融合引导概率的语音识别解码算法研究

语音帧在声学特征空间中的位置信息可以辅助解码器对潜在路径进行筛选。传统的语音识别系统缺乏利用这种位置信息。针对这种不足,本文提出一种引导概率模型,用于描述语音帧属于声学特征空间不同局部的概率,并将其用于识别。使用引导概率后,解码器更强调对声学特征空间中最有希望的局部进行搜索,保留并扩展通过此局部空间的路径,同时弱化不经过此局部空间的路径。实验结果显示,融合引导概率的解码算法在不显著增加解码复杂度的情形下,使汉字相对错误率下降10.95%。结果分析表明,融合了语音帧声学位置信息的解码方法能够更有效地鉴别潜在路径,从而降低误识率。      

A beam tracing method for interactive architectural acoustics

A difficult challenge in geometrical acoustic modeling is computing propagation paths from sound sources to receivers fast enough for interactive applications. This paper describes a beam tracing method that enables interactive updates of propagation paths from a stationary source to a moving receiver in large building interiors. During a precomputation phase, convex polyhedral beams traced from the location of each sound source are stored in a "beam tree" representing the regions of space reachable by potential sequences of transmissions, diffractions, and specular reflections at surfaces of a 3D polygonal model. Then, during an interactive phase, the precomputed beam tree(s) are used to generate propagation paths from the source(s) to any receiver location at interactive rates. The key features of this beam tracing method are (1) it scales to support large building environments, (2) it models propagation due to edge diffraction, (3) it finds all propagation paths up to a given termination criterion without exhaustive search or risk of under-sampling, and (4) it updates propagation paths at interactive rates. The method has been demonstrated to work effectively in interactive acoustic design and virtual walkthrough applications.     

基于双边信息的分布式视频压缩感知模型研究

针对传统视频编码技术计算量大和复杂度高的缺点,提出一种基于双边信息的分布式视频压缩感知算法。该算法将压缩感知技术与分布式视频编码技术相结合,把视频序列分为Key帧和CS帧,Key帧运用传统的帧内编码和解码,CS帧编码端运用压缩感知编码,解码端运用视频块内与视频块间的双边信息和梯度投影算法进行优化重构。通过双边信息的运动估计和压缩编码器的设计,实现基于双边信息的分布式视频压缩感知模型的构建。仿真结果表明该模型既可以实现高效编码,又可以实现复杂度由编码端向解码端转移,在较低的采样率下,提高视频的压缩能力和传输速度。     

Multiple-array passive acoustic source localization in urban environments

In many situations of interest, obstacles to acoustic wave propagation such as terrain or buildings exist that provide unique challenges to localization. These obstacles introduce multiple propagation paths, reflections, and diffraction into the propagation. In this paper, matched field processing is proposed as an effective method of acoustic localization in a two dimensional scattering environment. Numerical techniques can be used to model complex propagation in a space where analytical solutions are not feasible. Realistically, there is always some uncertainty in model parameters that in turn can adversely affect localization ability. In particular, uncertainty in array location, sound speed, and various parameters affecting inter-array coherence only are investigated. A spatially distributed, multiarray network is shown to mitigate the effects of uncertainty. Multiarray inverse filter processing techniques are evaluated through perturbation of uncertain model parameters. These techniques are more accurate and flexible to implement than other matched field processing methods such as time reversal.     

基于Polar码的水声通信信源信道联合译码方法

水声信道时延扩展较长,频域选择性衰落严重,导致水声通信提升可靠性困难.同时,水声通信实际发送的信源中通常存在残留冗余,传统方法难以利用这部分冗余,从而导致一定的带宽浪费.针对该问题,提出了一种基于Polar码的水声通信信源信道联合译码方法.该方法根据Polar码的译码结构,以信源状态转移关系为基础构建信源信道联合译码网...     

