声学模型区分性训练中的动态加权数据选取方法

提出了一种基于动态加权的数据选取方法, 并应用到连续语音识别的声学模型区分性训练中. 该方法联合后验概率和音素准确率选取数据, 首先, 采用后验概率的Beam算法裁剪词图, 在此基础上依据候选词所在候选路径的错误率, 基于后验概率动态的赋予候选词不同的权值; 其次, 通过统计音素对之间的混淆程度, 给易混淆音素对动态地加以不同的惩罚权重, 计算音素准确率; 最后, 在估计得到弧段期望准确率分布的基础上, 采用高斯函数形式对所有竞争弧段的期望音素准确率软加权.实验结果表明, 与最小音素错误准则相比, 该动态加权方法识别准确率提高了0.61%, 可有效减少训练时间.     

连续语音解码阈值参数的优化

针对目前连续语音识别解码过程中剪枝阈值的确定不能兼顾解码速度与精度的不足,文中提出一种多维剪枝阈值参数联合优化算法。该算法主要研究全局阈值、词尾阈值、激活模型数、令牌数四维剪枝阈值参数的优化,其优化的主要过程是首先应用多目标优化理论对这四维阈值参数进行联合优化,然后根据优化结果采用分段动态阈值的方法进行后处理。实验结果表明,采用该方法优化后的阈值参数进行一遍解码,解码器的剪枝性能得到明显改善,搜索空间的大小得到有效控制,达到预期的速度与精度权衡的优化效果。     

血管紧张素受体阻滞剂治疗慢性进展性肾病的研究进展

肾素-血管紧张素系统(RA)S 在慢性肾病的进展中起着非常重要的作用。阻断RAS作用已成为慢性进展性肾病(CPRD)治疗的主要手段之一。血管紧张素受体阻滞剂(ARB) 以其副反应小、安全性高等特点受到广泛关注。近十年的研究表明ARB可以从多方面确实有效地延缓慢性肾病的进展,并为CPRD的临床治疗提供了循证医学证据。     

A high-efficiency nonlinear frequency converter with a built-in amplitude modulator

We report simultaneous frequency conversion and amplitude modulation in an optical second-harmonic generator by electrooptically controlling the relative phase between the 1064-nm fundamental and the 532-nm second-harmonic fields in a dispersion crystal section between two periodically poled lithium niobate (PPLN) sections. Theoretical derivation and experimental demonstration were carried out for two novel crystal configurations, including a linear cascaded configuration in which a 1-cm dispersion section is sandwiched between two 2-cm PPLN sections, and a folding-crystal high-efficiency configuration in which the mixing waves traverse twice in a 2-cm PPLN section through total internal reflections in a 1.5-cm dispersion section. Due to the coherence enhancement in the constructive phase between the two second-harmonic generation (SHG) fields in the two PPLN sections, we measured a 30% increase in conversion efficiency compared to a 4-cm continuous-grating PPLN under the same condition. The measured half-wave voltage for the amplitude modulation is 1.1 V /spl times/ d (/spl mu/m)/l/sub d/ (cm), where d is the separation of the electrodes and l/sub d/ is the length of the electrodes.     

Dyeing bacterial cellulose pellicles for energetic heteroatom doped carbon nanofiber aerogels

The energy crisis and environmental pollution are serious challenges that humanity will face for the long-term. Despite tremendous efforts, the development of environmentally friendly methods to fabricate new energy materials is still challenging. Here we report, for the first time, a new strategy to fabricate various doped carbon nanofiber （CNF） aerogels by pyrolysis of bacterial cellulose （BC） pellicles which had adsorbed or were dyed with different toxic organic dyes. The proposed strategy makes it possible to remove the toxic dyes from waste-water and then synthesize doped CNF aerogels using the dyed BC pellicles as precursors. Compared with other reported processes for preparing heteroatom doped carbon （HDC） nanomaterials, the present synthetic method has some significant advantages, such as being green, general, low-cost and easily scalable. Moreover, the as-prepared doped CNF aerogels exhibit great potential as electrocatalysts for the oxygen reduction reaction （ORR） and as electrode materials for supercapacitors.     

