Experiments in rapid development of accurate phonetic alignments for TTS in Afrikaans

Abstract

The quality of corpus-based text-to-speech (TTS) systems depends on the accuracy of phonetic annotation (alignments) which directly influences the process of acoustic modelling. In this paper the author discusses the rapid development of accurate alignments for new languages in an under-resourced context based on an investigation of automatically obtained alignments of an Afrikaans speech corpus. It is shown that certain classes of inaccuracies can be effectively detected by acoustic analyses.