融合改进梅尔谱特征和深信念网络的语音测谎算法

为实现非接触式谎言检测,特提出了以语谱特征为线索,结合深度学习的谎言检测方法。为提取谎言中微颤抖所引起的语谱局部能量变化,算法先对梅尔频谱进行了Hu矩处理,然后进行离散余弦变换去除相关性。该特征利用了Hu矩的正交不变性和平移不变性,能较好的体现出语谱中局部能量的集中方式。然后将所提取的特征作为改进深信念网络输入进行谎言识别。为提高受限玻尔兹曼机的并行回火训练算法中相邻温度链之间的交换率,训练算法先对Markov链的状态能量进行等能量的划分,使得每个能量环内的状态具有相似的能量,然后再进行交换以提高交换率从而优化整个网络的训练。在Columbia-SRI-Colorado数据库上的实验表明,谎言识别率达到了71.47%,比梅尔倒谱系数特征的识别率提高了3%,比传统的BayesNet分类算法提高了7%。      

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BLCD分解与ACO-DBN网络的光纤振动信号识别

针对光纤振动信号受噪声干扰严重、特征提取单一和识别时间长的问题,提出了改进的局部特征尺度分解和蚁群算法优化深度置信网络的识别方法。首先,采用三次B样条函数插值拟合均值曲线改进局部特征尺度分解算法,并对原始信号进行分解得到一系列内禀尺度分量之和。其次,利用峭度因子和能谱系数构成融合指标筛选有效分量。然后,分别提取有效分量在时域、频域和时-频域的熵值特征进行融合并降维。最后,将综合特征向量馈入蚁群优化后的深度置信网络进行训练和识别,提高算法效率和识别率。采用实测数据进行实验验证,结果表明,信噪比平均提升8 dB,信号平均识别率可达95.83%,平均识别时间为0.715 s。     

一种基于音素模型感知度的发音质量评价方法

为了提高发音质量判别精度,提出了一种基于音素模型感知度的发音质量评价方法。它采用不同语音样本集合下样本声学特征的对数后验概率期望差作为音素模型对变异发音的感知度,并以此为基础,生成各音素对应的识别模型候选集。实验表明,所提出的方法使语音识别网络候选音素模型集合尺寸减少约95%;在非母语语音数据库上,该方法评分与人工专家打分相关性为0.828,基于该方法得到的声韵母错误检出率为70.8%,声调错误检出率为42.5%,均优于其它方法。      

决策树结构对说话人自适应影响的研究

提出一种利用自适应语料和训练语料对模型状态结构调整的算法。该算法在易混淆的状态间参数共享,提高了模型对样本的后验概率和对自适应语料的利用率,并间接地调整了系统决策树的结构。识别实验结果表明,在不同数量的自适应语句下,调整后的系统识别率比基线系统获得了一致的提高,结合使用MLLR说话人自适应,调整的系统识别率平均提高了 15．60％,有效地减少了测试语料与训练语料决策树结构不匹配造成的系统识别率降低。     

汉语发音质量评估的实验研究

研究了发音评估系统中通用的置信度测度——后验概率算法,针对它存在的不足,提出了两种改进方案。首先,为了降低计算复杂度,传统算法采用了求最大值算法代替求和算法,在被测发音偏离目标音素集的情况下,这会严重降低后验概率的计算精度,本文提出基于扩展的音素混淆网络的后验概率算法。其次,为使置信度能评估不同语音段长的发音质量优劣,传统算法采用了后验概率的段长规整策略,研究分析发现声学似然值与时间的关系更为紧密,所以本文提出了基于声学似然值的时间规整方案。试验结果表明:与传统算法相比,采用改进的置信度算法能使平均打分错误率相对降低35%左右,有效地改善了计算机辅助语言学习系统的性能。      

基于受限玻尔兹曼机与密集采样迭代加权的图像动作识别算法

针对当前动作识别技术中正确识别率不高,易受到环境变化的影响等问题,提出了一种基于受限玻尔兹曼机与密集采样特征迭代加权融合的动作识别算法。避免单个特征对图像序列的表达力不强,引入了受限玻尔兹曼机(RBM)特征与密集采样(DT)特征分别对行为动作进行特征提取,得到RBM特征和DT特征;定义一种迭代加权函数,将RBM特征与DT特征进行加权融合,形成描述能力更强的RBM-DT特征;基于K-近邻(KNN)算法,对RBM-DT特征进行分类学习,完成动作识别的决策判断。通过在KTH、Hollywood数据集上实验表明:与当前动作识别技术比较,提出的新算法能够有效识别各种行为动作,对各类型动作均具有更高的正确识别率与鲁棒性。     

