基于正弦谐波模型和BP神经网络的语音变换算法及实现

提出利用语音的正弦谐波模型和BP神经网络实现语音变换算法.提取语音的基音频率,进行高斯建模和变换,实现说话人韵律特征的变换.提取语音的巴克域倒谱参数,作为说话人的频谱特征,送入BP神经网络进行训练,利用训练好的BP网络实现频谱特征的变换.给出了变换实现的系统框图,进行了仿真实验,利用ABX测试对实验结果进行了评测.测试结果表明文中所述算法,较好地实现了语音变换.     

基于AR-HMM在线能量调整的语音增强方法

针对单通道语音增强技术对非平稳噪声的跟踪不准确、噪声抑制效果较差的问题,本文提出一种基于在线能量调整的语音增强方法.该方法以归一化临界带能量为特征,采用高斯混合模型对背景噪声进行分类,利用对应类型噪声的自回归隐马尔可夫模型(Auto-Regressive Hidden Markov Model,AR-HMM)和纯净语音的AR-HMM,在最小均方误差准则下估计语音和噪声的功率谱.考虑到非平稳环境中训练集和测试集的差异性,需在线调整语音模型和噪声模型中的能量,语音模型的能量调整采用迭代的期望最大化算法;噪声模型的能量调整则利用的是模型训练过程中的能量重估方法,并以最小值控制的递归平均算法确定噪声能量调整的初始值.在ITU-T G.160标准下对算法进行性能测试,测试结果表明,本文方法对非平稳噪声的跟踪效果较好,对噪声衰减量较大,收敛时间较短.     

一种数字音频水印嵌入算法的研究

结合人耳听觉模型,利用人耳掩蔽特性,提出了一种基于小波分解和复倒谱变换的音频数字水印算法。该算法通过对原始语音信号进行三级小波分解,提取出小波近似分量,从近似分量中选取重要系数。利用复倒谱变换系数的不相关性．对重要系数进行复倒谱变换实现解相关运算,最终把水印嵌入到语音复倒谱域中,实现了在语音信号中嵌入二值图像,提高了水印嵌入算法的不可听性。实验证明该方法在抗常见的信号处理如加噪、滤波、重采样、有损压缩等方面具有较好的稳健性。     

基于多元Laplace语音模型的语音增强算法

传统的短时谱估计语音增强算法通常假设语音谱分量相互独立,没有考虑语音谱分量间的相关性。针对这一问题,该文提出一种新的基于多元Laplace分布模型的短时谱估计算法。首先,假设语音的离散余弦变换(DCT)系数服从多元Laplace分布,以此利用谱分量间的相关性;在此基础上,利用多元随机矢量的高斯尺度混合模型表示,推导得到语音DCT系数矢量的最小均方误差(MMSE)估计的解析表达式;并进一步推导了基于该分布模型的语音存在概率,对最小均方误差估计子进行修正。实验结果表明,该算法在抑制背景噪声和减少语音失真等方面优于传统的语音增强方法。     

基于矢量泰勒级数的模型自适应算法

在实际环境中,由于测试环境与训练环境的不匹配,语音识别系统的性能会急剧恶化。模型自适应算法是减小环境失配影响的有效方法之一,它通过测试环境下的少量自适应数据,将HMM模型的参数变换到测试环境下。该文将矢量泰勒级数用于模型自适应,同时对HMM模型的均值向量和协方差矩阵进行变换,使其与实际环境相匹配。实验证明,该文算法优于MLLR算法和基于矢量泰勒级数的特征补偿算法,在低信噪比环境中性能提高尤为明显。     

基于循环神经网络的语音识别研究

语音识别作为人工智能研究中不可或缺的一部分已经逐渐渗透到人们的日常生活中。针对传统语音识别方法不能很好地实现并识别复杂多变、非特定人语音的问题,文中提出利用在时间序列上关联性较强的循环神经网络(RNN)建立语音识别模型。考虑到语音信号丰富的时频信息表达,在特征提取环节进行改进,利用具有较好时频分辨率的小波变换(WT)取代快速傅里叶变换(FFT)作为该模型的输入;然后,采用随时间展开的反向传播算法(BPTT)进行特征学习与训练。在实验测试中,首先,对比分析了基于小波变换的特征提取对识别效果的影响;其次,通过与传统的HMM模型及BP神经网络的识别率做对比,验证RNN神经网络可提高语音识别准确率和稳定性。     

