Towards Prosodic Phrasing of Spontaneous and Reading Speech for Romanian Corpora

This paper proposes a framework for developing an automatic annotation tool of Romanian prosody for spontaneous and reading speech and a set of acoustic cues at the prosodic word level, necessary to accurately discriminate the prosodic phrases. Even though many approaches have considered the silence pause as an important acoustic cue in the automatic detection of the prosodic phrase boundaries, our research results show that listeners perceive prosodic boundaries mainly through the embodiment of F0 reset and tonal contrasts between adjacent words. The silence pauses in spontaneous conversational and reading speech help to locate the prosodic boundaries only when they are accompanied by F0 and energy cues. To discriminate the prosodic phrases, we extracted the following acoustic features for each prosodic word: minimum, mean, maximum, standard deviation and regression of F0 and energy. Using these acoustic features led to 90% accuracy in prosodic phrases discrimination.