Abstract: Our aim is to perform a comparative evaluation of potential of oral versus nasal sounds in a European Portuguese Speaker Verification system. For that, we report, in this paper, the work on the necessary detection of the relevant segments. Implemented detection and classification consists in a typical cascade of speech framing, feature extraction and the use of classifiers. A total of 31 different features - including a subset of the features used recently by Pruthi and coworkers for American English nasal vowels detection - were extracted from each frame. Taking into account the small dataset restriction, we selected three classifiers: the well known SVM; the, more recent, Naive Credal Classifier 2 (from the Naive Bayes family of classifiers) and a Metaclassifier based on boosting (MultiBoostAB). Results, using a small database, showed as the best classifier the MultiBoostAB. Best results for Recall, Precision and F-measure, of 87.04, 88.0 and 87.5 %, were obtained for this classifier when trained with an equal number of samples of each class and non-including the first 40 % of the production of the nasal vowels.
Index Terms: Nasal vowels, Blind Segmentation, Naive Credal Classifier 2, MultiBoost, European Portuguese.