Spoken documents, such as podcasts or lectures, are a growing presence in everyday life. Being able to automatically
identify their discourse structure is an important step to understanding what a spoken document is about. Moreover,
finer-grained units, such as paragraphs, are highly desirable for presenting and analyzing spoken content. However, little
work has been done on discourse based speech segmentation below the level of broad topics. In order to examine how
discourse transitions are ...
Spoken documents, such as podcasts or lectures, are a growing presence in everyday life. Being able to automatically
identify their discourse structure is an important step to understanding what a spoken document is about. Moreover,
finer-grained units, such as paragraphs, are highly desirable for presenting and analyzing spoken content. However, little
work has been done on discourse based speech segmentation below the level of broad topics. In order to examine how
discourse transitions are cued in speech, we investigate automatic paragraph segmentation of TED talks using lexical
and prosodic features. Experiments using Support Vector Machines, AdaBoost, and Neural Networks show that models
using supra-sentential prosodic features and induced cue words perform better than those based on the type of lexical
cohesion measures often used in broad topic segmentation. Moreover, combining a wide range of individually weak
lexical and prosodic predictors improves performance, and modelling contextual information using recurrent neural
networks outperforms other approaches by a large margin. Our best results come from using late fusion methods that
integrate representations generated by separate lexical and prosodic models while allowing interactions between these
features streams rather than treating them as independent information sources. Application to ASR outputs shows that
adding prosodic features, particularly using late fusion, can significantly ameliorate decreases in performance due to
transcription errors.
+