (differing vowel segment changes meaning: /æ/ vs. /ɛ/).

  • tan vs. tad (differing final segment changes meaning: /n/ vs. /d/).

Articulatory Phonetics: Categorizing Consonants

• Linguists categorize consonant speech sounds across three fundamental articulatory dimensions:

  1. Place of Articulation: The specific anatomical location in the vocal tract where airflow obstruction or restriction occurs.

  • Bilabial: Restricting airflow using both lips (e.g., [p], [b], [m], [w], [ʍ]).

  • Labiodental: Restricting airflow using the lower lip and upper teeth (e.g., [f], [v]).

  • Interdental: Placing the tongue tip between the upper and lower teeth (e.g., [θ], [ð]).

  • Alveolar: Restricting airflow by placing the tongue against the alveolar ridge directly behind the upper front teeth (e.g., [t], [d], [s], [z], [n], [l], [ɹ]).

  • Alveo-palatal: Positioning the tongue in the area just behind the alveolar ridge (e.g., [ʃ], [ʒ], [ʧ], [ʤ]).

  • Palatal: Raising the tongue body against the hard palate (e.g., [j]).

  • Velar: Raising the back of the tongue against the soft palate or velum (e.g., [k], [g], [ŋ]).

  • Glottal: Restricting airflow at the vocal folds/glottis (e.g., [ʔ], [h]).

  1. Manner of Articulation: The precise spatial configuration and physical nature of the airflow obstruction.

  • Stop (Plosive): Airflow is completely blocked and subsequently released as a rapid burst (e.g., [p], [b], [t], [d], [k], [g], [ʔ]).

  • Fricative: Airflow is squeezed through a narrow anatomical channel, producing continuous acoustic friction or turbulence (e.g., [f], [v], [θ], [ð], [s], [z], [ʃ], [ʒ], [h]).

  • Affricative: Starts as a complete stop closure immediately followed by a fricative release at the same place of articulation (e.g., [ʧ], [ʤ]).

  • Nasal: The velum is lowered, completely closing off oral airflow and allowing air to escape through the nasal cavity (e.g., [m], [n], [ŋ]).

  • Lateral Liquid: Airflow passes around one or both sides of the tongue (e.g., [l]).

  • Retroflex Liquid: The tongue tip is curled upward and backward toward the hard palate (e.g., [ɹ]).

  • Glide: Articulators undergo a smooth, unobstructed movement similar to a vowel transition (e.g., [ʍ], [w], [j]).

  1. Voicing: Whether the vocal folds inside the larynx vibrate during the production of the sound.

  • Voiceless: Vocal folds remain held apart; air flows freely through the larynx without vibration (e.g., [s], [f], [p], [t], [k]).

  • Voiced: Vocal folds are drawn together and vibrate rapidly due to subglottal air pressure (e.g., [z], [v], [b], [d], [g]). Placing fingers directly over the larynx provides physical feedback of this vibration (e.g., comparing continuous sss vs. zzz).

Voicing, Voice Onset Time (VOT), and Aspiration

• Consonants are distinguished by their voicing state. Vowels are typically voiced sound categories.

• For oral stop consonants (e.g., /p/ vs. /b/), voicing is operationalized as Voice Onset Time (VOT):

  • VOT is defined as the temporal delay between the articulatory release of an oral stop obstruction and the onset of vocal fold vibration.

  • If vocal fold vibration begins within approximately $20\,\text{ms}$ following stop release, the segment is classified as a voiced stop (e.g., [b]).

  • If the delay between stop release and vocal fold vibration extends significantly past $20\,\text{ms}$, the segment is classified as a voiceless stop (e.g., [p]).

!Waveforms displaying voice onset time (VOT) and aspiration duration for 'bought' [b] versus 'pot' [pʰ]

• Aspiration:

  • Aspiration refers to an audible acoustic burst of air that accompanies the release of certain unvoiced stop consonants (e.g., [t] vs. [tʰ]).