不相关匹配追踪的分段区分性特征变换方法

为了提高基于分帧特征变换方法的稳定性,提出了一种基于分段的区分性特征变换方法.该方法将特征变换当成高维信号的稀疏逼近问题,采用状态绑定的方法训练得到基于域划分的线性变换矩阵（Region Dependent Linear Transform,RDLT）和基于最小音素错误准则均值补偿的特征（mean-offset feature Minimum Phone Error,m-fMPE）变换矩阵,将两者的特征变换矩阵构成过完备的字典;采用强制对齐的方式对语音信号进行分段,以似然度最大化作为目标函数,利用匹配追踪算法对目标函数迭代优化,自动地确定各语音信号段中的变换矩阵及其系数.为保证特征变换的稳定性,在选择变换矩阵过程中引入相关度测量,去除相关的特征基矢量.实验结果表明,相比于传统的RDLT方法,当声学模型分别采用最大似然和区分性准则训练时,识别性能分别可以提高1.63%和2.23%.该方法同时能应用于语音增强和模型区分性训练中.     

动态时间规正与差别子空间相结合变异语音识别的在线训练方法

应力影响下的变异语音(由于说话人受到重力加速度变化而产生)可以用动态时间规正与差别子空间相结合的方法进行识别,但是该方法空间开销很大,而且训练算法极为复杂。针对该方法提出特征矢量替换法和特征矢量求平均法两种基于聚类思想的简单在线训练方法,取一个较小的初始训练集,然后用在线训练的方法使系统识别率达到最大。实验结果表明,其中特征矢量求平均的在线训练方法非常有效,保证系统整体识别率的同时,大大减小了系统的空间开销,适合于训练数据有限的变异语音识别。     

基于长时信息的自适应话音激活检测

语音信号的长时信息应用于话音激活检测中表现优越.利用三种听觉滤波器组,对语音信号进行非线性的谱分解,本文提出了六种基于听觉滤波器组的长时信息,并提出了基于长时信息的自适应话音激活检测算法.该算法无需训练数据,根据多种长时信息,直接在待测信号中挑选出类别明确的信号,然后利用这些信号训练分类模型,对待测信号按帧进行语音-非语音分类.在TIMIT语音库和NOISEX-92噪声库上的实验表明,该算法在极低信噪比环境下,仍表现出更高的准确性和更强的稳健性.同时,在线实验表明,算法在实时处理中仍能取得优异的性能.     

基于LPC模型的DTW语音转换系统设计

文章重点研究了基于LPC模型的DTW语音转换方法.采用DTW技术进行模型特征参数对齐的优势是:经过数据对齐后,使得人工神经网络算法更好的训练特征参数,得到的映射规则能很好规范要转换的特征参数,使得语音转换质量更高.论文的仿真结果和数据分析表明,基于LPC模型的DTW语音转换系统转换出的语音自然度较高.该研究结论对于语音...     

层次语义类型树模型及其在汉英机器翻译中的应用

We introduce a novel Semantic-Category-Tree (SCT) model to present the semantic structure of a sentence for Chinese-English Machine Translation (MT). We use the SCT model to handle the reordering in a hi-erarchical structure in which one reordering is depend-ent on the others. Different from other reordering ap-proaches, we handle the reordering at three levels: sen-tence level, chunk level, and word level. The chunk-lev-el reordering is dependent on the sentence-level reorde-ring, and the word-level reordering is dependent on the chunk-level reordering. In this paper, we formally de-scribe the SCT model and discuss the translation strate-gy based on the SCT model. Further, we present an al-gorithm for analyzing the source language in SCT and transforming the source SCT into the target SCT. We apply the SCT model to a rule-based patent text MT to evaluate the ability of the SCT model. The experimental results show that SCT is efficient in handling the hierar-chical reordering operation in MT.     

Why do users not report spear phishing emails?

Cyber security training programs encourage users to report suspicious spear phishing emails, and most antiphishing software provide interfaces to assist in the reporting. Evidence, however, suggests that reporting is scarce. This research examined why this is the case. To this end, Social Cognitive Theory (SCT) was used to examine the influence of the triadic factors of perceived self-efficacy toward antiphishing behaviors, expected negative outcomes from reporting spear phishing emails, and cyber security self-monitoring, on individuals’ likelihood of reporting spear phishing emails. Based on recent research on phishing victims, the present study also incorporated cyber risk beliefs (CRBs) into the SCT framework. The model, tested using survey data (N = 386), revealed that the likelihood of reporting spear phishing emails is increased by perceived self-efficacy, expected negative outcomes, and cyber security self-monitoring. Furthermore, the CRBs directly influenced the three SCT factors and indirectly the individuals’ likelihood of reporting spear phishing emails. The findings add to our understanding of SCT and the science of cyber security.     

