Automatic Indexing of Speech Segments with Spontaneity Levels on Large Audio Database

摘要

Spontaneous speech detection from a large audio database can be useful for different applications. For example, processing spontaneous speech is one of the many challenges that Automatic Speech Recognition (ASR) systems have to deal with. Spontaneous speech detection can also be an informative descriptor for information retrieval. The main evidences characterizing spontaneous speech are dis-fluencies (filled pause, repetition, repair and false start) and many studies have focused on the detection and the correction of these disfluencies. In this study1 we define spontaneous speech as unprepared speech, in opposition to prepared speech where utterances contain well-formed sentences close to those that can be found in written documents. Disfluencies are of course very good indicators of unprepared speech, however they are not the only ones: un-grammaticality and language register are also important as well as prosodic patterns. This paper proposes a set of acoustic and linguistic features that can be used for characterizing and detecting spontaneous speech segments from large audio databases, and proposes a method to extract and to exploit these features in order to index audio documents with three speech spontaneity levels.