利用语音学知识和神经网络识别非特定人的复合元音

本文主要研究：以音素为识别基元，运用语音学知识，对非特定人的普通话复合元音进行识别。其特点是音素识别由神经网络（ＮＮ）完成，为了便于利用语音知识ＮＮ输入层的刺激采用语音的功率谱，用单元音训练的ＮＮ识另非特定人的普通话复合元音，识别率是５４％。而运用语音学知识后，其识别率提高到９０％。     

基于密集连接时延神经网络的说话人识别算法*

说话人识别技术是一项重要的生物特征识别技术。近年来,使用深度神经网络提取发声特征的说话人识别算法取得了突出成果。时延神经网络作为其中的典型代表之一已被证明具有出色的特征提取能力。为进一步提升识别准确率并节约计算资源,通过对现有的说话人识别算法进行研究,提出一种带有注意力机制的密集连接时延神经网络用于说话人识别。密集连接的网络结构在增强不同网络层之间的信息复用的同时能有效控制模型体积。通道注意力机制和帧注意力机制帮助网络聚焦于更关键的细节特征,使得通过统计池化提取出的说话人特征更具有代表性。实验结果表明,在VoxCeleb1测试数据集上取得了1.40%的等错误率（EER）和0.15的最小检测代价标准（DCF）,证明了在说话人识别任务上的有效性。     

汉语耳语音孤立字识别研究

耳语音识别有着广泛的应用前景,是一个全新的课题.但是由于耳语音本身的特点,如声级低、没有基频等,给耳语音识别研究带来了困难.本文根据耳语音信号发音模型,结合耳语音的声学特性,建立了一个汉语耳语音孤立字识别系统.由于耳语音信噪比低,必须对其进行语音增强处理,同时在识别系统中应用声调信息提高了识别性能.实验结果说明了MFCC结合幅值包络可作为汉语耳语音自动识别的特征参数,在小字库内用HMM模型识别得出的识别率为90.4%.     

超声探伤信号的人工神经网络识别

粗晶奥氏体不锈钢的超声探伤受到能否有效区分有用信号与背影噪声的限制，目前人们大多倾向使用频率分隔来提高缺陷回波比例。本文则介绍一种用傅里叶变换作特征提取，用前馈网络自动识别奥氏体钢中缺陷信号的方法。在作者的实验中，这种方法的正确识别率达到９０％。     

Study on the method and system implementation of an isolated word recognition system used in noisy environment

I.IntroductionRecentlytherearemanykindsofsystemsandproductsforspeechrecognition,butalmostallofthemareworkinginquietenvironment,theperformancearedegradedorevencan'tworkwhenitisoperatedinhighnoisyenvironmentssuchasincockpits,vehicle,workshopsetc.SonoiserobustnesshasbecomeoneofthemainobstaclesfortherealaPplicationsoftheautomaticspeechrecognizersanditattractstheattentionofresearchersinspeechtechnologyareas.Since1978,substantialeffortshavebeendevotedtotestandevaluatethespeechrecognizersusedinfight…     

神经网络在声发射信号模式识别中的应用

介绍了人工神经网络中的BP网络、RBF网络、Hamming网络、BP－Hamming网络在声发射信号模式识别中的应用现状，并对这些方法的优缺点进行了比较。     

复合深度神经网络在直升机声目标识别中的研究

针对直升机探测中目标运动过程连续识别的鲁棒性问题,提出了一种基于复合深度神经网络的直升机声学特征提取和识别框架。复合深度神经网络由卷积神经网络和长短时记忆神经网络以并行结构组合,进行直升机声学特征的优化,完成直升机类型识别。针对直升机声信号特性,对卷积神经网络进行了改进,使得该复合深度神经网络在信号短时谱基础上优化声信号特征表征并提取前后帧之间的相关信息,弥补通常声目标识别方法不能充分利用目标信号时间历程信息的缺陷。真实外场实验数据测试结果显示:相较于传统识别方法,该算法显著提升了直升机进入有效探测范围后连续识别的鲁棒性和目标识别正确率。     

