基于层次化Conformer的语音合成

语音合成需要将输入语句的文本转换为包含音素、单词和语句的语音信号。现有语音合成方法将语句看作一个整体,难以准确地合成出不同长度的语音信号。通过分析语音信号中蕴含的层次化关系,分别设计基于Conformer的层次化文本编码器和基于Conformer的层次化语音编码器,并提出了一种基于层次化文本-语音Conformer的语音合成模型。首先,该模型根据输入文本信号的长度,构建层次化文本编码器,包括音素级、单词级、语句级文本编码器3个层次,不同层次的文本编码器描述不同长度的文本信息;并使用Conformer的注意力机制来学习该长度信号中不同时间特征之间的关系。利用层次化的文本编码器,能够找出语句中不同长度需要强调的信息,有效实现不同长度的文本特征提取,缓解合成的语音信号持续时间长度不确定的问题。其次,层次化语音编码器包括音素级、单词级、语句级语音编码器3个层次。每个层次的语音编码器将文本特征作为Conformer的查询向量,将语音特征作为Conformer的关键字向量和值向量,来提取文本特征和语音特征的匹配关系。利用层次化的语音编码器和文本语音匹配关系,可以缓解不同长度语音信号合成不准确的问题。所提模型的层次化文本-语音编码器可以灵活地嵌入现有的多种解码器中,通过文本和语音之间的互补,提供更为可靠的语音合成结果。在LJSpeech和LibriTTS两个数据集上进行实验验证,实验结果表明,所提方法的梅尔倒谱失真小于现有语音合成方法。

基于改进Conformer的新闻领域端到端语音识别

目前,开源的中文语音识别数据集大多面向通用领域,缺少面向新闻领域的开源语音识别语料库,因此该文构建了面向新闻领域的中文语音识别数据集CH_NEWS_ASR,并使用ESPNET-0.9.6框架的RNN、Transformer和Conformer等模型对数据集的有效性进行了验证,实验表明,该文所构建的语料在最好的模型上CER为4.8%,SER为39.4%。由于新闻联播主持人说话语速相对较快,该文构建的数据集文本平均长度为28个字符,是Aishell_1数据集文本平均长度的2倍;且以往的研究中训练目标函数通常为基于字或词水平,缺乏明确的句子水平关系,因此该文提出了一个句子层级的一致性模块,与Conformer模型结合,直接减少源语音和目标文本的表示差异,在开源的Aishell_1数据集上其CER降低0.4%,SER降低2%;在CH_NEWS_ASR数据集上其CER降低0.9%,SER降低3%,实验结果表明,该方法在不增加模型参数量的前提下能有效提升语音识别的质量。

基于时频感知双路径Conformer的语音增强

近年来,Conformer在语音领域的应用表现较为突出。该模块通过结合多头自注意力机制和卷积神经网络,能够同时关注短时和长时序列信息,从而在语音处理任务中表现出卓越的性能。在此基础上提出了一种基于时频感知双路径Conformer的语音增强网络(TFDPCNet)。首先,该网络将改进的Conformer结构作为核心,采用双路径结构,构成时频感知的双路径Conformer模块(TFDP-Conformer),增强了整体网络的时频提取能力;同时,为了减小时频特征融合的难度,提出了注意力门控交叉融合模块(AGCF),通过额外的注意力门进一步增强了网络训练过程中时频特征的交互,提高了时频特征的利用率;最后,引用度量鉴别器,并对其进行适当剪枝,使得增强后的音频和原始音频在量化评价指标上保持更高的一致性。实验结果表明,相比于TSTNN算法,TFDPCNet在主观和客观指标上都有明显提高。