Distinguishing between the interior pressures induced from two independent sources within a room using the probabilistic approach

In this paper, a probabilistic approach is introduced and used to distinguish between the interior pressures that are induced from two independent sound sources within a rectangular room. One source is a vibrating wall of the room and the other is an interior point source. The model is set up using the modal analysis method. In the probabilistic method, the probability densities of the uncertain acoustic model parameters are computed. The values of highest probability density are identified and assigned to the uncertain parameters in the acoustic model. The contributions of the sound sources to the total pressure are then obtained from the acoustic model input with the identified parameters. The effects of the modal truncations, the modeling errors, and the measurement locations on the accuracies of the identification process are studied.     

Time-reversing array retrofocusing in noisy environments

Acoustic time reversal is a robust means of retrofocusing acoustic energy, in both time and space, to the original sound-source location. However, noise may limit the performance of a time-reversing array (TRA) at long source-array ranges, or when the original-source or TRA-element power levels are low. The operation of a TRA requires two steps (reception and transmission) so both TRA-broadcast noise and ambient noise must be taken into account. In this paper, predictions are made for how a simple omnidirectional noise field influences the probability that the signal amplitude from a narrow-band TRA will exceed the noise at the TRA's retrofocus. A general formulation for the probability of TRA retrofocusing, which can be used for TRA design, is developed that includes: the variance of the noise field, the original source strength, the TRA's element output power, the number of TRA elements (N), and the propagation characteristics of the environment. This formulation predicts that a TRA's array gain (in dB) at the retrofocus may be as high as + 10log10(N) to + 20 log10(N) depending on the relative strengths of the original source and the TRA's elements. Monte Carlo simulations in both a free-space environment and a shallow-ocean sound-channel environment compare well to this probability formulation even when simple approximate parametric relationships for the appropriate Green's functions are used. The dominant deviation between theory and simulation in the sound channel is caused by acoustic absorption.     

Att-U-Net：融合注意力机制的U-Net骨导语声增强*

近年来大量全卷积网络、U-Net等编解码网络结构应用于语音增强,它们具有计算复杂度低、模型参数少等优势。然而,与长短时记忆模型等方法相比,这些编解码结构仍存在不能充分利用先后时间之间和高低频率之间的关联信息等缺点,尤其对于长序列数据的输入,编解码结构存在信息丢失的问题。为保持计算效率的同时考虑更充分的时频关联信息建模,本文提出一种融合注意力机制的U-Net网络的骨导语音增强方法（Att-U-Net）,通过在跳跃连接中引入注意力机制,生成一个权重矩阵,将编码层中的全局信息根据权重融入对应的解码层中,使网络在编解码过程中能够关注输入数据中与增强目标相关程度高的重要信息,同时抑制不相关的信息。在骨导语音数据集上的实验表明,融合注意力机制的U-Net网络能在保持模型轻量化的同时有效提升骨导语音的增强效果,增强后的语音在各项客观评价指标上均优于基线模型。通过对编解码网络中间层的可视化分析发现,在解码过程中注意力机制有效地保留了有声段的信息,滤除了骨导语音由于骨导传声特性带来的中频共振,从而使得增强后的骨导语音具有较好的听觉效果。     

A ray technique to calculate reflected and transmitted waves in layered media

This paper considers the problem of calculating the propagation of acoustic waves within an ideal isotropic multilayer plate structure. In such a situation the process of mode conversion as the wave interacts with each interface of the plate creates an ever increasing number of waves to track, and to perform calculations on, as the wave propagates within the layered media. Exploring this problem by examining the ray paths of the multiple reflections within the plate structure, it is possible to show that upon careful consideration many of these paths will travel equivalent distances in time and space becoming coincident. The principle of superposition can then be used to combine these coincident paths, this superposition reduces the number of waves to track, and simplifies the problem so that the necessary calculations can be performed in a time efficient manner.     