A semi-analytical approach using artificial neural network for dielectrophoresis generated by parallel electrodes

In this paper, we present a semi-analytical approach to obtain the DEP force generated by parallel electrodes. By solving the electric potential equation using the separation of variables method, a solution was found in the form of a Fourier series with unknown coefficients. The unknown coefficients were determined by training a linear artificial neural network with the appropriate data satisfied on the boundary. This results of calculated electric field and DEP force for both planar electrode system and 3D electrode system are validated by comparison with the numerical results obtained using the commercial software CFD-ACE+.     

Recent progress on two-dimensional neuromorphic devices and artificial neural network

Mimicking biological synapses with microelectronic devices is widely considered as the first step in hardware building artificial neuromorphic networks, which is also the basis of brain-inspired neuromorphic computing. Numerous artificial neurons and synapses making up an artificial neuromorphic network have been gained wide attention due to their powerful and efficient data processing capabilities. Recently, artificial synapses, especially memristor-type and transistor-type synapses based on multifarious two-dimensional (2D) materials have been paid much attention. The unique properties of 2D materials make devices perform well in learning ability and power efficiency when mimicking synaptic behaviors, which highlights the feasibility of 2D neuromorphic devices in constructing artificial neuromorphic networks. Herein, the basic structures and principles of biological synapses are introduced, and the definitions of synaptic behaviors in synaptic electronic devices are discussed. Then, the progress of 2D memristor-type and transistor-type neuromorphic devices involving their device architecture, neuromorphic operational mechanism, and promising applications is reviewed. Finally, the future challenges of artificial synaptic devices based on 2D materials are discussed briefly.     

Prediction of the plasma distribution using artificial neural network

In this work,an artificial neural network (ANN) model is established using a back-propagation training algorithm in order to predict the plasma spatial distribution in an electron cyclotron resonance (ECR)—plasma-enhanced chemical vapor deposition (PECVD) plasma system. In our model, there are three layers:the input layer, the hidden layer and the output layer. The input layer is composed of five neurons: the radial position, the axial position, the gas pressure, the microwave power and the magnet coil current. The output layer is our target output neuron: the plasma density. The accuracy of our prediction is tested with the experimental data obtained by a Langmuir probe, and ANN results show a good agreement with the experimental data. It is concluded that ANN is a useful tool in dealing with some nonlinear problems of the plasma spatial distribution.     

Inversion for acoustic impedance of a wall by using artificial neural network

A new approach for measuring acoustic impedance is developed by using artificial neural network (ANN) algorithm. Instead of using impedance tube, a rectangular room or a box is simulated with known boundary conditions at some boundaries and an unknown acoustic impedance at one side of the wall. A training data basis for the ANN algorithm is evaluated by similar source method which was developed earlier by Too and Su [Too G-PJ, Su T-K. Estimation of scattering sound field via nearfield measurement by source methods. Appl Acoust. 1999;58:261-81 (SCI) (EI)] for the estimation of interior and exterior sound field. The training data basis is constructed by evaluating of acoustic pressure at a field point with various acoustic impedance conditions at one side of the wall. Then, the inversion for unknown acoustic impedance of a wall is performed by measuring several field data and substituting these data into ANN algorithm. The simulation result indicates that the prediction of acoustic impedance is very accurate with error percentage under 1%. In addition, one field point measurement in the present approach for acoustic impedance provides more straightforward and easier evaluation than that in the two point measurement of impedance tube.     