基于Conformer的实时多场景说话人识别模型

为解决在多场景(跨域、长时以及噪声干扰语音场景)下说话人确认系统性能较差的问题,提出了一种基于Conformer构建的、实时多场景鲁棒的说话人识别模型——PMS-Conformer。PMS-Conformer的设计灵感来自于先进的模型MFA-Conformer。PMS-Conformer对MFA-Conformer的声学特征提取器、网络组件和损失函数计算模块进行了改进,其具有新颖有效的声学特征提取器,以及鲁棒的、具有较强泛化能力的声纹嵌入码提取器。基于VoxCeleb1&2数据集实现了PMS-Conformer的训练;开展了PMS-Conformer与基线MFA-Conformer以及ECAPA-TDNN在说话人确认任务上的性能对比评估实验。实验结果表明在长语音SITW、跨域VoxMovies以及加噪处理的VoxCeleb-O测试集上,以PMS-Conformer构建的说话人确认系统的性能比用这两个基线构建的说话人确认系统更有竞争力;并且在声纹嵌入码提取器的可训练参数(Params)和推理速度(RTF)方面,PMS-Conformer明显优于ECAPA-TDNN。实验结果说明了PMS-Conformer在实时多场景下具有良好的性能。

基于Conformable分数阶导数的灰色Bernoulli模型

为增强灰色Bernoulli模型对各种实际数据序列的适应性,借助分数阶微积分在描述复杂系统中的优势,提出了一种基于Conformable分数阶导数的灰色Bernoulli模型。研究发现,可通过改变结构参数将模型转换为不同的经典灰色预测模型,体现了其统一性。此外,采用粒子群优化算法求解规划模型,获取了模型的最优超参数。最后,用所提模型和5个竞争模型对3个真实案例进行了预测建模,结果表明,所提模型的2项评估指标均优于5个竞争模型,验证了所提模型的有效性和可行性。

Laccase/caffeic acid-catalyzed crosslinking coupled with galactomannan alters the conformational structure of ovalbumin and alleviates Th2-mediated allergic asthma

Ovalbumin(OVA)is the major allergenic protein that can induce T helper 2(Th2)-allergic reactions,for which current treatment options are inadequate.In this study,we developed a polymerized hypoallergenic OVA product via laccase/caffeic acid(Lac/CA)-catalyzed crosslinking in conjunction with galactomannan(Man).The formation of high molecular weight crosslinked polymers and the Ig G-binding were analyzed by sodium dodecyl sulfate-polyacrylamide gel electrophoresis(SDS-PAGE)and Western blotting.The study indicated that Lac/CA-catalyzed crosslinking plus Man conjugation substantially altered secondary and tertiary structures of OVA along with the variation in surface hydrophobicity.Gastrointestinal digestion stability assay indicated that crosslinked OVA exhibited less resistance in simulated gastric fluid(SGF)and simulated intestinal fluid(SIF).Mouse model study indicated that Lac-Man/OVA ameliorated eosinophilic airway inflammatory response and efficiently downregulated the expression of Th2-related cytokines(interleukin(IL)-4,IL-5,and IL-13),and upregulated IFN-γand IL-10 expression.Stimulation of bone marrow-derived dendritic cells with Lac-Man/OVA suppressed the expression of phenotypic maturation markers(CD80 and CD86)and MHC class II molecules,and suppressed the expression levels of proinflammatory cytokines.The knowledge obtained in the present study offers an effective way to acquire a hypoallergenic OVA product that can have a therapeutic effect in alleviating OVA-induced allergic asthma.

Classifying rockburst with confidence:A novel conformal prediction approach

The scientific community recognizes the seriousness of rockbursts and the need for effective mitigation measures.The literature reports various successful applications of machine learning(ML)models for rockburst assessment;however,a significant question remains unanswered:How reliable are these models,and at what confidence level are classifications made?Typically,ML models output single rockburst grade even in the face of intricate and out-of-distribution samples,without any associated confidence value.Given the susceptibility of ML models to errors,it becomes imperative to quantify their uncertainty to prevent consequential failures.To address this issue,we propose a conformal prediction(CP)framework built on traditional ML models(extreme gradient boosting and random forest)to generate valid classifications of rockburst while producing a measure of confidence for its output.The proposed framework guarantees marginal coverage and,in most cases,conditional coverage on the test dataset.The CP was evaluated on a rockburst case in the Sanshandao Gold Mine in China,where it achieved high coverage and efficiency at applicable confidence levels.Significantly,the CP identified several“confident”classifications from the traditional ML model as unreliable,necessitating expert verification for informed decision-making.The proposed framework improves the reliability and accuracy of rockburst assessments,with the potential to bolster user confidence.

