• The Journal of the Acoustical Society of Korea
• /
• v.42 no.2
• /
• pp.94-101
• /
• 2023

A deep neural network composed of encoders and decoders, such as U-Net, used for speech enhancement, concatenates the encoder to the decoder through skip-connection. Skip-connection helps reconstruct the enhanced spectrum and complement the lost information. The features of the encoder and the decoder connected by the skip-connection are incompatible with each other. In this paper, for complex-valued spectrum based speech enhancement, Self-Attention (SA) method is applied to skip-connection to transform the feature of encoder to be compatible with the features of decoder. SA is a technique in which when generating an output sequence in a sequence-to-sequence tasks the weighted average of input is used to put attention on subsets of input, showing that noise can be effectively eliminated by being applied in speech enhancement. The three models using encoder and decoder features to apply SA to skip-connection are studied. As experimental results using TIMIT database, the proposed methods show improvements in all evaluation metrics compared to the Deep Complex U-Net (DCUNET) with skip-connection only.

• Journal of the Korea Academia-Industrial cooperation Society
• /
• v.12 no.11
• /
• pp.5150-5155
• /
• 2011

Alzheimer's disease (AD) and vascular dementia (VD) are the most common dementia. In this paper, we proposed a region selection for classification of AD, VD and normal (NOR) based on micro-Raman spectra from platelet. The preprocessing step is a smoothing followed by background elimination to the original spectra. Then we applied the minmax method for normalization. After the inspection of the preprocessed spectra, we found that 725-777, 1504-1592 and 1632-1700 $cm^{-1}$ regions are the most discriminative features in AD, VD and NOR spectra. We applied the feature transformation using PCA (principal component analysis) and NMF (nonnegative matrix factorization). The classification result of MAP(maximum a posteriori probability) involving 327 spectra transformed features using proposed local region showed about 92.8 % true classification average rate.

• Journal of the Korean Institute of Intelligent Systems
• /
• v.10 no.4
• /
• pp.304-309
• /
• 2000

For understanding the information processing in human brain, we analyze the EEG, a spontaneous electric activity on the scalp of the human. In this paper, we used the mutual information to analyze EEG. The mutual information is used to show the stochastic correlation between signals which are generated in the communication and information theory. The used EEG is evoked by each auditory stimulus in positive and negative emotional states. As a result, we found thet there is some difference at the mutual information in each emotional state.

• Proceedings of the Korean Institute of Intelligent Systems Conference
• /
• 2000.05a
• /
• pp.215-219
• /
• 2000

인간의 뇌 정보처리를 이해하기 위한 일환으로서, 많은 연구자들이 사람의 두피에서 자발적으로 발생하는 전기 활동인 뇌파(EEG)를 분석하였다. 측정된 뇌파는 잡음처럼 보이는 비선형적인 거동으로 인하여 단순한 관찰만으로는 그 특징을 분석하기가 매우 어렵다. 따라서 이러한 뇌파를 분석하고 이해하기 위한 방법으로 파워스펙트럼, 바이스펙트럼 등과 같은 스펙트럼 분석과 상관차원, 프랙탈 차원과 같은 비선형 카오스 분석 등과 같은 해석법들이 활발히 연구되어왔다. 본 논문에서는 이러한 기존의 방법 외에 두 신호사이의 통계적 의존성을 측정하는 양인 상호정보를 이용하여 뇌파의 특징을 분석하였다. 뇌파간의 상호정보 분석을 통해 두뇌에서의 정보의 흐름에 관한 특징을 알아보았고, 감성자극에 반응하는 두뇌의 활동영역을 알 수 있었다.

• Journal of the Institute of Electronics Engineers of Korea CI
• /
• v.48 no.4
• /
• pp.21-26
• /
• 2011

Feature ranking is useful to gain knowledge of data and identify relevant features. In this study, we proposed a use of feature ranking for classification of neuro-degeneration and vascular dementia in micro-Raman spectra of platelet. The entire region of the spectrum is divided into local region including several peaks, followed by Gaussian curve fitting method in the region to be modeled. Local minima select from the subregion and then remove the background based on the position by using interpolation method. After preprocessing steps, significant features were selected by feature ranking method to improve the classification accuracy and the computational complexity of classification system. PCA (principal component analysis) transform the selected features and the overall features that is used classification with the number of principal components. These were classified as MAP (maximum a posteriori) and it compared with classification result using overall features. In all experiments, the computational complexity of the classification system was remarkably reduced and the classification accuracy was partially increased. Particularly, the proposed method increased the classification accuracy in the experiment classifying the Parkinson's disease and normal with the average 1.7 %. From the result, it confirmed that proposed method could be efficiently used in the classification system of the neuro-degenerative disease and vascular dementia of platelet.