一种递归神经网络空间分集均衡器

水声信道中的衰变多途特性常引起难以克服的码间干扰(ISI)。为了降低误码率、改善系统性能,本文提出了一种递归神经网络空间分集均衡器。它在传统的空间分集结构中融入了多层感知器,该结构能够充分利用训练信号的信息,在可调权数较少的情况下,能够得到较好的均衡效果;针对该结构,采取了初始设定权值的办法,从而使该均衡器的收敛速度得到大大提高。仿真与湖试试验结果表明,该均衡器结构合理,对空间分集均衡性能有一定的改善。     

一种基于模式识别的多路盲语音提取方法

盲分离算法能在缺少混合系统参数的条件下仅由观测信号估计初始源,但分离信号存在固有的排列模糊性,这往往导致两次批处理过程中同一信号"对不准",因此很难获得连续的源信号。本文针对盲声源分离中存在的相同问题,根据语音和其他音频信号的特征差异,提出一种修正的自相关函数并以其值作为一个特征基元来表征声音信号的时序相关特性,同时用平均声门波形状参数作为另一个特征基元来表征语音产生的生理效应。以这两个参数作为识别不同音频信号的二维模式特征,采用一种模糊聚类算法提取多路盲分离语音。本方法有效克服了批处理盲声源分离中的信号排列顺序的不确定性,并通过选择合适的阈值提取多路连续语音。仿真给出了5路混合音频信号中盲提取两路连续语音的实验结果。     

束晕-混沌的神经网络自适应控制

 理论分析了强流离子束在周期磁场聚焦通道中传输时产生的束晕 混沌动力学行为,给出了近似反映实际聚焦磁场的余弦函数形式。然后利用神经网络方法对非线性复杂系统控制的优越性,提出前馈反传神经网络方法对强流离子束中束晕 混沌进行自适应控制。通过适当选择的神经网络控制结构和线性反馈系数以及自适应调整神经网络的权系数,可将强流离子束的包络半径达到束匹配半径的控制目标,且束包络的抖动大小明显减少,束晕 混沌现象得到了明显的抑制。     

基于时延神经网络模型的舰船辐射噪声目标识别*

水声目标被动识别是水声信号处理领域的研究热点之一。海洋环境中存在的不规则噪声干扰,使得基于传统方法的水声目标被动识别技术在实际的应用场景中效果不佳。本文采用一种基于时延网络(Time Delay Neural Network,TDNN)模型的舰船辐射噪声目标识别方法,该方法利用目标的短时平稳特性和长时关联特性对目标的声纹特征进行建模,使用梅尔谱图提取目标信号的初级特征,再通过融合注意力机制和时延神经网络的深度学习模型实现高级特性提取,最后再利用余弦相似度实现不同目标的类别划分。该方法在ShipsEar数据集和自行采集的数据进行测试验证,目标识别准确率分别达到79.2%和73.9%,可证明本文方法的有效性。     

Prediction of outdoor sound transmission loss with an artificial neural network

An artificial neural network is developed for rapid prediction of sound transmission loss (TL) during propagation outdoors. The network predicts TL for a nonturbulent atmosphere from inputs involving the source/receiver propagation geometry (height range: 0-5 m, horizontal separation distance: 100-900 m), source frequency (range: 20-200 Hz), ground properties, and atmospheric refractive profile characteristics. A parabolic equation (PE) code generates the training and test data sets for the network. To ensure that a minimal set of input parameters is used in the network training, a nondimensional version of the PE and accompanying boundary, initial, and atmospheric conditions is developed. A total of 10 independent, nondimensional input parameters are found to be necessary for the training. Approximately 27,000 random cases involving these 10 parameters are generated used to train networks with varying numbers of neurons. The root mean square (RMS) error between random test cases solved by the PE and corresponding neural network predictions was 2.42 dB when a sufficient number of neurons (about 44) are included in the hidden layer. Also, only 18% of the cases resulted in RMS errors that were greater than 2 dB.     