Gelatin-Based Metamaterial Hydrogel Films with High Conformality for Ultra-Soft Tissue Monitoring

Implantable hydrogel-based bioelectronics(IHB)can precisely monitor human health and diagnose diseases.However,achieving biodegradability,biocompatibility,and high conformality with soft tissues poses significant challenges for IHB.Gelatin is the most suitable candidate for IHB since it is a collagen hydrolysate and a substantial part of the extracellular matrix found naturally in most tissues.This study used 3D printing ultrafine fiber networks with metamaterial design to embed into ultra-low elastic modulus hydrogel to create a novel gelatin-based conductive film(GCF)with mechanical programmability.The regulation of GCF nearly covers soft tissue mechanics,an elastic modulus from 20 to 420 kPa,and a Poisson’s ratio from-0.25 to 0.52.The negative Poisson’s ratio promotes conformality with soft tissues to improve the efficiency of biological interfaces.The GCF can monitor heartbeat signals and respiratory rate by determining cardiac deformation due to its high conformability.Notably,the gelatin characteristics of the biodegradable GCF enable the sensor to monitor and support tissue restoration.The GCF metamaterial design offers a unique idea for bioelectronics to develop implantable sensors that integrate monitoring and tissue repair and a customized method for endowing implanted sensors to be highly conformal with soft tissues.

Recent advances in protein conformation sampling by combining machine learning with molecular simulation

The rapid advancement and broad application of machine learning(ML)have driven a groundbreaking revolution in computational biology.One of the most cutting-edge and important applications of ML is its integration with molecular simulations to improve the sampling efficiency of the vast conformational space of large biomolecules.This review focuses on recent studies that utilize ML-based techniques in the exploration of protein conformational landscape.We first highlight the recent development of ML-aided enhanced sampling methods,including heuristic algorithms and neural networks that are designed to refine the selection of reaction coordinates for the construction of bias potential,or facilitate the exploration of the unsampled region of the energy landscape.Further,we review the development of autoencoder based methods that combine molecular simulations and deep learning to expand the search for protein conformations.Lastly,we discuss the cutting-edge methodologies for the one-shot generation of protein conformations with precise Boltzmann weights.Collectively,this review demonstrates the promising potential of machine learning in revolutionizing our insight into the complex conformational ensembles of proteins.

3D‑Printed Carbon‑Based Conformal Electromagnetic Interference Shielding Module for Integrated Electronics

Electromagnetic interference shielding(EMI SE)modules are the core com-ponent of modern electronics.However,the tra-ditional metal-based SE modules always take up indispensable three-dimensional space inside electronics,posing a major obstacle to the integra-tion of electronics.The innovation of integrating 3D-printed conformal shielding(c-SE)modules with packaging materials onto core electronics offers infinite possibilities to satisfy ideal SE func-tion without occupying additional space.Herein,the 3D printable carbon-based inks with various proportions of graphene and carbon nanotube nanoparticles are well-formulated by manipulating their rheological peculiarity.Accordingly,the free-constructed architectures with arbitrarily-customized structure and multifunctionality are created via 3D printing.In particular,the SE performance of 3D-printed frame is up to 61.4 dB,simultaneously accompanied with an ultralight architecture of 0.076 g cm^(-3) and a superhigh specific shielding of 802.4 dB cm3 g^(-1).Moreover,as a proof-of-concept,the 3D-printed c-SE module is in situ integrated into core electronics,successfully replacing the traditional metal-based module to afford multiple functions for electromagnetic compatibility and thermal dissipa-tion.Thus,this scientific innovation completely makes up the blank for assembling carbon-based c-SE modules and sheds a brilliant light on developing the next generation of high-performance shielding materials with arbitrarily-customized structure for integrated electronics.