一种区分性互补系统构造与融合的语音识别方法

在区分性训练的框架下,提出了一种基于混淆信息加权的互补系统构造方法。首先通过统计音素对的混淆信息,利用混淆信息给音素对加以不同的惩罚权重,分别以基线系统中的3个最优识别结果作为参考,计算混淆信息加权后的音素准确率,同时以正确的标注为参考计算标准的音素准确率。然后通过同时最大化混淆信息加权后的音素准确率和最小化标准音素准确率,构建模型层互补系统,并进一步通过结合RDLT (region-dependent linear transform)特征变换过程构造特征层的互补系统。实验结果表明,与互补最小音素错误准则相比,融合模型层互补系统后识别率提高了0.76%,同时融合特征层和模型层的互补系统识别率提高了1.35%。本方法可以增大互补系统间的差异性,提高系统融合后的识别性能。      

基于多重分形谱的光纤周界振动信号识别

为了有效识别光纤周界系统的振动信号,提出一种多重分形谱参数和改进概率神经网络相结合的光纤振动信号识别方法.该方法能够避免特征提取过程中需要选择经验阈值和模式识别过程中需要确定平滑因子的不足.首先,检验分析光纤振动信号多重分形的存在性和有效性.然后,计算和提取光纤振动信号的多重分形谱参数,构成能够准确描述信号非线性和复杂性特性的特征向量.最后,采用改进的概率神经网络算法进行自适应地学习和分类,实现对不同光纤振动信号的识别.采用现场实验采集的四种振动信号对该方法进行验证,结果表明,平均识别率达到96.25%,识别时间为1.63s.该方法在正确识别率方面优于传统的概率神经网络算法.     

说话人识别使用遗传RBF网络

针对RBF网络普遍采用的一种训练算法所存在的局部最佳问题,本文将遗传算法用于RBF网络训练过程,增强了该网络的全局寻优能力,提高了RBF网络的模式识别性能.说话人识别实验显示,改进训练算法后,RBF网络的说话人识别率有稍许提高.     

Theoretical analysis and simulation of conjugate heights for dual-conjugate AO system in lidar

A multi-conjugate adaptive optics （MCAO） can offer a possibility of widening field of view （FOV） characterized by the isoplanatic angle, and the choose of conjugate height becomes a basic problem for MCAO, which influences the size of iosplanatic angle. Considering the application of lidar, the isoplanatic angle＇s expressions of two deformable mirrors （DMs） MCAO for uplink and downlink are deduced. The effects of conjugate heights for dual-conjugate AO are thoughtfully discussed, and the isoplanatic angles are further analyzed. The results show that the isopanatic angle varies with the conjugate height and reaches the maximum as the conjugate height is at the optimal altitude. Moreover, the optimal conjugate height changes with the propagation distance.     

Quick single-photon detector with many avalanche photo diodes working on the time division

Due to the limit of response speed of the present single-photon detector, the code rate is still too low to come into practical use for the present quantum key distribution (QKD) system.A new idea is put up to design a quick single-photon detector.This quick single-photon detector is composed of a multi-port optic-fiber splitter and many avalanche photo diodes (APDs).Au of the ports with APDs work on the time division and cooperate with a logic discriminating and deciding unit driven by the clock signal.The operation frequency lies on the number N of ports, and can reach N times of the conventional single-photon detector.The single-photon prompt detection can come true for high repetition-rate pulses.The applying of this detector will largely raise the code rate of the QKD, and boost the commercial use.     

Characterization of PolyA and PolyC mismatches by Raman spectroscopy

A.C mismatches are studied by Raman spectral characterization of PolyA, PolyC, and their equimolar complex in solution of 0.14 mol/L Na ,pH7.0.Experimental results show that A·C mismatches occur to be A/B (mainly A) conformers, and unlike Watson-Crick base pairing, this kind of mismatches is stabilized by only one hydrogen bond involving cytosine N4H2 and adenine N7.The formation of A·C complex makes the base stacking interactions much stronger, and conformation of the backbone more ordered, which leads to obvious Raman hypochromic effect with some shifts in corresponding bands.     

