Cross-language phrase boundary detection

摘要: We describe models of prosodic phrasing trained on multiple languages to identify boundaries in an unseen language. Our goal is create from High Resource languages, which hand-annotated phrase are available, use identifying a Low language, with little or no training material. train American English, Italian, Mandarin, and German test each these languages. find that, while pause the most important feature for boundary prediction all examined, role identification varies by annotator relative importance other features significantly also that different acoustic correlates characterize In some, silence > pitch intensity duration, more than pitch. These differences do not appear be attributable language family, since, e.g. English display patterns.