  • Aspiration is produced by an extra-long Voice Onset Time, extending up to $80\,\text{ms}$.

  • The physical acoustic difference between unextracted [t] and aspirated [tʰ] ($\approx 60\,\text{ms}$ VOT difference) is equal in magnitude to the acoustic distance separating /t/ and /d/ ($\approx 20\,\text{ms}$ vs. $80\,\text{ms}$ VOT).

  • Although the acoustic gap for aspiration [t] vs. [tʰ] is identical in size to the voicing gap /t/ vs. /d/, native English speakers perceive /t/ vs. /d/ as distinct phonemes while treating [t] vs. [tʰ] as equivalent allophones.

  • This demonstrates that human speech perception is not merely a reflection of raw sensory distance, but is modulated by functional categories within a language system.

Cross-Linguistic Phonemic Variability

• Phonemes are not universally fixed across human languages; different languages partition the continuous acoustic space in distinct ways.

• Comparison of phonetic distinctions across select languages:

  • Voicing Contrast (/t/ vs. /d/, VOT $0\,\text{ms}$ vs. $20+\text{ms}$):

  • English: Functions as separate phonemes (e.g., tie vs. die).

  • Thai: Functions as separate phonemes.

  • Mandarin: Functions as allophones.

  • Aspiration Contrast ([t] vs. [tʰ], VOT $20\,\text{ms}$ vs. $80\,\text{ms}$):

  • English: Functions as allophones (e.g., Stan [t] vs. tan [tʰ]).

  • Thai: Functions as separate phonemes.

  • Mandarin: Functions as separate phonemes.

  • Pitch/Tone Modulation on Vowels:

  • English: Functions as allophones (used for intonation/pragmatics, not core lexical meaning).

  • Thai: Functions as separate phonemes (tonal language distinction).

  • Mandarin: Functions as separate phonemes (tonal language distinction).

  • Liquid Contrast (/r/ vs. /l/):

  • English: Functions as separate phonemes (e.g., rice vs. lice).

  • Japanese: Functions as allophones of a single liquid category.

  • Korean: Functions as allophones of a single liquid category.

Phonetics vs. Phonology

• Phonetics:

  • The study of the physical, physiological, and acoustic properties of human speech sounds.

  • Provides universal descriptions applicable across all human languages.

• Phonology:

  • The study of the abstract cognitive systems, structural patterns, and formal rules that organize speech sounds within a specific language.

  • Highly language-specific.

Development of Speech Perception in Infants

• High-Amplitude Sucking (HAS) Paradigm (Eimas et al., 1971):

  • Measures an infant's sucking behavior on a pacifier connected to a pressure sensor to assess perceptual discrimination.

  • Procedure: Repeatedly present a sound stimulus (e.g., [pa]) to an infant. Sucking rate initially rises, then systematically declines as the infant habituates to the repeated stimulus. Once habituated, a novel sound stimulus (e.g., [pʰa]) is presented.

  • Inference: If the infant's sucking rate increases (dishabituation) upon hearing the new sound, it demonstrates that the infant perceives a difference between the two sound categories.

  • Findings: Human newborns exhibit discrimination capabilities for both voicing contrasts ([ba] vs. [pa]) and aspiration contrasts ([pa] vs. [pʰa]), regardless of the language spoken in their home environment.

• Universal Acoustic Sensitivity in Infants:

  • Newborn infants easily discriminate fine acoustic contrasts that native adult speakers of other languages cannot perceive, such as:

  • Hindi plain dental /t/ vs. retroflex /ʈ/ (Werker & Tees, 1984).

  • Czech raised fricative ř vs. postalveolar ž.

  • French oral vs. nasalized vowels.

• Perceptual Narrowing (Retuning):

  • By approximately $10\text{ to }12\,\text{months}$ of age, infants cease discriminating phonetic contrasts that are not present in their ambient language environment, responding only to native phonemic distinctions.

  • This developmental change is not sensory loss or a decline in peripheral auditory acuity.

