A model for varying speaking style in TTS systems

Abstract

This paper aims to enhance the performance of a TTS system\nby generating various speaking styles. First we describe three\nspeaking styles (Radio News, Political Address and\nConversation) and compare the prosodic features found in\nthese authentic styles with the prosody in “neutral” speech\nuttered by the eLite TTS system. Differences concern\nabout 20 prosodic characteristics (F0 span, speech rate, pauses\nand hesitation, primary and secondary accentuation, schwa\ndeletion, etc.). In order to make the neutral speech similar to a\ntypical speaking style, prosodic characteristics are\nimplemented within the TTS system itself or during a postprocessing\nstep. The quality of the “stylized” synthesis is\nevaluated by comparing it to the original style.