Stress effects upon Korean vowels were investigated with reference to rhythm. We measured three acoustic correlates (Duration: VOT, Vowel Duration; F0; Intensity) of stress from the seven pairs of stressed vs. unstressed Korean vowels /i, ${\varepsilon}(e)$, a, o, u, i, e/. The results of the experiment revealed that stress gave only inconsistent and weak effects on duration, which supports that Korean is not a stress-timed language as far as strong stress effects on duration are still considered crucial in stress-timing. On the other hand, Korean stressed vowels were most characterized with higher F0 and next with stronger intensity. But speakers generally showed tactics to reversely use F0 and intensity in stressing an utterance rather than proportionately strengthening both of the two acoustic correlates of stress. There was found great inter-speaker variability especially in the variations of duration.

This study was designed to explore the analogical reasoning and metaphoric understanding in typically developing children and language impaired children. 13 Language-impaired children were matched to 16 typically developing children on the basis of receptive vocabulary age. All 29 children were enrolled in the 1st to 3rd grade in regular elementary schools. All were administered analogical reasoning and metaphoric tasks. Results indicated that the children with language disabilities did not perform as well as the receptive vocabulary matched group on the two tasks. In addition, we found that both of children with and without language disabilities did not have relationship between analogical reasoning and metaphoric understanding.

The aim of this study is to show the differences in acoustic parameters between a pathological voice /a/ caused by vocal polyp and a normal voice /a/ produced after LMS (Laryngeal Microscopic Surgery). It was expected that voices of two kinds could be analyzed effectively in terms of HNR in specific frequency bands than in all frequency bands. For this study, 10 patients' voice were recorded before and after LMS and then were manipulated in terms of four acoustic parameter. It was found out that (a) frequency bands of 500Hz in the range of 1,000Hz to 4,000Hz were very useful to obtain HNR values; (b) frequency bands in the range of 1,248Hz to 5,500Hz on a log scale were very useful to obtain HNR values; (c) F0 dropped after LMS but not significantly; (d) the bandwidth of the second formant (B2) decreased significantly after LMS, while that of the first formant (B1) decreased after LMS but not significantly.

The purpose of this study is to blow the characteristics of disfluency in childhood. The subjects were 144 normal children at the age of between 3 to 8 years who lived in Seoul. All the subjects provided spontaneous conversational speech samples during free-play interactions with their friends. We investigated the patterns and the frequency of disfluency and its relevance with subject's age, speaking rate and MLU(mean length of utterance). The results of this study can be summarized as follows. (1) There was no difference in the frequency of disfluency with the speaker's age or speaking rate. (2) Interjection was the most frequently occurring pattern of disfluency. (3) Prolongation, revision, interjection increased with age while part-word repetition, single-syllable word repetition, multi-syllable word repetition decreased gradually. (4) A significant effect of MLU on the frequency of disfluencies were demonstrated. The regression analysis has shown that more disfluencies occurred in utterances of children whose MLU is longer.

In this paper, we investigate the relation between Bayes factor and likelihood ratio test (LRT) approaches and apply the neighborhood information of Bayes factor to building an alternate hypothesis model of the LRT system. To consider the neighborhood approaches, we contemplate a distance measure between models and algorithms to be applied. We also evaluate several methods to improve performance of utterance verification using neighborhood information. Among these methods, the system which adopts anti-models built by collecting mixtures of neighborhood models obtains maximum error rate reduction of 17% compared to the baseline, linear and weighted combination of neighborhood models.

Basic components for multimodal interface, such as speech recognition, speech synthesis, gesture recognition, and multimodal fusion, have their own technological limitations. For example, the accuracy of speech recognition decreases for large vocabulary and in noisy environments. In spite of those technological limitations, there are lots of applications in which speech-oriented multimodal user interfaces are very helpful to users. However, in order to expand application areas for speech-oriented multimodal interfaces, we have to develop the interfaces focused on usability. In this paper, we introduce usability and user-centered design methodology in general. There has been much work for evaluating spoken dialogue systems. We give a summary for PARADISE (PARAdigm for Dialogue System Evaluation) and PROMISE (PROcedure for Multimodal Interactive System Evaluation) that are the generalized evaluation frameworks for voice and multimodal user interfaces. Then, we present usability components for speech-oriented multimodal user interfaces and usability testing guidelines that can be used in a user-centered multimodal interface design process.

