Almost all ASR front ends use an amplitude-independent representation of spectral shape as the primary feature vector, obtained via some combination of normalization, logarithms, or AR modeling. They also typically represent total power or loudness as a separate feature. These ideas are fine to first order, and have gotten ASR to where it is today. But they totally punt on the issue of what is ...

This paper presents a segmentation scheme for images containing both smooth regions and textures. It is based on a vector-valued anisotropic diffusion on a combined color/Gabor feature space, followed by a hierarchical segmentation using dynamics of multiscale ’generalized gradient’ watersheds. The proposed method gives good segmentation results and is shown to be more effective than its counte...

Classifiers based on probability density estimates can be used to find posterior probabilities for the objects to be classified. These probabilities can be used for rejection or for combining classifiers. Posterior probabilities for other classifiers, however, have to be conditional for the classifier., i.e. they yield class probabilities for a given value of the classifier outcome instead for ...

So far, applied research aiming at computer-assisted pronunciation training has normally concentrated on segmental aspects. Here, we present a database with realizations of nonnative English speakers with German, French, Spanish, and Italian as native language. We concentrate on the acoustic-prosodic modelling of word accent position and use a large prosodic feature vector to automatically reco...

This paper addresses detection of imperfections in repetitive regular structures (textures). Humans can easily find such defects without prior knowledge of the ‘good’ pattern. In this study, it is assumed that structural defects are detected as irregularities, that is, locations of lower regularity. We define pattern regularity features and find defects by robust detection of outliers in the fe...

Function tag assignment has been studied for English and Spanish. In this paper, we address the question of assigning function tags to parsed sentences in Chinese. We show that good performance for Chinese function tagging can be achieved by using labeling method, extending the work of Blaheta (2004). In this method, the objects being modeled are syntax trees which require some mechanism to con...

This paper describes a new feature vector classification method for speaker identification. Purpose of this paper is constructing robust speaker models which only use meaningful feature vectors and discard confusing feature vectors. To construct robust speaker model, proposed method classifies feature vectors using log-likelihood estimation. Experimental results, with various segments ranging f...

To improve the performance of the subspace classi er, it is e ective to reduce the dimensionality of the intersections between subspaces. For this purpose, the feature space is mapped implicitly to the in nite dimensional Hilbert space and the subspace classi er is applied in the Hilbert space.

This paper presents two methods for adding domain knowledge to similarity-based learning through feature construction, a form of representation change in which new features are constructed from relationships detected among existing features. In the first method, domain-knowledge constraints are used to eliminate less desirable new features before they are constructed. In the second method, doma...

We present a method to quantify the political legitimacy of a populace using public Twitter data. First, we represent the notion of legitimacy with respect to k-dimensional probabilistic topics, automatically culled from the politically oriented corpus. The short tweets are then converted to a feature vector in k-dimensional topic space. Leveraging sentiment analysis, we also consider the polar...