Automatic intonation-based keyword extraction from academic discourse

Abstract

This paper examines the perspectives of intonation processing for automatic keyword extraction. Based on a discourse intonation model from D. Brazil, automatic tone pattern recognition in speech stream is performed. It is shown that automatic classification of tone patterns can be done using simple polynomials and correlation. The original software tool PitchKey-wordExtractor (PKE) was applied to academic discourse (on-line lectures) to extract keywords. The results were compared to the output of popular tools for speech analytics: VoiceBase and IBM Watson. All the records were processed also with Praat software and annotated by human experts. Experiments show that none of the automatic systems outperforms the others and PKE, VoiceBase and IBM Watson have the identical error rates with respect to human expert opinion. It motivates further research and supports the tendency to integrate intonation and, more generally, prosody processing in automatic keyword extraction.