Stress annotated Urdu speech corpus to build female voice for TTS

Abstract

This research describes the stress annotation process for the two hours of Urdu speech corpus containing 18,640 words and 28,866 syllables to build a natural voice for Text-to-speech (TTS) system. For the stress annotation of speech corpus, two algorithms i.e. phonological and acoustic stress marking algorithms have been tested in comparison to perceptual stress marking. Urdu phonological stress markings algorithm [1] reports 70% accuracy whereas Urdu acoustic stress marking algorithm developed through this research reports 81.2% accuracy. This acoustic stress marking algorithm is then used to annotate two hours of Urdu speech corpus. It is a semi-automatic acoustic stress marking algorithm, which annotates 54% data automatically using duration cue whereas 46% data is marked manually using the acoustic cues of pitch, glottalization and intensity.