The purpose of this study was to investigate the characteristics of fundamental frequency(F0) and amplitude distributions in persons with cerebral palsy(CP) in the reading task. Participants were divided into three groups: 6 persons with spastic CP, 6 persons with athetoid CP and 6 normal persons who are around 15-20 years old. On the results of this study, firstly, in F0 distributions, most of the spastic CPs tended to appear narrow distributions on the basis of mode, but most of the athetoid CPs were opposite, and both of the CP groups tended to distribute highly on lower and higher frequencies than mean and mode. On the other hand, normal persons had a tendency to appear narrow distributions on the basis of mode. Finally, in amplitude distributions, the spastic CPs showed a tendency that there are little differences between the distribution of mode and the others, and most of the athetoid CPs showed a tendency that the distributions of mode were higher than the others. In addition to, the normal persons had a tendency that the distributions of mode were remarkably higher than both of the CP groups.

The purpose of this study is to analyze the intonational slope characteristics of the Korean news utterances. Prosodic phrases were analyzed in terms of the K-ToBI labeling system. In addition, the change of intonation contour that occurs throughout the sentences was discussed in terms of types of media and gender. Results showed that the overall declination of the intonation contour of radio and male revealed a gentler slope than that of TV and female, respectively. While the regression of the top line slope showed male's higher $R^2$ with the number of words, the base line slope of the radio and female was proved to be highly influenced from the number of syllables, words, and prosodic phrases. A lot more independent variables statistically affected to the base line slope. This means that the base line slope was strongly related to the variables, the top line slope, otherwise, could be more freely fluctuated due to the light correlation with them.

Knowledge in linguistic rhythm of the target language plays a major role in foreign language proficiency. This study attempts to discover valid rhythm features that can be utilized in automatic assessment of non-native English pronunciation. Eight previously proposed and two novel rhythm metrics are investigated with 360 English read speech tokens obtained from 27 Korean learners and 9 native speakers. It is found that some of the speech-rate normalized interval measures and above-word level metrics are effective enough to be further applied for automatic scoring as they are significantly correlated with speakers' proficiency levels. It is also shown that metrics need to be dynamically selected depending upon the structure of target sentences. Results from a preliminary auto-scoring experiment through a Multi Regression analysis suggest that appropriate control of unexpected input utterances is also desirable for better performance.

This paper proposes a method to improve pathological and normal voice classification performance by combining multiple features such as auditory-based and higher-order features. Their performances are measured by Gaussian mixture models (GMMs) and linear discriminant analysis (LDA). The combination of multiple features proposed by the frame-based LDA method is shown to be an effective method for pathological and normal voice classification, with a 87.0% classification rate. This is a noticeable improvement of 17.72% compared to the MFCC-based GMM algorithm in terms of error reduction.

Beamforming is one of the spatial filtering techniques which extract only desired signals from noisy environments using microphone arrays. Fixed beamforming is a simple concept and easy to implement. However, it does not show good performance in real noisy conditions. As an adaptive beamforming, Frost algorithm can be a good candidate. It uses the concept of the linearly constrained minimum variance (LCMV) algorithm. The difference between the Frost and the LCMV algorithm is the error correction scheme which is very effective feature in the aspect of performance. In this paper, as quadrature mirror filtering (QMF)-based filterbank is utilized as the pre-processing of the Frost beamformning, the filter length and the learning rate of each band is optimized to improve the performance. The performance is measured by the signal-to-noise ratio (SNR) and the Bark's scale spectral distortion (BSD).

We describe a new active learning method in conditional random fields (CRFs) framework for keyphrase extraction. To save elaboration in annotation, we use diversity and representative measure. We select high diversity training candidates by sentence confidence value. We also select high representative candidates by clustering the part-of-speech patterns of contexts. In the experiments using dialog corpus, our method achieves 86.80% and saves 88% training corpus compared with those of supervised method. From the results of experiment, we can see that the proposed method shows improved performance over the previous methods. Additionally, the proposed method can be applied to other applications easily since its implementation is independent on applications.

In this research, we propose a speaker identification system using a kernel method which is expected to model the non-linearity of speech features well. We have been using principal component analysis (PCA) successfully, and extended to kernel PCA, which is used for many pattern recognition tasks such as face recognition. However, we cannot use kernel PCA for speaker identification directly because the storage required for the kernel matrix grows quadratically, and the computational cost grows linearly (computing eigenvector of $l{\times}l$ matrix) with the number of training vectors I. Therefore, we use greedy kernel PCA which can approximate kernel PCA with small representation error. In the experiments, we compare the accuracy of the greedy kernel PCA with the baseline Gaussian mixture models using MFCCs and PCA. As the results with limited enrollment data show, the greedy kernel PCA outperforms conventional methods.

This paper presents a method for automatically detecting Korean prosodic boundaries using both acoustic and grammatical information for the performance improvement of speech information processing systems. While most of previous works are solely based on grammatical information, our method utilizes not only grammatical information constructed by a Maximum-Entropy-based grammar model using 10 grammatical features, but also acoustical information constructed by a GMM-based acoustic model using 14 acoustic features. Given that Korean prosodic structure has two intonationally defined prosodic units, intonation phrase (IP) and accentual phrase (AP), experimental results show that the detection rate of AP boundaries is 82.6%, which is higher than the labeler agreement rate in hand transcribing, and that the detection rate of IP boundaries is 88.7%, which is slightly lower than the labeler agreement rate.