QUANTIFYING TIMBRE: AN ATTENTION-BASED PROSODIC ANALYSIS OF TEXT

The article is dedicated to a problem of the structural integrity of the text, where prosody plays a critical meaning-distinguishing role. It is evident that there is intrinsic prosody in the text that a reader perceives through linguistic cues, which the author purposefully uses. This intended prosodic emotional-expressive coloring of a text is timbre, which is predetermined by its linguistic characteristics, that reflects an author’s hidden purport. This article introduces a quantitative method to timbre analysis by integrating attention mechanisms from the Longformer model. The prosodic index (PI), exponential index (EI), and integral index (II) are calculated. The goal is to objectively measure word prominence, structural variability, and semantic/emotional buildup through the indices and to classify texts into distinct timbre types (coherent, dispersed, dynamic, or static). By applying this framework to text analysis, it is evident that PI scores reliably exhibit the prominence of words, while EI and II capture the development of meaning across the text. Attention-based metrics can be applied as a tool for philological research to further improve NLP models, translation and interpretation.