Improvement of four-wave mixing-based wavelength conversion efficiency in dispersion shifted fiber by 40-GHz clock pumping

40-GHz clock modulated signal as a pump to improve the efficiency of four-wave mixing (FWM)-based wavelength conversion in a 26.5-km dispersion shifted fiber (DSF) is investigated. The experimental results demonstrate that the conjugated FWM component has higher intensity with the clock pumping than that with the continuous-wave (CW) light pumping. The improvement of FWM-based wavelength conversion efficiency is negligible when the pump power is less than Brillouin threshold. But when the pump power is greater than Brillouin threshold, the improvement becomes significant and increases with the increment of pump power. The improvement can increase up to 9 dB if pump power reaches 17 dBm.     

Multi-pulse operation of a Kerr-lens mode-locked femtosecond laser

Our experimental results show that the presence of a proper amount of negative group velocity dispersion is essential to multi-pulse operation of a Kerr-lens mode-locked femtosecond laser. We demonstrate that the pulse separations and the number of pulses contained within a cavity round trip are strongly dependent on the initial perturbations. The results allow us to get a better understanding on the influences of the convoluted self-phase modulation and intra-cavity dispersions on the stable multi-pulse oscillation in a Kerr-lens mode-locked femtosecond laser.     

Propagation of Helmholtz-Gauss beams in weak turbulent atmosphere

Based on the Rytov approximation of light propagation in weak turbulent atmosphere,the closed-form expressions of field and average irradiance of each one of the four fundamental families of Helmholtz-Gauss (HzG)beams:cosine-Gauss beams,stationary Mathieu-Gauss beams,stationary parabolic-Gauss beams,and Bessel-Gauss beams,which are propagating in weak turbulent atmosphere,are obtained.The results show that the field and average irradiance can be written as the product of four factors:complex amplitude depending on the z-coordinate only,a Ganssian beam.a factor of complex phase perturbation induced by atmospheric turbulence,and a complex scaled version of the transverse shape of the non-diffracting beam.The effect of weak atmospheric turbulence on irradiance distribution of the HzG beam can be ignored.     

An effective method for reducing speckle noise in digital holography

An effective method for reducing the speckle noise in digital holography is proposed in this paper.Different from the methods based on classical filtering technique,it utilizes the multiple holograms which are generated by rotating the illuminating light continuously.The intensity images reconstructed by a series of holograms generated by rotating the illuminating light possess different speckle patterns.Hence by properly averaging the reconstructed intensity fields,the speckle noises can be reduced greatly.Experimental results show that the proposed method is simple and effective to reduce speckle noise in digital holography.     

Influence of purity of HfO2 on reflectance of ultraviolet multilayer

The impurities in two kinds of HfO2 materials and in their corresponding single layer thin films were determined through glow discharge mass spectrum technology and secondary ion mass spectrometry(SIMS) equipment respectively.It was found that ZrO2 was the main impurity in the two kinds of HfO2 either in the original HfO2 materials or in the electron beam deposited films.In addition,the difference of Zr content in the two kinds of HfO2 single laver films was much laxger than that of the other impurities such as Ti and Fe.which showed that it was just ZrO2 that made the difference between the optical performance of the film products including the two kinds of HfO2.With these two kinds of HfO2 and the same kind of SiO2.we deposited HfO2/SiO2 multilayer reflective coatings at the wavelength of 266 nm.Experimental results showed that the reflectances of these two mirrors were about 99.85% and 99.15% respectively,which agreed well with the designed results what were based on the optical constants obtained from the corresponding single layer thin films.     

Speech enhancement by array crosstalk resistant ANC and spectrum subtraction

Microphone array-based speech enhancement has great importance for speech communications and speech recognition. To reduce the aperture of the microphone array and to increase the effect of the speech enhancement will greatly broaden the application areas of the microphone array. An array crosstalk resistant adaptive noise cancellation method is therefore presented. And then an improved spectral subtraction algorithm is further cascaded to obtain better enhancement results. Theoretic analysis and experiments indicate that the proposed scheme needs only a very small microphone array while it simultaneously achieves a higher SNR improvement. Besides, the proposed scheme can be used in many noisy environments and is easy for real-time implementation.