Contribution of natural durations to speech synthesized by rule

Abstract

Rules for speech synthesis using the FOVE program at Haskins Laboratories incorporate some durational rules which attempt to produce differences in durations similar to those in natural speech. A recent set of rules produce speech in which the words are approximately 85% intelligible in relatively difficult sentences. To determine whether further attention to durational values would prove profitable in revising the rules, an experiment was conducted in which synthetic sentences by rule were modified by changing only the durations to match those of a reading of the same sentences by a human speaker. Frequencies and amplitudes of the speech by rule were unchanged. Listeners' performance on ten nonsense sentences of the type “The (adjective) (noun) (verb) and (noun)” improved from 77% to 85% word intelligibility when natural durations were used. It seems clear that further refinement of the rules for duration as well as for frequency cues is necessary. [Work supported by Veterans Administration contract V101 (134) p-342.]