Momentum as Translations at Conformal Infinity

Although General Relativity is the classic example of a physical theory based on differential geometry, the momentum tensor is the only part of the field equation that is not derived from or interpreted with differential geometry. This work extends General Relativity and Einstein-Cartan theory by augmenting the Poincaré group with projective (special) conformal transformations, which are translations at conformal infinity. Momentum becomes a part of the differential geometry of spacetime. The Lie algebra of these transformations is represented by vectorfields on an associated Minkowski fiber space. Variation of projective conformal scalar curvature generates a 2-index tensor that serves as linear momentum in the field equations of General Relativity. The computation yields a constructive realization of Mach’s principle: local inertia is determined by local motion relative to mass at conformal infinity in each fiber. The vectorfields have a cellular structure that is similar to that of turbulent fluids.

Exact Solutions and Finite Time Stability of Linear Conformable Fractional Systems with Pure Delay

We study nonhomogeneous systems of linear conformable fractional differential equations with pure delay.By using new conformable delayed matrix functions and the method of variation,we obtain a representation of their solutions.As an application,we derive a finite time stability result using the representation of solutions and a norm estimation of the conformable delayedmatrix functions.The obtained results are new,and they extend and improve some existing ones.Finally,an example is presented to illustrate the validity of our theoretical results.

On Fuzzy Conformable Double Laplace Transform with Applications to Partial Differential Equations

The Laplace transformation is a very important integral transform,and it is extensively used in solving ordinary differential equations,partial differential equations,and several types of integro-differential equations.Our purpose in this study is to introduce the notion of fuzzy double Laplace transform,fuzzy conformable double Laplace transform(FCDLT).We discuss some basic properties of FCDLT.We obtain the solutions of fuzzy partial differential equations(both one-dimensional and two-dimensional cases)through the double Laplace approach.We demonstrate through numerical examples that our proposed method is very successful and convenient for resolving partial differential equations.

What Is the Conformity-Nonconformity Dichotomy?

This paper argues that the conformity-nonconformity dichotomy is a false dilemma.This study first critically reviews the basic philosophical,ethical,social psychological,and pedagogical literature related to the two concepts.It then outlines the way to overcome the phenomena of conformism and nonconformism together.The description of conformity and nonconformity as deprivation of freedom becomes stronger in 20th century philosophy and special literature from Heidegger through Fischer’s definition up to Cooley,and Crutchfield.Conformity is the sinking of the Self into the Anyone,the unprincipled alignment to the opinion of group mates,and nonconformity is the unprincipled resistance to it.But what is beyond conformity and nonconformity together as a group?There is a real community,in it the transformation of our pedagogical culture in a both useful and reasonable manner to allow the youth to accept the world by denying it and to deny the world by accepting it.The real community involves the virtue of goodness.Educate for goodness,because we possibly are the honest and humane man,who disregards the sinking of the self into the anyone and the self-contained rebellion.

基于多尺度阶梯时频Conformer GAN的语音增强算法

针对频率域语音增强算法中因相位混乱产生人工伪影,导致去噪性能受限、语音质量不高的问题,提出一种基于多尺度阶梯型时频Conformer生成对抗网络(MSLTF-CMGAN)的语音增强算法。将语音语谱图的实部、虚部和振幅谱作为输入,生成器首先在多个尺度上利用时间-频率Conformer学习时域和频域的全局及局部特征依赖;其次,利用Mask Decoder分支学习振幅掩码,而Complex Decoder分支则直接学习干净的语谱图,融合这两个Decoder分支的输出可得到重建后的语音;最后,利用指标判别器判别语音的评价指标得分,通过极大极小训练使生成器生成高质量的语音。采用主观评价平均意见得分(MOS)和客观评价指标在公开数据集VoiceBank+Demand上与各类语音增强模型进行对比,结果显示,所提算法的MOS信号失真(CSIG)和MOS噪声失真(CBAK)比目前最先进的方法CMGAN(基于Conformer的指标生成对抗网络语音增强模型)分别提高了0.04和0.07,尽管它的MOS整体语音质量(COVL)和语音质量的感知评估(PESQ)略低于CMGAN,但与其他对比模型相比在多项主客观语音质量评估方面的评分均处于领先水平。

