Break Index (BI) annotated speech corpus for Urdu TTS

This study presents 10 hours of speech corpus annotated at Break Index (BI) level for Urdu text to speech system (TTS). For the speech corpus annotation, 5-steps scale of BI is presented in this study, which assists us to understand the prosodic relationship between word boundaries and phrase boundaries in Urdu. Moreover, an algorithm for Urdu TTS is designed to detect the 5-steps scale of BI automatically using acoustic cues such as duration of phrasal breaks, intonational patterns, glottalization and phrase lengthening. This algorithm annotates 10 hours of speech automatically and reports 98.4% accuracy compared to perceptual marked BI. This research also discusses the controversial issue of prosody to syntax mapping and syntax to prosody mapping in Urdu. The results indicate: (i) in the context of noun phrase and case phrase (NP-KP), prosodic break occurs between NP-KP to eliminate syntactic ambiguity; (ii) string of KPs does not require explicit prosodic break because case markers act as phrase separators in KP-KP context (iii) verb phrase (VP) in an utterance is preceded by a prosodic break to differentiate it from other phrases in an utterance and (iv) NPs of complex predicates prosodically act as single unit in an utterance.