Computational phonology investigates the sound systems of natural languages using formal mathematical tools and computational algorithms. Phonology — the study of how languages organize sounds into systematic patterns — encompasses processes such as assimilation, dissimilation, vowel harmony, tone assignment, and syllabification. Computational approaches formalize these patterns, enabling automatic phonological analysis, text-to-speech synthesis, speech recognition preprocessing, and theoretical investigations into the computational complexity of natural language sound systems.
Formal Models of Phonology
"A becomes B in the context after C and before D"
Vowel nasalization: V → Ṽ / _ N
(a vowel becomes nasal before a nasal consonant)
Formal: φ → ψ / λ _ ρ
where φ, ψ are feature matrices and λ, ρ are contexts
The generative phonology tradition, established by Chomsky and Halle's The Sound Pattern of English (1968), models phonology as ordered sequences of rewrite rules that transform underlying representations into surface forms. Kaplan and Kay (1994) proved the foundational result that any ordered sequence of context-sensitive rewrite rules of this type can be compiled into a finite-state transducer, establishing that generative phonology is regular in the formal language sense. This result connects phonology to the broader theory of finite-state computation and enables efficient implementation of phonological rule systems.
Constraint-Based Approaches
Optimality Theory (OT), introduced by Prince and Smolensky in 1993, reconceptualized phonology as the resolution of conflicts between ranked, violable constraints rather than the sequential application of rules. Computational implementations of OT use generation functions (GEN) to produce candidate surface forms and evaluation functions (EVAL) to select the optimal candidate according to the constraint ranking. While OT itself has been shown to exceed finite-state power in the general case, restricted versions of OT that are relevant to natural language remain within finite-state bounds.
A major research program in computational phonology investigates where phonological patterns fall within the subregular hierarchy — proper subclasses of the regular languages. Heinz (2011) showed that most phonotactic constraints (restrictions on which sounds can co-occur) are strictly local or tier-based strictly local, while most phonological processes (alternations) are input strictly local. These subregular characterizations explain why phonological patterns are learnable from positive data alone, resolving a tension with Gold's theorem that regular languages in general are not identifiable in the limit from positive examples.
Machine Learning in Phonology
Machine learning approaches to phonological analysis include learning phonological rules from data, inducing phonological features from acoustic properties, and discovering phonotactic generalizations from corpus statistics. Neural network models have been applied to phonological learning, including recurrent networks that learn to produce correct outputs for processes like vowel harmony and consonant assimilation. These models provide computational demonstrations of how phonological patterns might be learnable from data, complementing formal analyses of learnability.
Computational phonology has practical applications in grapheme-to-phoneme conversion (predicting pronunciation from spelling), phonological awareness tools for language learning, and the design of writing systems for previously unwritten languages. The field continues to bridge theoretical linguistics and practical engineering, using formal methods to understand the computational nature of sound patterns and applying this understanding to speech and language technology.