Introduction
Speech synthesis requires converting text into phonemes, a task complicated by the irregularities of English. While dictionaries handle known words, "letter-to-sound" rules are needed for the rest. Storing every variation of a word (e.g., form, formal, formality) is inefficient. Morphological decomposition—breaking words into morphemes—addresses this but introduces the challenge of stress assignment, as suffixes often shift the stress of the root (e.g., divine vs. divinity).
Methodological Refinements
Kenneth Church proposes a sophisticated framework for morphological decomposition that moves beyond simple concatenation.
Aronoff-style Truncation and Statistical Derivation
The system utilizes "truncation rules" to capture the asymmetry of English morphology. For instance, presidency is derived from president by truncating -ent and adding -ency. To determine these relationships, Church employs a statistical model based on dictionary data.
As shown in the data above, there is a high probability (e.g., 43% for -ability to -able) that words with specific suffixes share a common root. This data-driven approach helps the system avoid "spurious analyses" and correctly identify primary stems.
Lexical Marking and Compound Theory
The paper argues that stress shifts are determined by syllable weight, an inherent property of stems that should be marked in the lexicon. Furthermore, Church challenges the traditional view of Level 2 affixes (like un-, anti-, -ness) as "stress neutral." Instead, he proposes they behave like compounds, where stress is assigned based on the word's grammatical category.
Results and Insights
By treating Level 2 prefixes as compound elements, the system successfully predicts stress patterns that previous theories missed. For example, it distinguishes the stress in the noun underdog from the verb undergo. The system also generated extensive "forests" of related words, identifying 38 words related to the stem port and 36 related to infect, aligning closely with human linguistic intuition.
Conclusion
This framework provides a more linguistically grounded approach to speech synthesis. By incorporating asymmetric truncation and category-based stress rules, it significantly improves the accuracy of phoneme conversion, leading to more natural-sounding synthetic speech. Future work remains in automating lexical marking for syllable weight and refining semantic nuances in compound stress.
Summary updated