@inproceedings {pub3161pub3244,
	title = {COMPARING SPEAKER INDEPENDENT AND SPEAKER ADAPTED CLASSIFICATION FOR WORD PROMINENCE DETECTION},
	author = {Andrea Schnall AND Martin Ernst Heckmann},
	year = {2016},
	month = {December},
	abstract = { Prominence is an essential tool in human communication to e.g. express important information. But since there are large differences how prominence is expressed by different speakers, it is not easy to extract and to incorporate it in a speech recognition system. In speech processing, for the handling of new data in a speaker independent trained model, adaptation techniques has been established. Most speaker adaptation techniques are developed for GMM-HMM based classifiers. We formerly showed that for such problem cases like the prominence detection a discriminative classifier like the SVM performance better. Therefore, we introduced a SVM-RBF based adaptation technique with a Gaussian regularization, GrfSVMA. One problem of most speaker adaptation techniques is that a sufficient amount of labeled speaker data is needed. Through adding an additional regularization term to our technique we stabilize the prediction and reduce the amount of necessary data while keeping same performance. },
	publisher = {IEEE},
	booktitle = {2016 IEEE Workshop on Spoken Language Technology}
}