基于MCE准则的语音识别特征线性判别分析

提出了一种基于最小分类错误（Minimum classification error,MCE）准则的线性判别分析方法（Linear discriminant analysis,LDA）,并将其应用到连续语音识别中的特征变换.该方法采用非参数核密度估计方法进行数据概率分布估计;根据得到的概率分布,在最小分类错误准则下,采用基于梯度下降的线性搜索算法求解判别分析变换矩阵.利用判别分析变换矩阵对相邻帧梅尔滤波器组输出拼接的超矢量变换降维,得到时频特征.实验结果表明,与传统的MFCC特征相比,经过本文判别分析提取的时频特征其识别准确率提高了1.41%,相比于HLDA（Heteroscedastic LDA）和近似成对经验正确率准则（Approximate pairwise empirical accuracy criterion,aPEAC）判别分析方法,识别准确率分别提高了1.14%和0.83%.     

Temperature-Invariant Superelastic and Fatigue Resistant Carbon Nanofiber Aerogels

Superelastic and fatigue-resistant materials that can work over a wide temperature range are highly desired for diverse applications. A morphology-retained and scalable carbonization method is reported to thermally convert a structural biological material (i.e., bacterial cellulose) into graphitic carbon nanofiber aerogel by engineering the pyrolysis chemistry. The prepared carbon aerogel perfectly inherits the hierarchical structures of bacterial cellulose from macroscopic to microscopic scales, resulting in remarkable thermomechanical properties. In particular, it maintains superelasticity without plastic deformation even after 2 × 10⁶ compressive cycles and exhibits exceptional temperature-invariant superelasticity and fatigue resistance over a wide temperature range at least from −100 to 500 °C. This aerogel shows unique advantages over polymeric foams, metallic foams, and ceramic foams in terms of thermomechanical stability and fatigue resistance, with the realization of scalable synthesis and the economic advantage of biological materials.     

The Southeast Collaboratory for Structural Genomics: a high-throughput gene to structure factory

The Southeast Collaboratory for Structural Genomics consists of four working groups. The protein production group supplies/develops high-output production of Pyrococcus furiosus, Caenorhabditis elegans, and selected human proteins. The X-ray crystallography group conducts high-throughput structure production in parallel with production-related research/development in nanocrystallization robotics, capillary crystallization cassette, synchrotron/home X-ray instrumentation, sample mounting robotics, data processing and pipelined structure analysis, combined refinement/validation protocols, and direct use of unlabeled native crystals (Direct Crystallography). The NMR group emphasizes/develops sample screening and backbone structure determination from residual dipolar coupling data. The bioinformatics group implements/develops local database interfaces, pipelined sequence/structure information search/updates, and database/bioinformatics toolkits.     

A Supervised Learning Approach to Flashlight Detection

Shot boundary detection is a fundamental step of video indexing. One crucial issue of this step is the discrimination of abrupt shot change from flashlight, because flashlight often induces a false shot boundary. Support vector machine (SVM) is a supervised learning technique for data classification. In this paper, we propose a SVM-based technique to detect flashlights in video. Our approach to flashlight detection is based on the facts that the duration of flashlight is short and the video contents before and after a flashlight should be similar. Therefore, we design a sliding window in temporal domain to monitor the instantaneous video variation and extract color and edge features to compare the visual contents between two video segments. Then, a SVM is employed to classify the luminance variation into flashlight or shot cut. Experimental results indicate that the proposed approach is effective and outperforms some existing techniques.     

Printable elastic silver nanowire-based conductor for washable electronic textiles

Printable elastic conductors promote the wide application of consumable electronic textiles (e-textiles) for pervasive healthcare monitoring and wearable computation. To assure a clean appearance, the e-textiles require a washing process to clean up the dirt after daily use. Thus, it is crucial to develop low-cost printable elastic conductors with strong adhesion to the textiles. Here, we report a composite elastic conductor based on Ag nanowires (NWs) and polyurethane elastomer. The composite could be dispersed into ink and easily printed onto textiles. One-step print could form robust conductive coatings without sealing on the textiles. Interestingly, the regional concentration of Ag NWs within the polyurethane matrix was observed during phase inversion, endowing the elastic conductor with a low percolation threshold of 0.12 vol.% and high conductivity of 3,668 S·cm⁻¹. Thanks to the high adhesion of the elastic conductors, the resulted e-textiles could withstand repeated stretching, folding, and machine washing (20 times) without obvious performance decay, which reveals its potential application in consumable e-textiles.