@inproceedings {pub2521,
	title = {Prosodic, Spectral and Visual Features for the Discrimination of Prominent and Non-prominent words},
	author = {Martin Ernst Heckmann},
	year = {2013},
	month = {March},
	abstract = {Amongst other things, speakers use prosodic cues to signal the importance words take in an utterance. When we correct what someone has said before, we usually make the corrected word very prominent. Different prosodic features have been proposed in the past to capture this word prominence. In this paper we compare prosodic cues as duration, energy and fundamental frequency as well to spectral cues commonly used for speech recognition, i.e. RASTA-PLP features, as to visual features extracted from the speaker{\textquoteright}s mouth. The evaluation is done on a dataset containing 13 speakers which interacted in a small game with a computer. The design of the game elicited correction utterances from the users. Hence it contains words uttered normally and with high prominence. From these interactions we made audio-visual recordings. Our results show that all of the features we investigated can be used to differentiate prominent from non-prominent words and that their combination is beneficial yielding discrimination rates of 76\%-92\%, depending on the speaker. },
	publisher = {DEGA},
	booktitle = {40. Jahrestagung f{\"u}r Akustik (DAGA)},
	address = {Oldenburg}
}