基于四节点RSSI的三维空间定位算法

针对提高室内三维空间的定位精度和速度的问题,从减少节点数量、建立特定坐标系的角度提出了一种基于四节点和特定坐标系的三维空间定位算法。该算法引入位置相对固定的四个节点,提高解算速度;充分考虑有效反射区对衰减指数影响,根据伯鲁斯特角将天线辐射场区分为叠加加强区和叠加减弱区,建立三维坐标系进行定位,结合信号强度指示(Received Signal Strength Indicator, RSSI)衰减模型,在叠加减弱区定位,把三维空间问题转化到二维空间,降低定位算法的复杂度。文中分别阐述方法的原理与验证,结果表明,四节点定位算法的解算速度比降低复杂度(Complexity-Reduced Trilateration Approach , COLA)定位算法有较大提高,该算法平均定位误差小于2米。     

Prediction of noise inside an acoustic cavity of elongated shape using statistical energy analyses with spatial decay consideration

In this paper, Statistical Energy Analysis (SEA) is used to predict the interior noise of an acoustic cavity of elongated shape. The disadvantage of the conventional SEA method, which quantifies the response in terms of the energy averaged over each subsystem, is overcome by introducing a one-dimensional spatial decay relation, through which information about the acoustic energy variation in the elongated direction is taken into account. The modified SEA is experimentally validated using a 1:5 scaled space station prototype, having the longitudinal dimension much larger than the cross-sectional dimension. It is also compared with a model reported in the literature. It is shown that, in the region where the acoustic pressure level decays at a constant rate, the two models agree well with each other and are capable of estimating the acoustic pressure variation along the space station cabin. However, near the end walls where the decay rate of the acoustic pressure level is not constant, the proposed model provides better accuracy.     

针对周期偏移声信标的二阶高精度水声定位模型

针对传统长基线水声定位模型在以飞机黑匣子等周期偏移声信标为目标时定位精度显著下降的问题,提出了一种不依赖于信号周期信息的新型长基线定位模型。该模型将传统时延差定位模型拓展至二阶,通过求取到达时间的二阶导数,消去未知声信标周期。为避免由阶数提高引起的解算非收敛问题,引入了改进牛顿迭代算法,有效提高了新模型的收敛概率。所提出新模型的仿真试验及湖上试验的定位误差分别为3.14 m和1.19 m,试验结果表明,该模型实现了周期偏移声信标的高精度定位解算。     

On the performance of cooperative systems with distributed linear block coding

Small-scale fading is one of the main problems in wireless communication systems. Multiple transmit/receive antennas, providing spatial diversity, are a common solution to combat fading, but practical constraints at the user location may limit their use. User cooperation is an efficient technique to introduce spatial diversity when multiple antennas are not suitable. In this paper we study the physical-layer performance of a cooperative system based on distributed linear block coding. Analytical results in terms of bit error rate and outage probability are presented when perfect decoding at the user location is assumed. Simulation results in terms of bit error rate are shown, taking into account the impact of errors on decoding and channel estimation at both the user location and the receiver location. Two scenarios are considered, representing uplink communications from static users to a static or mobile base station.     

Explicit mapping of acoustic regimes for wind instruments

This paper proposes a methodology to map the various acoustic regimes of wind instruments. The maps can be generated in a multidimensional space consisting of design, control parameters, and initial conditions. The boundaries of the maps are obtained explicitly in terms of the parameters using a Support Vector Machine (SVM) classifier as well as a dedicated adaptive sampling scheme. The approach is demonstrated on a simplified clarinet model for which several maps are generated based on different criteria. Examples of computation of the probability of occurrence of a specific acoustic regime are also provided. In addition, the approach is demonstrated on a design optimization example for optimal intonation.     

Bayesian localization in an uncertain ocean environment

In order to improve the ability to localize a source in an uncertain acoustic environment,a Bayesian approach,referred to here as Bayesian localization is used by including the environment in the parameter search space.Genetic algorithms are used for the parameter optimization.This method integrates the a posterior probability density(PPD) over environmental parameters to obtain a sequence of marginal probability distributions over source range and depth,from which the most-probable source location and localization uncertainties can be extracted.Considering that the seabed density and attenuation are less sensitive to the objective function of matched field processing,we utilize the empirical relationship to invert those parameters indirectly.The broadband signals recorded by a vertical line array in a Yellow Sea experiment in 2000 are processed and analyzed.It was found that,the Bayesian localization method that incorporates the environmental variability into the processor,made it robust to the uncertainty in the ocean environment.In addition,using the empirical relationship could enhance the localization accuracy.     

