Xiao Yao

Harmonic potential field based routing protocol for 3D underwater sensor networks

The local minima has been deemed as a challenging issue when designing routing protocols for 3D underwater sensor networks. Recently, harmonic potential field method has been used to tackle the issue of local minima which was also a major bottleneck in path planning and obstacle avoidance of robotics community. Inspired by this, this paper proposes a harmonic potential field based routing protocol for 3D underwater sensor networks with local minima. More specifically, the harmonic potential field is calculated using harmonic functions and Dirichlet boundary conditions are used for the local minima, sink(or seabuoy) and sending node. Numerical results show the effectiveness of the proposed routing protocol.

JM-MAC: A JSW-based multi-channel MAC protocol in underwater acoustic sensor networks

Constrained by the interior properties of underwater acoustic channels such as limited bandwidth, time-varying multipath propagation, and large propagation delay, the design of MAC protocols for underwater acoustic sensor networks (UWSNs) poses significant challenges. Existing MAC protocols for UWSNs exhibit very low performance in channel utilization, throughput and end-to-end delay, as well as energy consumption since 1) they are based on the conventional stop-and-wait (SW) transmission scheme, and 2) fewer has error-control mechanism at data link layer. In this paper, we assume that the UWSN has a multi-hop topology, where there exist a sink node and some common sensor nodes; all the nodes adopt Time Division Multiple Access (TDMA) to access channel. First, we classify all the nodes into different groups according to the numbers of hops between these nodes themselves and the sink node; the nodes belonging to different groups adopt TDMA frames with different starting time. We then propose a new MAC protocol by incorporating the multi-channel technique and the juggling-like stop-and-wait (JSW) transmission scheme, which is called the JM-MAC. Finally, simulation results are presented to highlight its good performance in terms of channel utilization, throughput and delay.

GMM based classification of speech under stress using physical features

Physical stress from workload for speaker exerts limitations on his speech production in physiological system causing speech variability, and thereby reduces speech system performance. The speech under stress presents a marked difference from the speech under neural condition. The distribution for the stress samples in the feature space shows the discontinuity because of the existence of different stress levels and different physiological characteristics for each speaker under the stress condition. In this paper, we use a Gaussian Mixture Models (GMM) framework with the physical features derived from the speech production model for neutral/stress speech classification. Cluster analysis is performed within stress class, and several cluster areas can be found for the classification, which is following Gaussian distribution. Experimental results show that GMM outperforms other classifiers for differentiating neutral speech from stress.

Dynamic features of vocal folds based on speech production model for detection of stressed speech

In this paper, dynamic characteristics representing physiological variations for speakers are explored for speech under stress. A speech production model is used to estimate the physical parameters representing stiffness of vocal folds and subglottal pressure of trachea. We analyze the short-time and long-time temporal information respectively to catch dynamic feature in physical parameters using different window lengths, and the dynamic parameters are proposed in order to classify recorded samples into neutral and stressed speech. Results show the proposed dynamic features corresponding short-time and long-time window lengths could be able to lead to improvements in classification for the stressed speech.

Comparison Analysis of Classifiers for Speech under Stress

In this paper, we focus on the classification of neutral and stressed speech. The parameters representing airflow patterns in physiological system are achieved using a physical model. Speech features were modeled using Gaussian Mixture Models (GMM) and Support Vector Machines (SVM). A comparison is made of different classifiers to determine their performance in stressed speech classification. Results show that SVM outperforms the standard GMM and linear classifiers, because SVM can better solve the small sample size problem, which often occurs in stressed speech classification tasks.

Peak-error-constrained sparse FIR filter design using iterative L 1 optimization

In this paper, a novel algorithm is presented for the design of sparse linear-phase FIR filters. Compared to traditional l1-optimization-based methods, the proposed algorithm minimizes l1 norm of a portion (instead of all) of nonzero coefficients. In this way, some nonzero coefficients at crucial positions are not affected by l1 norm utilized in the objective function. The proposed algorithm employs an iterative procedure and the index set of these crucial coefficients is updated in each iteration. Simulation results demonstrate that the proposed algorithm can achieve better design results than both greedy methods and traditional l1-optimization-based methods.

Stress classification in speech based on stress levels

In this study, we focus on the discontinuity of samples distribution in the feature space due to the existence of different stress levels. Some parameters from stress samples, calculated using a mathematical model, are clustered to different virtual classes according to degrees of stress. Based on different stress level cluster analysis is proposed within stress class. Experimental results illustrate that the performance for differentiating different speaking styles has been improved.

Voice conversion based on empirical conditional distribution in resource-limited scenarios

In this paper, a computationally efficient voice conversion system has been designed in order to improve the performance in resource-limited scenarios. First, mixtures of Gaussians (MoGs) at fixed locations of Mel frequencies have been used to represent the spectrum of STRAIGHT compactly. Second, the key conditional distributions for prediction are approximated by building histograms of aligned features empirically. Experiments have confirmed that our proposed method can obtain fairly good results compared to the traditional method without huge computational costs.