Performance degradation of an actual speaker recognition system is due to many factors. Mismatches in speaking-style of a same speaker between training and testing data are importance factors. This paper is based on a...
详细信息
Performance degradation of an actual speaker recognition system is due to many factors. Mismatches in speaking-style of a same speaker between training and testing data are importance factors. This paper is based on a database with multiple speaking-style variations and comes up with a series of speaking-style-related score normalization solutions (that is, S-Norms) for situations without speaking-style restrictions. Experimental results show that the integral performance is significantly improved, compared with the baseline system by using these normalization methods, including SZ-Norm, ST-Norm and SZT-Norm. The equal error rate drops by 27%, which indicates that score normalization-based methods are efficient.
Due to the features of high synchronization error and high power consumption, the classic synchronization scheme of two-way packets exchange is unfit for some applications in wireless sensor networks, especially for n...
详细信息
Aiming at the tracking control of underactuated surface vessels ( USV) with parametric modeling uncertainties, an dynamic neural-fuzzified model (DNFM) based nonlinear control algorithm was presented. After translatin...
详细信息
Aiming at the tracking control of underactuated surface vessels ( USV) with parametric modeling uncertainties, an dynamic neural-fuzzified model (DNFM) based nonlinear control algorithm was presented. After translating the underactuated problem into fullactuated one by model transformation, a nonlinear controller (NC) was designed. The DNFM partly identified the inverse dynamics of USV sufficiently while the structure and parameters are adjusted simultaneously. Well-trained DNFM is then parallel connected with NC for path following of USV, and the weights of DNFM are available adjusted online. Simulation results validated the effectiveness of the proposed algorithm.
Automatic facial features localization is capable of affording accurate face location which is highly desirable for face analysis. Active shape model (ASM) and active appearance model (AAM) are efficient frameworks fo...
详细信息
Automatic facial features localization is capable of affording accurate face location which is highly desirable for face analysis. Active shape model (ASM) and active appearance model (AAM) are efficient frameworks for facial features localization. However, the performance of these conventional methods is unsatisfactory under pose, illumination and facial expression changes. To overcome these drawbacks, an automatic and accurate facial features localization algorithm is proposed which has two improvements: 1) utilizing efficient machine learning methods (random forest classifier and pair-compared feature) to construct local appearance model to deal with illumination and expression changes;2) incorporating classifier outputs and shape constraint into the quadratic optimization. Experimental results over many images with obvious pose, expression and illumination changes have shown the accuracy and efficiency of our method.
An effective computational approach for novel non-coding RNA (ncRNA) gene prediction based on fuzzy neural networks is presented. It consists of three modules, including a genome sequence preprocessor, a fuzzy neural ...
详细信息
An effective computational approach for novel non-coding RNA (ncRNA) gene prediction based on fuzzy neural networks is presented. It consists of three modules, including a genome sequence preprocessor, a fuzzy neural network with structure learning predictor (FNNSL), and a postprocessor. The preprocessor converts the input sequence alignment to a sequence of sliding windows, and extracts effective features for each window. The five-layer structure of Takagi-Sugeno type fuzzy neural network is adopted for the predictor. Based on the fuzzy partitioning of the features and the definition of membership functions of fuzzy subsets, the prediction results are obtained by the computation of the input layer, the fuzzifying layer, the firing strength layer, the normalized firing strength layer, and the output layer. Furthermore, a structure learning algorithm is introduced to decrease parameter dimensions, enhance the computational efficiency, and avoid the over-learning. Finally, the postprocessor stitches overlapping predictions together. The experimental results show that the prediction accuracy of this method is higher than other ncRNA gene prediction tools.
Based on the policy search algorithm in partially observable Markov decision process (POMDP), an optimal policy search algorithm is proposed. An algorithm leading to matching law is then derived from the optimal algor...
详细信息
Based on the policy search algorithm in partially observable Markov decision process (POMDP), an optimal policy search algorithm is proposed. An algorithm leading to matching law is then derived from the optimal algorithm. The aim of the subject can find a policy parameter that can maximize the expected value of a value function, and the policy parameter is updated on the experience of the subject. Due to the Markov assumption for the environment, the optimal policy algorithm can be obtained from computing the gradient of the expected value of the value function. Theoretical analysis and simulation results show that the decision behavior achieved by this algorithm is able to reach matching law. The matching law can be met if one subject tries to maximize the expected value of the value function under the simple assumption that past choice behaviors do not affect the expected value of the value function and the current policy. It reveals the relationship between the matching behavior and the optimal policy search algorithm, and suggests that the matching behavior is a suboptimal decision behavior.
Knowing patterns of relationship in covert (illegal) networks is very useful for law enforcement agencies and intelligence analysts to investigate collaborations among criminals. Previous studies in network analysis h...
详细信息
Knowing patterns of relationship in covert (illegal) networks is very useful for law enforcement agencies and intelligence analysts to investigate collaborations among criminals. Previous studies in network analysis have mostly dealt with overt (legal) networks with transparent structures. Unlike conventional data mining that extracts patterns based on individual data objects, network structure mining is especially suitable for mining a large volume of association data to discover hidden structural patterns in criminal networks. Covert networks share some features with conventional (real world) networks, but they are harder to identily because they mostly hide their illicit activities. After the September 11, 2001 attacks, social network analysis (SNA) has increasingly been used to study criminal networks. However, Finding out who is related to whom on a large scale in a covert network is a complex problem. In this paper we will discuss how network structure mining is applied in the domain of terrorist networks using structural (indices) measures or properties from social network analysis (SNA) and web structural mining research and proposed an algorithm for network disruption. Structural properties are determined by the graph structure of the network. These structural properties are used for locating and isolating core members by using importance ranking score and thereby analyzing the effect to remove these members in terrorist networks. The discussion is supported with a case study of Jemma Islamiah (JI) terrorist network.
This paper describes a database of emotional speech variations named CESD. The database contains 600 utterances in the form of dialogues with 20 emotional variation modes consisting of 3 different emotions including a...
详细信息
This paper describes a database of emotional speech variations named CESD. The database contains 600 utterances in the form of dialogues with 20 emotional variation modes consisting of 3 different emotions including anger, impatience, neutral, joy, and happiness. Besides the utterances, the database also includes the corresponding label files which include silence or effective speech segments, emotional classes, emotional variation segments, and emotional quality. 67 normal acoustical features are extracted based on the Praat tool and stored in the database. Subjective assessments of the emotional variations demonstrate that the database is suitable for research on speech emotion recognition and emotional variations.
暂无评论