Combining text and prosodic analysis for prominent word detection

This paper presents an approach that considers both the corpus level (global) information as well as localized acoustic patterns to discover prominent words in an audio conversations. The global information is extracted by using text analysis techniques, in particular latent Dirichlet allocation (LDA), that extracts domain specific prominent words and also arranges them in a set of topics. The domain specific terms thus extracted are used for training an acoustic classifier that considers prosodic patterns around these terms and automatically learns features that are important for distinguishing prominent words from non-prominent ones. At the evaluation time, we extract prosodic feature and query the acoustic classifier for its prominent state. We evaluate our proposed system using a set of 40 manually annotated audio conversations. Experiments show that the proposed approach has a recall of up to 51% and a precision of 33% in detecting the prominent words. Note that both the lexical and the prosodic classifiers are trained in a purely unsupervised manner.