Human speech is a continuous, highly variable physical signal produced by the fluid movement of vocal organs. Yet, human language and speech recognition systems organize this fluid stream into discrete, symbolic categories—such as voiced vs. unvoiced, stop vs. fricative, or monophthong vs. diphthong.
This mismatch between continuous acoustics and discrete mental representations creates a fundamental challenge: category ambiguity. Across many languages and contexts, certain speech sounds sit precisely on the knife's edge between categories, defying binary classification. Understanding why these ambiguities occur, and how the brain and computational models resolve them, requires moving away from static boxes and embracing gradient, dynamic representations.
The Illusion of Binary Boundaries
Linguistic frameworks often force researchers to make binary choices about the structure of a sound. Two of the most common structural ambiguities are deciding whether a sound is a single complex unit or a sequence of two sounds, and whether a feature is a standalone segment or a "suprasegmental" layer spread across a syllable.
Units vs. Sequences
Consider the affricate (like the "ch" sound in church). Is it a single phoneme, or a rapid sequence of a stop followed by a fricative? Julia Hoelterhoff demonstrates this tension in German obstruents, noting that the linguistic status of affricates remains an open question that directly impacts how automatic speech recognition systems must extract features (Recognition of German obstruents).
A similar ambiguity plagues vowel systems. In the Taiyuan Jin Chinese dialect, Liping Xia et al. show that falling diphthongs behave like spectrally dynamic single units, whereas rising diphthongs function more like sequences of distinct vowels (Vowels and Diphthongs in the Taiyuan Jin Chinese Dialect). Forcing both into a single "diphthong" category obscures their underlying phonetic reality.
Segments vs. Suprasegmentals
When a sound carries a distinct quality like nasalization or glottalization, does that quality constitute its own consonant segment, or is it a suprasegmental feature coloring the entire vowel? In the Amazonian language Korebaju, prior research has been deeply divided. Jenifer Vega Rodríguez documents how this lack of consensus leaves the language's phonemic inventory poorly defined (The Vowel System of Korebaju).
To resolve why this ambiguity exists, Jenifer Vega Rodriguez et al. utilized electroglottographic (EGG) signaling to look directly at vocal fold physics, revealing that glottalization dynamically shifts between functioning as a discrete glottal stop segment and an acoustic vowel quality (creaky voice) depending on context (Glottal Sounds in Korebaju).
Hybrid Articulations and Intermediate States
Ambiguity also arises because the physical vocal tract can produce stable, intermediate articulations that sit squarely between traditional phonetic categories.
Stops vs. Fricatives
A stop consonant requires complete closure of the vocal tract, while a fricative requires a narrow constriction that generates turbulent noise. But what happens when a sound does both poorly? Mary Stevens et al. investigate "semi-fricatives" in Sienese Italian—sounds characterized by a closure followed by a weak, burstless release (Spirantization of /p t k/ in Sienese Italian and so-called semi-fricatives). Rather than being accidental "weakened" variants, these are stable, intermediate acoustic states. Similarly, John Hajek et al. show that ejective stops in Waima'a cannot be cleanly split into discrete "strong" or "weak" binary classes, pointing instead to a continuous spectrum of glottalic pressure (On the acoustic characterization of ejective stops in Waima'a).
Vowels vs. Fricatives
Vowels are typically open and resonant, but "apical vowels" (or syllabic fricatives) in Mandarin challenge this definition. They share their physical articulation with preceding fricative consonants but serve as the acoustic nucleus of a syllable. Jingyi Sun et al. show that these sounds are acoustically much shorter than regular vowels due to the aerodynamic difficulty of sustaining simultaneous voicing and friction (Apical vs. Regular Vowel Duration: A Corpus-based Analysis of Contextual Influences in Standard Mandarin). Because they blend the properties of consonants and vowels, Lei Wang argues they are best reclassified as syllabic approximants, resolving a long-standing classification failure in traditional Chinese phonology (The Phonology and Phonetics of Kaifeng Mandarin Vowels).