应用ResNet和CatBoost检测重放语声*

针对短语声指令声频信息少、不适用句子级重放语声检测的问题以及近距离录声后用高质量重放设备重放的语声难以检测的问题，提出了一种适用于词级重放语声检测的模型。首先，利用短时傅里叶变换、低频平均能量计算和帧排序等方法选择声频帧，然后提取这些帧的伽马通频率倒谱系数。其次，用基于自注意机制的残差网络模型进一步提取伽马通频率倒谱系数中的信息，并转化为特征向量。最后，将提取后的特征向量用CatBoost分类，从而提高检测性能。在POCO数据集上的实验结果表明，提出的方法可以以87.54%的准确率和12.53%的等错误率检测重放语声，优于基线和现有的方法。该文提出的方法在ASVspoof2019 PA数据集上的等错误率与串联检测代价函数分别为4.92%和0.1418，证明该文方法也适用于多种设置的重放语声检测。     

概率盒框架下混合认知不确定声场的响应预测

针对含概率盒-证据混合认知不确定参数声场的响应预测问题,提出了一种概率盒框架下的改进区间蒙特卡洛方法。该方法首先将混合认知不确定参数转换为纯概率盒形式,然后结合有限元方法推导出混合认知不确定声场的盖根鲍尔多项式代理模型,再采用蒙特卡洛方法求解代理模型得到声压响应。以含概率盒-证据混合认知不确定参数的二维管道声场模型和卡车乘客舱声腔模型为例,计算结果表明混合认知不确定参数影响下的声压响应为概率盒形式,其包括声压响应极值和相应的概率信息,并且所提方法较常规混合离散方法效率更优,较基于一阶摄动法的区间蒙特卡洛方法准确性更高。研究结果表明:所提方法可以有效预测混合认知不确定声场的声压响应,并可进行声学性能的风险和保守估计。      

Effect of diffusive and nondiffusive surfaces combinations on sound diffusion

One of room acoustic goals, especially in small to medium rooms, is sound diffusion in low frequencies, which have been the subject of lots of researches. Sound diffusion is a very important consideration in acoustics because it minimizes the coherent reflections that cause problems. It also tends to make an enclosed space sound larger than it is. Diffusion is an excellent alternative or complement to sound absorption in acoustic treatment because it doesn’t really remove much energy, which means it can be used to effectively reduce reflections while still leaving an ambient or live sounding space. Distribution of diffusive and nondiffusive surfaces on room walls affect sound diffusion in room, but the amount, combination, and location of these surfaces are still the matter of question. This paper investigates effects of these issues on room acoustic frequency response in different parts of the room with different source-receiver locations. Room acoustic model based on wave method is used (implemented) which is very accurate and convenient for low frequencies in such rooms. Different distributions of acoustic surfaces on room walls have been introduced to the model and room frequency response results are calculated. For the purpose of comparison, some measurements results are presented. Finally for more smooth frequency response in small and medium rooms, some suggestions are made.     

海洋声场中分布式无源定位系统的节点配置方法

为了提高海洋声场中分布式无源定位系统的定位精度,提出一种基于自适应遗传算法的节点配置方法。首先,基于到达时间TOA(Time of Arrival)定位算法推导出均匀物理场与海洋声场中定位误差的CRLB(Cramer Rao Lower Bound)。之后,利用BELLHOP模型对海洋声场进行建模,获得任意位置处目标声源与节点接收信号相关的传递函数并对等效声速进行计算.以目标声源在观测区域内服从均匀分布为例,将定位误差的平均CRLB最小为优化准则,采用自适应遗传算法对节点进行优化配置。结果表明,该方法能够有效降低海洋声场中分布式无源定位系统的定位误差,并给出定位误差随节点个数增加呈非线性递减的变化趋势,可为工程应用提供理论指导。      

FS水声遥控系统

本文介绍一种可实用的ＦＳ式编码的水声遥控系统，该系统采用９个频率点进行移频编码，可有效抗多途对译码的影响，编码设计能使遥控指令的码间距做得较大，通过自适应滤波器有效译码，可将误码率控制在希望值以下。理论分析和试验结果均表明，本系统误码率接近０，解码率１５００ｍ以内大于０．９，近距离大于０．９６本系统已应用于水下物理场测量系统中。