Non-linearly Separable Classes Research Articles

Extension of a Kernel-Based Classifier for Discriminative Spoken Keyword Spotting

A keyword spotter is considered as a binary classifier that separates a class of utterances containing a target keyword from utterances without the keyword. These two classes are not inherently linearly separable. Thus, linear classifiers are not completely suitable for such cases. In this paper, we extend a kernel-based classification approach to separate the mentioned two non-linearly separable classes so that the area under the Receiver/Relative Operating Characteristic (ROC) curve (the most common measure for keyword spotter evaluation) is maximized. We evaluated the proposed keyword spotter under different experimental conditions on TIMIT database. The results indicate that, in false alarm per keyword per hour smaller than two, the true detection rate of the proposed kernel-based classification approach is about 15 % greater than that of the linear classifiers exploited in previous researches. Additionally, area under the ROC curve (AUC) of the proposed method is 1 % higher than AUC of the linear classifiers that is significant due to confidence levels 80 and 95 % obtained by t-test and F-test evaluations, respectively. In addition, we evaluated the proposed method in different noisy conditions. The results indicate that the proposed method show a good robustness in noisy conditions.

Novel Fisher discriminant classifiers

At the present, several applications need to classify high dimensional points belonging to highly unbalanced classes. Unfortunately, when the training set cardinality is small compared to the data dimensionality (“small sample size” problem) the classification performance of several well-known classifiers strongly decreases. Similarly, the classification accuracy of several discriminative methods decreases when non-linearly separable, and unbalanced, classes are treated. In this paper we firstly survey state of the art methods that employ improved versions of Linear Discriminant Analysis (LDA) to deal with the above mentioned problems; secondly, we propose a family of classifiers based on the Fisher subspace estimation, which efficiently deal with the small sample size problem, non-linearly separable classes, and unbalanced classes. The promising results obtained by the proposed techniques on benchmark datasets and the comparison with state of the art predictors show the efficacy of the proposed techniques.