A one-dimension convolutional neural network based interference classification method

Interference is a common problem in wireless communication, navigation and radar systems. A wide variety of interferences are used to degrade the communication quality especially in electronic warfare environment. In modern military communication systems, interference classification is an important module for its ability to obtain prior interference information before adopting related anti-interference method. This paper proposes a deep learning based interference classification method, which applies one-dimension convolutional neural networks to automatically extract interference features for classification. Computer simulations show better classification performance and lower computational complexity. Meanwhile, this proposed method is implied on software defined radios (SDR) hardware, more than 99% correct classification probability can be achieved with limited samples of the received signal, which verifies the robustness of this proposed method.     

In vitro microemboli classification using neural network models and RF signals

Emboli classification is of high clinical importance for selecting appropriate treatment for patients. Several ultrasonic (US) methods using Doppler processing have been used for emboli detection and classification as solid or gaseous matter. We suggest in this experimental study exploiting the Radio-Frequency (RF) signal backscattered by the emboli since they contain additional information on the embolus than the Doppler signal. The aim of the study is the analysis of RF signals using Multilayer Perceptron (MLP) and Radial-Basis Function Network (RBFN) in order to classify emboli.Anthares scanner with RF access was used with a transmit frequency of 1.82 MHz at two mechanical indices (MI) 0.2 and 0.6. The mechanical index is given as the peak negative pressure (in MPa) divided by the square root of the frequency (in MHz). A Doppler flow phantom was used containing a 0.8 mm diameter vessel surrounded by a tissue mimicking material. To imitate gas emboli US behaviour, Sonovue microbubbles were injected at two different doses (10μl and 5μl) in a nonrecirculating at a constant flow. The surrounding tissue was assumed to behave as a solid emboli. In order to mimic real clinical pathological situations, Sonovue concentration was chosen such that the fundamental scattering from the tissue and from the contrast were identical. The amplitudes and bandwidths of the fundamental and the 2nd harmonic components were selected as input parameters to the MLP and RBFN models. Moreover the frequency bandwidths of the fundamental and the 2nd harmonic echoes were approximated by Gaussian functions and the coefficients were used as a third input parameter to the neural network models. The results show that the Gaussian coefficients provide the highest rate of classification in comparison to the amplitudes and the bandwidths of the fundamental and the 2nd harmonic components. The classification rates reached 89.28% and 92.85% with MLP and RBFN models respectively.This short communication demonstrates the opportunity to classify emboli based on a RF signals and neural network analysis.     

A packet routing strategy using neural networks on scale-free networks

We propose a dynamic packet routing strategy by using neural networks on scale-free networks. In this strategy, in order to determine the nodes to which the packets should be transmitted, we use path lengths to the destinations of the packets, and adjust the connection weights of the neural networks attached to the nodes from local information and the path lengths. The performances of this strategy on scale-free networks which have the same degree distribution and different degree correlations are compared to one another. Our numerical simulations confirm that this routing strategy is more effective than the shortest path based strategy on scale-free networks with any degree correlations and that the performance of our strategy on assortative scale-free networks is better than that on disassortative and uncorrelated scale-free networks.     

An artificial neural network for proton identification in HERMES data

The HERMES time-of-flight (TOF) system is used for proton identification, but must be carefully calibrated for systematic biases in the equipment. This paper presents an artificial neural network (ANN) trained to recognize protons from Λ0 decay using only raw event data such as time delay, momentum, and trajectory. To avoid the systematic errors associated with Monte Carlo models, we collect a sample of raw experimental data from the year 2000. We presume that when for a positive hadron (assigned one proton mass) and a negative hadron (assigned one π- mass) the reconstructed invariant mass lies within the Λ0 resonance, the positive hadron is more likely to be a proton. Such events are assigned an output value of one during the training process; all others were assigned the output value zero.The trained ANN is capable of identifying protons in independent experimental data, with an efficiencyequivalent to the traditional TOF calibration. By modifying the threshold for proton identification, a researchercan trade off between selection efficiency and background rejection power. This simple and convenient methodis applicable to similar detection problems in other experiments.