Contextual and Perceptual Confounding
Even when phonological categories are theoretically distinct, physical coarticulation—how adjacent sounds bleed into one another—can render them acoustically ambiguous to listeners.
Coarticulatory Bleeding
In Brazilian Portuguese, a vowel next to a nasal consonant naturally becomes nasalized due to the physical lag of the velum. How do listeners distinguish a "truly" nasal vowel phoneme from one that is merely accidentally nasalized? Luciana Marques et al. discovered that listeners do not just listen for nasality; they rely heavily on shifts in vowel quality (formant frequencies) caused by the nasalization (Perception and Acoustics of Vowel Nasality in Brazilian Portuguese). The ambiguity is resolved because the physical category shift alters the entire acoustic landscape of the vowel, not just a single feature.
Acoustic Overlap
Sometimes, distinct lexical items collapse into the exact same acoustic space. Mathilde Hutin et al. demonstrated that when controlling for stress, the French "schwa" vowel is acoustically identical to its full, mid-front rounded neighbors /ø/ and /œ/ (French schwa is not acoustically distinct from its two lexical neighbors /ø/ and /œ/). This reveals that the perceived identity of a sound often depends entirely on structural context rather than its inherent physical signal.
When historical sound changes erase a consonant category, the system adapts by shifting the distinction elsewhere. In Punjabi, Jasmeen Kanwal et al. show that the loss of historically voiced aspirated consonants left an ambiguous acoustic signal that listeners now resolve by tracking pitch contours on the following vowel, effectively transforming a consonantal category distinction into a tonal one (The Role of Pitch in Punjabi Word Identification).
From Binary Categories to Gradient and Dynamic Models
If speech is inherently ambiguous and continuous, why do traditional speech models fail, and how do modern approaches succeed? The breakthrough lies in abandoning binary, static classification in favor of gradient activation and dynamic modeling.
Gradient Grammar
Traditional phonology struggles with word-boundary phenomena like French liaison, where a transitional consonant appears between two words. Does the consonant belong to the end of Word 1 or the start of Word 2? Forcing a binary choice fails because the sound exhibits properties of both.
To solve this, Eric Rosen developed the EDGAR algorithm, which models linguistic symbols as having continuous, gradient activation levels rather than binary presence or absence (Learning a gradient grammar of French liaison). By allowing a consonant to be "partially active" in both lexical slots simultaneously, the model resolves long-standing theoretical conflicts and naturally accounts for complex linguistic exceptions.
Functional Data Analysis
When speech undergoes reduction (such as words compressing in casual conversation), is the change gradual or categorical? Static measurements—like taking a single average pitch or duration value—cannot answer this because the data clusters overlap heavily.
Static Measurement: [ Category A ] <--- Overlap ---> [ Category B ]
(Ambiguity Zone)
Dynamic Trajectory: Category A ───► Gradient Transition ───► Category B
(Captured as a continuous curve)
To capture this fluidity, Michele Gubian et al. introduced Functional Data Analysis (FDA) to speech dynamics (Functional data analysis as a tool for analyzing speech dynamics - a case study on the French word c'était). Instead of treating speech as a sequence of static points, FDA treats an entire acoustic trajectory as a continuous mathematical function.
This approach succeeds because it preserves the time-varying velocity and acceleration of the vocal tract. By analyzing the entire curve rather than isolated points, models can map the subtle, continuous paths speakers take between categories, proving that phonetic ambiguity is not a failure of the system, but an inherent property of a continuous physical medium being mapped to a flexible mind.
Key Foundational Work
Recognition of German obstruents (2005) — contributed 2 concept elements to this direction.
Czech voiced labiodental continuant discrimination from basic acoustic data (2005) — contributed 1 concept elements to this direction.
Spirantization of /p t k/ in Sienese Italian and so-called semi-fricatives (2005) — contributed 1 concept elements to this direction.