Combining energy and cross-entropy analysis for nuclear segments detection

Features related to rhythmic patterns are involved in the representation of the intonational content for spoken language analysis. Among others, speech rate is one of the most used measures extracted by systems using prosodic analysis and is typically measured in syllables per second. Automatic approaches designed to estimate this measure in absence of manual annotations usually mark the position of syllable nuclei as a single point in time. Approaches extracting duration features using automatic segmentation in units shorter than words but larger than phones tend to detect syllables. To represent the prosodic contents of an utterance, especially from the rhythmic point of view, automatic positioning of nuclear boundaries may, however, be more informative than syllable boundaries. In this paper we present a method combining the analysis of the energy envelope and of the cross-entropy profile to obtain a segmentation into nuclear and inter-nuclear segments, showing that the proposed method can be used to obtain a reliable estimate of speech rate and that accuracy in nuclear boundary positioning allows the extraction of segmental features useful for automatic prosodic analysis.