基于Conformer的时域多通道语音分离方法

多通道语音中的空间特征信息为说话人分离提供了重要的线索,为了更好地提取通道间信息并有效降低网络的处理时延,提出一种多通道时域语音分离方法。利用多层编码器实现语音特征提取并挖掘通道间信息,在逐层编码过程中获得不同时间分辨率的语音特征并降低特征时间维度;引入Conformer结构对语音全局时间关系进行建模,在解码阶段使用特征加权跳跃连接融合对应编码层的输出特征进行解码,并将高维语音特征恢复为时域信号。在基于LibriSpeech仿真的多通道混响带噪语音数据集中进行实验,实验结果表明,所提方法通过多层编解码机制充分利用了多通道语音信息并降低了网络处理时延,通过Conformer实现并行数据处理和全局时间关系建模,在推理速度、分离语音质量和语音感知质量方面均优于基线单通道和多通道时域语音分离算法。

使用Conformer增强的混合CTC/Attention端到端中文语音识别

最近,基于自注意力的Transformer结构在不同领域的一系列任务上表现出非常好的性能。探索了基于Transformer编码器和LAS(listen,attend and spell)解码器的Transformer-LAS语音识别模型的效果,并针对Transformer不善于捕捉局部信息的问题,使用Conformer代替Transformer,提出Conformer-LAS模型。由于Attention过于灵活的对齐方式,使得在嘈杂环境中的效果急剧下降,采用连接时序分类(connectionist temporal classification,CTC)辅助训练以加快收敛,并加入音素级别的中间CTC损失联合优化,提出了效果更好的Conformer-LAS-CTC语音识别模型。在开源中文普通话Aishell-1数据集上对提出来的模型进行验证,实验结果表明,Conformer-LAS-CTC相对于采用的基线BLSTM-LAS和Transformer-LAS模型在测试集上的字错率分别相对降低了22.58%和48.76%,模型最终字错误率为4.54%。

基于U-Conformer的多特征融合鸟鸣声分离方法

针对多个鸟类个体同时发声导致的鸣声混叠问题,本文提出了一种融合录音通道间空间特征的鸟类鸣声分离方法.该方法将混叠鸣声信号的声谱特征和空间特征作为分离模型的输入,提出深度学习模型U-Conformer来预测每个鸣声源方向的幅值谱掩膜(spectral magnitude mask,SMM),通过模型估计的SMM从混叠鸣声信号中恢复每个鸣声源信号.由多源混叠鸟类鸣声数据的实验结果表明,本文提出的分离方法较其他深度学习模型结构具有更好的分离效果,有助于更好地分析野外鸟类鸣声录音.

On Nonlinear Conformable Fractional Order Dynamical System via Differential Transform Method

This article studies a nonlinear fractional order Lotka-Volterra prey-predator type dynamical system.For the proposed study,we consider the model under the conformable fractional order derivative(CFOD).We investigate the mentioned dynamical system for the existence and uniqueness of at least one solution.Indeed,Schauder and Banach fixed point theorems are utilized to prove our claim.Further,an algorithm for the approximate analytical solution to the proposed problem has been established.In this regard,the conformable fractional differential transform(CFDT)technique is used to compute the required results in the form of a series.Using Matlab-16,we simulate the series solution to illustrate our results graphically.Finally,a comparison of our solution to that obtained for the Caputo fractional order derivative via the perturbation method is given.

语音识别中的Conformer模型压缩研究

针对使用Conformer模型的语音识别算法在实际应用时设备算力不足及资源缺乏的问题,提出一种基于Conformer模型间隔剪枝和参数量化相结合的模型压缩方法。实验显示,使用该方法压缩后,模型的实时率(real time factor, RTF)达到0.107614,较基线模型的推理速度提升了16.2%,而识别准确率只下降了1.79%,并且模型大小也由原来的207.91MB下降到72.69MB。该方法在模型准确率损失很小的情况下,较大程度地提升了模型的适用性。