  • It represents cognitive perceptual narrowing: a functional retuning of selective attention toward category distinctions relevant to the native language.

  • Adults retain raw sensory capability to detect physical sound differences, but the auditory system ignores non-phonemic differences for higher-level linguistic processing (analogous to recognizing a colleague's identity while ignoring day-to-day changes in clothing).

Structural Sound Patterns: Complementary Distribution and Assimilation

• The Acquisition Puzzle:

  • Infants begin organizing and sorting sound categories around $6\,\text{months}$ of age, prior to building a vocabulary.

  • Because infants do not yet possess semantic word knowledge, they cannot rely on minimal pairs (semantic differences) to learn which sound contrasts matter in their language.

• Complementary Distribution:

  • Allophones of a phoneme typically occur in complementary distribution: they appear in mutually exclusive phonetic environments, where the presence of one allophone precludes the presence of another.

  • Distribution of voiceless stop allophones in English:

  • Aspirated stops ([pʰ], [tʰ], [kʰ]) appear word-initially or at the start of stressed syllables (e.g., pit, cat, baTTAlion, comPAssion).

  • Unaspirated stops ([p], [t], [k]) appear directly after a consonant (e.g., spit, scream) or at the start of unstressed syllables (e.g., CAnter, AMputate, elLIPtical).

  • Diagnostic Rule: If the choice between two sounds is entirely predictable based on surrounding phonetic context, the acoustic variation carries no independent semantic information. Therefore, the sounds are complementary allophones regulated by physical production factors, rather than distinct phonemes.

• Assimilation:

  • Complementary distributions frequently arise due to assimilation: the process by which a speech sound modifies its physical articulation to become more similar to neighboring sounds, minimizing articulatory effort.

  • Example: Articulation of the voiceless velar stop /k/ in English:

  • Palatal stop [c] is produced before front vowels (/e/, /ɪ/, /i/): cape, kill, keel, kick.

  • Velar stop [k] is produced before back vowels (/ɑ/, /o/, /u/): con, cope, can, coop.

Learning Allophones via Statistical Distribution: White et al. (2008)

• Research Question: Can infants use contextual statistical distribution (complementary vs. overlapping contexts) to infer whether two sounds are allophones or distinct phonemes without knowing word meanings?

• Experimental Design:

  • White et al. (2008) exposed infants to artificial language sequences consisting of two-word combinations containing precise distributional rules:

  • Stop consonants (p, b, t, d) underwent contextual assimilation: voiceless stops occurred only after rot (rot pevi, rot pogu, rot tula, rot tizu), whereas voiced stops occurred only after na (na bevi, na bogu, na dula, na dizu).

  • Fricative consonants (f, v, s, z) did not assimilate: both voiced and voiceless fricatives occurred freely following both rot and na (rot zuma, na suma, rot suma; rot zobi, na sobi, rot sobi; rot veda, na feda, rot feda; rot vadu, na fadu, rot fadu).

• Target Structural Deductions:

  • For stops ([b] vs. [p]): Present in complementary distribution $\rightarrow$ Choice is predictable from context $ ightarrow$ Allophones of a single phoneme. Words like bevi and pevi represent contextual variants of the same word.

  • For fricatives ([s] vs. [z]): Present in overlapping distribution $\rightarrow$ Choice is unpredictable from context $ ightarrow$ Separate phonemes. Words like savi and zavi represent two distinct words.

• Methodology & Results:

  • Evaluated using the Head-Turn Preference Paradigm with $12\,\text{month-old}$ infants.

  • After familiarization, infants listened significantly longer to allophonic alternations (poli / boli) compared to phonemic alternations (zadu / sadu).

  • Conclusion: $12\,\text{month-old}$ infants track environmental sound distributions to determine which sound differences are meaningful (phonemes) and which are conditioned non-meaningful variations (allophones).

Constraints and Learning Biases in Sound Pattern Acquisition

• Human languages do not form arbitrary or random sound rules (e.g., a rule stating that words starting with /k/ must end with /m/).

