The derivation of prosody for text-to-speech from prosodic sentence structure

Abstract

Suprasegmental phenomena in synthetic speech should reflect the linguistic structure of the input text. An algorithm is described, which establishes the prosodic sentence structure (PSS). This can be achieved without exhaustive syntactic parsing, using a dictionary of 550 function words. Subsequently, phrase and accent locations are derived from the PPS; accentuation is also affected by some semantic and contextual information. Comparison of the resulting sentence prosody with that of a human (professional) speaker shows that more detailed syntactic analysis may be necessary. Most of the accentuation errors are caused by semantic, pragmatic and contextual factors. These factors can only partly be imitated (using heuristics), since the relations between linguistic representations and real-world knowledge are not yet fully understood.