DISCRIMINATION OF SPEECH REGISTERS BY PROSODY
Kikuo Maekawa · ICPhS · 2011
Normalized frequency data of X-JToBI prosodic labels were used to automatically discriminate 4 speech registers –academic presentation, simulated public speaking, dialogue, and reproduction speech– of the Corpus of Spontaneous Japanese (CSJ). It turned out that the use of prosodic label frequency information and speaking rate could achieve more than 85% accuracy (closed data). It also turned out that the prosodic cues contributing to the classification were distributed pervasively throughout speech.