To estimate directions of multi-sound sources, we consider an approach based on zero crossings which provided more robust results to diffuse noise than the conventional cross-correlation-based method [6][7]. In reverberant environments, the performance of source direction estimation can be improved by using signal components through direct paths from sources to microphones. Since a cepstral prefiltering technique [8] removes the effect of reverberation, we propose a source direction estimation method which can find out intervals of the direct-path components by comparing original and cepstral-prefiltered envelopes. Simulations demonstrate that the proposed method can improve the performance of source direction estimation in reverberant environments.

In this paper, we propose a method for source localization in reverberant environments based on an adaptive eigenvalue decomposition (AED) algorithm which directly estimates channel impulse responses from a speaker to microphones. Unfortunately, the AED algorithm may suffer from whitening effects on channels estimated from temporally correlated natural sounds. The proposed method which applies sparse priors to the estimated channels can avoid the temporal whitening and improve the performance of source localization in reverberant environments. Experimental results show the effectiveness of the proposed method.

In spoken document retrievals (SDR), subword (typically phonemes) indexing term is used to avoid the out-of-vocabulary (OOV) problem. It makes the indexing and retrieval process independent from any vocabulary. It also requires a small corpus to train the acoustic model. However, subword indexing term approach has a major drawback. It shows higher word error rates than the large vocabulary continuous speech recognition (LVCSR) system. In this paper, we propose an probabilistic slot detection and n-gram based string matching method for phone based spoken document retrievals to overcome high error rates of phone recognizer. Experimental results have shown 9.25% relative improvement in the mean average precision (mAP) with 1.7 times speed up in comparison with the baseline system.

This paper describes a novel spectral envelope conversion method based on Gaussian mixture model (GMM). The core of this paper is rearranging source feature vectors in input space to the transformed feature vectors in feature space for the better modeling of GMM of source and target features. The quality of statistical modeling is dependent on the distribution and the dimension of data. The proposed method transforms both of the distribution and dimension of data and gives us the chance to model the same data with different configuration. Because the converted feature vectors should be on the input space, only source feature vectors are rearranged in the feature space and target feature vectors remain unchanged for the joint pdf of source and target features using KPCA. The experimental result shows that the proposed method outperforms the conventional GMM-based conversion method in various training environment.

In this paper, we propose a vocabulary coverage improvement method for embedded continuous speech recognition (CSR) using a part-of-speech (POS) tagged corpus. We investigate 152 POS tags defined in Lancaster-Oslo-Bergen (LOB) corpus and word-POS tag pairs. We derive a new vocabulary through word addition. Words paired with some POS tags have to be included in vocabularies with any size, but the vocabulary inclusion of words paired with other POS tags varies based on the target size of vocabulary. The 152 POS tags are categorized according to whether the word addition is dependent of the size of the vocabulary. Using expert knowledge, we classify POS tags first, and then apply different ways of word addition based on the POS tags paired with the words. The performance of the proposed method is measured in terms of coverage and is compared with those of vocabularies with the same size (5,000 words) derived from frequency lists. The coverage of the proposed method is measured as 95.18% for the test short message service (SMS) text corpus, while those of the conventional vocabularies cover only 93.19% and 91.82% of words appeared in the same SMS text corpus.

This paper addresses the problem of single-channel speech separation to extract the speech signal uttered by the speaker of interest from a mixture of speech signals. We propose to apply time-frequency smoothing to the existing statistical single-channel speech separation algorithms: The soft mask and the minimum-mean-square-error (MMSE) algorithms. In the proposed method, we use the two smoothing later. One is the uniform mask filter whose filter length is uniform at the time-Sequency domain, and the other is the met-scale filter whose filter length is met-scaled at the time domain. In our speech separation experiments, the uniform mask filter improves speaker-to-interference ratio (SIR) by 2.1dB and 1dB for the soft mask algorithm and the MMSE algorithm, respectively, whereas the mel-scale filter achieves 1.1dB and 0.8dB for the same algorithms.