• Sound rules in natural human languages systematically target natural classes.

• Natural Class: A set of speech sounds that share one or more distinctive phonetic features to the exclusion of all other sounds in that language (e.g., the set of all voiceless stops [p, t, k], or all voiced fricatives [v, z, ʒ]).

• Experimental Evidence for Natural Class Learning Biases:

  • Saffran & Thiessen (2003):

  • Tested $9\,\text{month-old}$ infants on artificial sound rules.

  • Infants successfully learned a pattern operating over a natural class (all voiceless stops), but failed to learn an identical pattern applied to an unnatural, arbitrary set of sounds.

  • Linzen & Gallagher (2017):

  • Exposed adult participants to a language rule where words began with a subset of voiceless stops.

  • After only $2\,\text{exposures}$, adult learners generalized the rule to all unencountered voiceless stops, demonstrating an implicit expectation that sound rules apply across natural classes.

Theoretical Implications: Nativist vs. Anti-Nativist Perspectives

• The existence of learning biases for natural classes informs theoretical debates on language acquisition:

  • Nativist Position:

  • Argues that learning biases stem from innate, domain-specific knowledge or genetic constraints specialized for language acquisition.

  • Suggests that human genetics hardwire an abstract linguistic architecture that constrains permissible sound systems.

  • Anti-Nativist Position:

  • Argues that learning biases reflect general cognitive, perceptual, and processing constraints operating across cognitive domains.

  • Suggests that languages adapt over time to fit domain-general human cognitive and perceptual limitations, making natural patterns easier to process and transmit.

Conceptual & Comprehension Checks

• Scenario 1: Language X

  • Condition: Swapping [s] for [ʃ] never changes word meaning, but swapping [s] for [z] does.

  • Correct Conclusion: [s] and [ʃ] are allophones of one phoneme; /s/ and /z/ are separate phonemes.

• Scenario 2: Cross-Linguistic Perception (Mandarin vs. English)

  • Condition: Mandarin speakers readily notice [t] vs. [tʰ] but struggle to discriminate [t] vs. [d]; English speakers display the reverse pattern.

  • Correct Conclusion: Which physical acoustic differences sound obvious depends directly on which contrasts are phonemic in a speaker's native language.

• Scenario 3: Language Y

  • Condition: [s] appears exclusively before vowels, and [ʃ] appears exclusively before consonants, making their occurrence fully predictable from context.

  • Correct Conclusion: [s] and [ʃ] are allophones of one phoneme. If a sound choice is fully fixed by context, it cannot carry independent semantic information.

• Scenario 4: Artificial Language Discrimination

  • Condition: An infant hears an artificial language where [g] occurs only after na, and [k] occurs only after rot, while [v] and [f] appear freely after both.

  • Correct Conclusion: The infant should treat vadu / fadu as two different words because [v] and [f] occur in overlapping distribution, indicating phonemic status. Conversely, goli / koli are in complementary distribution, indicating allophonic variants of one word.

• Scenario 5: Generalization Biases

  • Condition: Adult learners hear a made-up language where words begin with /k/, /t/, or /f/ (voiceless consonants).

  • Correct Conclusion: When presented with novel items, adults accept soma (initial voiceless fricative /s/) more readily than zoma (initial voiced fricative /z/), because /s/ shares the voiceless natural class feature.

Discussion Questions & Synthesis

• Question 1: Could any arbitrary sound pattern function as part of a human language? Would some patterns be inherently better or worse?

  • Synthesis: Arbitrary patterns (such as linking initial /k/ to final /m/) are harder to process and learn. Human speech systems favor patterns operating on natural classes and ease of articulation (assimilation).

• Question 2: Does empirical evidence for natural class learning biases favor nativist or anti-nativist theories?

  • Synthesis: Nativists interpret learning biases as evidence for innate universal grammar constraints. Anti-nativists interpret biases as domain-general perceptual constraints to which language systems evolutionary adapt.