基于空间相关性约束稀疏表示的高光谱图像分类

该文提出一种新的基于稀疏表示的高光谱图像分类方法。首先利用训练数据构造结构化字典,建立基于稀疏表示的高光谱图像分类模型;然后添加空间相关性约束项和训练数据的空间信息,提高稀疏表示模型分类的准确性;最后采用快速的交替方向乘子法求解模型。实验结果表明:该文方法能够有效提高分类精度,且分类结果稳定。     

Dropping False Packet to Increase the Network Lifetime of Wireless Sensor Network using EFDD Protocol

In Wireless Sensor Network (WSN) sensors are densely deployed where the intruders can compromise some sensor nodes and inject false data in order to raise false alarms, reduce network lifetime, utilize bandwidth resources and so on. False Data Injection can possibly occur in Data Aggregation (DA) and Data Forwarding (DF). This paper analyses EFDD Protocol-Early False Data Detection Protocol which addresses the two possibilities in a simple and secure way considering the constraints of sensor nodes. The main idea is the selection of the network structure; this protocol will work effectively in Spatial/Semantic Correlation Tree Structure (SCT). False Data Detection in DA is done using some monitoring nodes which will monitor the Data Aggregator. EFDD in SCT structure reduce the counterfeit data transmission when compared to other structure in a better way. The result shows that EFDD reduce data transmission by dropping false data earlier and it also reduces computation when compared to the existing schemes.     

Trajectory-based multi-dimensional outlier detection in wireless sensor networks using Hidden Markov Models

Wireless sensor networks (WSNs) have been increasingly available for monitoring the traffic, weather, pollution, etc. Outlier detection in WSNs is an essential step for many important applications, such as abnormal event detection, fraud analysis, etc. While existing efforts focus on identifying individual outliers from sensory data, the unsupervised high semantic outlier detection in WSNs is more challenging and has received far less attentions. In addition, the correlation between multi-dimensional sensory data has not yet been considered when detecting outliers in WSNs. In this paper, based on multi-dimensional Hidden Markov Models, we propose a trajectory-based outlier detection algorithm by model training and model-based likelihood estimation. Our data preprocessing, clustering, model training and model updating schemes are developed to reduce the computational complexity and enhance the detecting performance. We also explore the possibility and feasibility of adapting the proposed algorithm to real-time outlier detections. Experimental results show that our methods achieve good performance on detecting various kinds of abnormal trajectories composed of multi-dimensional data.     

Parameter re-estimation in semicontinuous hidden Markov modellingof speech with feedback to vector quantisation codebook

The parameters of the semicontinuous hidden Markov model (SCHMM) can be re-estimated by allowing the codebook to be updated, thus achieving an optimised codebook/model combination. With the optimised codebook, the SCHMM can offer improved recognition accuracy in comparison to both the continuous and the discrete hidden Markov model     

基于 Min-Sum-MSE的多用户MIMO训练序列与数据功率联合优化

本文研究了多用户MIMO系统中基于Sum-MSE最小准则预编码的训练序列与数据联合功率分配问题.考虑信道空间相关性,采用MMSE准则进行信道估计,在训练序列与数据总功率一定的条件下,通过分析上下行链路存在MSE对偶性,将下行Sum-MSE最小化转化为上行Sum-MSE最小化问题.然后推导了在发送端存在相关性时该功率联合...     

Analysis of error-gradient adaptive linear estimators for a class of stationary dependent processes

In many applications, the training data to be processed by an adaptive linear estimator can be assumed to have a finite correlation length. An exact analysis for this class of problems that yields the coefficient bias, coefficient correlation matrix, and mean square estimation error is obtained via a stochastic imbedding procedure. A power series expansion in the gain parameter is used to obtain simplified expressions of order one for the above statistical moments. These new expressions are shown to contain the terms that would result from an analysis based upon the assumption of independent training data plus additional terms arising from data correlation. Algorithm convergence properties are studied by identifying the appropriate matrix eigenvalues from the first-order theory.     

基于混合隐私的区块链高效模型协同训练共享方案

针对海量数据下,基于区块链的联邦学习数据共享平台面临的效率低下和隐私泄露问题,该文提出基于混合隐私的区块链高效模型协同训练共享方案。在该方案中,首先根据欧氏距离设计了一种基于相似度的训练成员选择算法来选择训练成员,组成联邦社区,即通过选取少量的高匹配训练节点来提高训练的效率和效果。然后,结合阈值同态加密和差分隐私,设计一种基于混合隐私技术的模型协同训练共享方案来保证训练和共享过程中的隐私性。实验结果和系统实现表明,所提方案可以在保证训练结果准确率的情况下,实现高效训练和隐私保护下的数据共享。