• Science of Emotion and Sensibility
• /
• v.16 no.3
• /
• pp.409-416
• /
• 2013

Attention is one of important cognitive functions in human affecting on the selectional concentration of relevant events and ignorance of irrelevant events. The discrimination of attentional and inattentional status is the first step to manage human's attentional capability using computer assisted device. In this paper, we newly combine the non-linear recurrence pattern analysis and spectrum analysis to effectively extract features(total number of 13) from the electroencephalographic signal used in the input to classifiers. The performance of diverse types of attention-inattention classifiers, including supporting vector machine, back-propagation algorithm, linear discrimination, gradient decent, and logistic regression classifiers were evaluated. Among them, the support vector machine classifier shows the best performance with the classification accuracy of 81 %. The use of spectral band feature set alone(accuracy of 76 %) shows better performance than that of non-linear recurrence pattern feature set alone(accuracy of 67 %). The support vector machine classifier with hybrid combination of non-linear and spectral analysis can be used in later designing attention-related devices.

• Journal of the Korea Institute of Information and Communication Engineering
• /
• v.6 no.6
• /
• pp.873-878
• /
• 2002

Stator current spectrum by the fast Fourier transform (FFT) of current signals has been widely used for fault detection in induction motors. In this paper, we propose efficient rotor fault detection of Induction motors using stator current spectrum monitoring. The proposed method utilizes the mean absolute difference (MAD) between a Predetermined reference vector and a feature vector extracted from the stator current spectrum. Our proposed approach requires a smaller amount of computations when compared to fault detection algorithms based on neural networks, since it uses simple MAD criterion to detect rotor faults related broken rotor bars. Experimental results show that our proposed method can successively detect the rotor fault of the induction motor.

• Journal of the Korea Society of Computer and Information
• /
• v.10 no.6 s.38
• /
• pp.47-54
• /
• 2005

In the most of speaker recognition system, speaker's characteristics is extracted from acoustic parameter by speech analysis and we make speaker's reference pattern. Parameters used in speaker recognition system are desirable expressing speaker's characteristics fully and being a few difference whenever it is spoken. Therefore we su99est following to solve this problem. This paper is proposed to use strong spectrum characteristic in non-noise circumstance and prosodic information in noise circumstance. In a stage of making code book, we make the number of data we need to combine spectrum characteristic and Prosodic information. We decide acceptance or rejection comparing test pattern and each model distance. As a result, we obtained more improved recognition rate than we use spectrum and prosodic information especially we obtained stational recognition rate in noise circumstance.

• The Journal of the Acoustical Society of Korea
• /
• v.26 no.5
• /
• pp.207-213
• /
• 2007

In this research, we propose an efficient Speech/Music discrimination method that uses spectrum analysis and neural network. The proposed method extracts the duration feature parameter(MSDF) from a spectral peak track by analyzing the spectrum, and it was used as a feature for Speech/Music discriminator combined with the MFSC. The neural network was used as a Speech/Music discriminator, and we have reformed various experiments to evaluate the proposed method according to the training pattern selection, size and neural network architecture. From the results of Speech/Music discrimination, we found performance improvement and stability according to the training pattern selection and model composition in comparison to previous method. The MSDF and MFSC are used as a feature parameter which is over 50 seconds of training pattern, a discrimination rate of 94.97% for speech and 92.38% for music. Finally, we have achieved performance improvement 1.25% for speech and 1.69% for music compares to the use of MFSC.

• Journal of the Korean Institute of Intelligent Systems
• /
• v.7 no.2
• /
• pp.34-42
• /
• 1997

This paper deals with the recognition of closed planar shape using a three dimensional spectral feature vector which is derived from the FFT(Fast Fourier Transform) spectrum of contour sequence and fuzzy ARTMAP neural network classifier. Contour sequences obtained from 2-D planar images represent the Euclidean distance between the centroid and all boundary pixels of the shape, and are related to the overall shape of the images. The Fourier transform of contour sequence and spectrum analyzer are used as a means of feature selection and data reduction. The three dimensional spectral feature vectors are extracted by spectrum analyzer from the FFT spectrum. These spectral feature vectors are invariant to shape translation, rotation and scale transformation. The fuzzy ARTMAP neural network which is combined with two fuzzy ART modules is trained and tested with these feature vectors. The experiments including 4 aircrafts and 4 industrial parts recognition process are presented to illustrate the high performance of this proposed method in the recognition problems of noisy shapes.