Disfluency Detection Across Domains

• Autor(a) principal: Moniz, Helena
• Data de Publicação: 2015
• Outros Autores: Ferreira, Jaime, Batista, Fernando, Trancoso, Isabel
• Tipo de documento: Artigo
• Idioma: eng
• Título da fonte: Repositório Científico de Acesso Aberto de Portugal (Repositórios Cientìficos)
• Texto Completo: http://hdl.handle.net/10451/31081
• topic: Disfluency detection; Acoustic-prosodic features; Cross-domain analisys; European Portuguese

This paper focuses on disfluency detection across distinct domains using a large set of openSMILE features, derived from the Interspeech 2013 Paralinguistic challenge. Amongst different machine learning methods being applied, SVMs achieved the best performance. Feature selection experiments revealed that the dimensionality of the larger set of features can be further reduced at the cost of a small degradation. Different models trained with one corpus were tested on the other corpus, revealing that models can be quite robust across corpora for this task, despite their distinct nature. We have conducted additional experiments aiming at disfluency prediction in the context of IVR systems, and results reveal that there is no substantial degradation on the performance, encouraging the use of the models in IVR domains.