提出一种约束条件下的结构化高斯混合模型及非平行语料语音转换方法.从源与目标说话人的原始非平行语料中提取出少量相同音节,在结构化高斯混合模型的训练过程中,利用这些相同音节包含的语义信息及声学特征对应关系对K均值聚类中心进行...提出一种约束条件下的结构化高斯混合模型及非平行语料语音转换方法.从源与目标说话人的原始非平行语料中提取出少量相同音节,在结构化高斯混合模型的训练过程中,利用这些相同音节包含的语义信息及声学特征对应关系对K均值聚类中心进行约束,并在(Expectation Maximum,EM)迭代过程中对语音帧属于模型分量的后验概率进行修正,得到基于约束的结构化高斯混合模型(Structured Gaussian Mixture Model with Constraint condition,CSGMM).再利用全局声学结构(Acoustic Universal Structure,AUS)原理对源和目标说话人的约束结构化高斯混合模型的高斯分布进行匹配对准,推导出短时谱转换函数.主观和客观评价实验结果表明,使用该方法得到的转换后语音在谱失真,目标倾向性和语音质量等方面均优于传统的结构化模型语音转换方法,转换语音的平均谱失真仅为0.52,说话人正确识别率达到95.25%,目标语音倾向性指标ABX平均为0.82,性能更加接近于基于平行语料的语音转换方法.展开更多
追踪研究当中,交叉滞后模型可以探究多变量之间往复式影响,潜增长模型可以探究个体增长趋势。对两类模型进行整合,例如同时关注往复式影响与个体增长趋势,同时可以定义测量误差、随机截距等变异成分,衍生出随机截距交叉滞后模型、特质−...追踪研究当中,交叉滞后模型可以探究多变量之间往复式影响,潜增长模型可以探究个体增长趋势。对两类模型进行整合,例如同时关注往复式影响与个体增长趋势,同时可以定义测量误差、随机截距等变异成分,衍生出随机截距交叉滞后模型、特质−状态−误差模型、自回归潜增长模型、结构化残差潜增长模型等。以交叉滞后模型和潜增长模型分别作为基础模型,从个体间/个体内变异分解的角度对上述各类模型梳理,整合出此类模型的分析框架,并拓展建立“因子结构化潜增长模型(factor latent curve model with structured reciprocals)”作为统合框架。通过实证研究(早期儿童的追踪研究−幼儿园版,ECLS-K),建立21049名儿童的阅读和数学能力的往复式影响与增长趋势。研究发现,分离了稳定特质的模型拟合最优。研究也对模型建模思路和模型选择提供了建议。展开更多
The optimum models of harvesting yield and net profits of large diameter trees for broadleaved forest were developed, of which include matrix growth sub-model, harvesting cost and wood price sub-models, based on the d...The optimum models of harvesting yield and net profits of large diameter trees for broadleaved forest were developed, of which include matrix growth sub-model, harvesting cost and wood price sub-models, based on the data from Hongshi Forestry Bureau, in Changbai Mountain region, Jilin Province, China. The data were measured in 232 permanent sample plots. With the data of permanent sample plots, the parameters of transition probability and ingrowth models were estimated, and some models were compared and partly modified. During the simulation of stand structure, four factors such as largest diameter residual tree （LDT）, the ratio of the number of trees in a given diameter class to those in the next larger diameter class （q）, residual basal area （RBA） and selective cutting cycle （C） were considered. The simulation results showed that the optimum stand structure parameters for large diameter trees are as follows： q is 1.2, LDT is 46cm, RBA is larger than 26 m^2 and selective cutting cycle time （C） is between 10 and 20 years.展开更多
文摘提出一种约束条件下的结构化高斯混合模型及非平行语料语音转换方法.从源与目标说话人的原始非平行语料中提取出少量相同音节,在结构化高斯混合模型的训练过程中,利用这些相同音节包含的语义信息及声学特征对应关系对K均值聚类中心进行约束,并在(Expectation Maximum,EM)迭代过程中对语音帧属于模型分量的后验概率进行修正,得到基于约束的结构化高斯混合模型(Structured Gaussian Mixture Model with Constraint condition,CSGMM).再利用全局声学结构(Acoustic Universal Structure,AUS)原理对源和目标说话人的约束结构化高斯混合模型的高斯分布进行匹配对准,推导出短时谱转换函数.主观和客观评价实验结果表明,使用该方法得到的转换后语音在谱失真,目标倾向性和语音质量等方面均优于传统的结构化模型语音转换方法,转换语音的平均谱失真仅为0.52,说话人正确识别率达到95.25%,目标语音倾向性指标ABX平均为0.82,性能更加接近于基于平行语料的语音转换方法.
文摘追踪研究当中,交叉滞后模型可以探究多变量之间往复式影响,潜增长模型可以探究个体增长趋势。对两类模型进行整合,例如同时关注往复式影响与个体增长趋势,同时可以定义测量误差、随机截距等变异成分,衍生出随机截距交叉滞后模型、特质−状态−误差模型、自回归潜增长模型、结构化残差潜增长模型等。以交叉滞后模型和潜增长模型分别作为基础模型,从个体间/个体内变异分解的角度对上述各类模型梳理,整合出此类模型的分析框架,并拓展建立“因子结构化潜增长模型(factor latent curve model with structured reciprocals)”作为统合框架。通过实证研究(早期儿童的追踪研究−幼儿园版,ECLS-K),建立21049名儿童的阅读和数学能力的往复式影响与增长趋势。研究发现,分离了稳定特质的模型拟合最优。研究也对模型建模思路和模型选择提供了建议。
基金This paper was supported by National Strategy Key Project, Research and Paradigm on Ecological Harvesting and Regeneration Tech-nique for Northeast Natural Forest (2001BA510B07-02)
文摘The optimum models of harvesting yield and net profits of large diameter trees for broadleaved forest were developed, of which include matrix growth sub-model, harvesting cost and wood price sub-models, based on the data from Hongshi Forestry Bureau, in Changbai Mountain region, Jilin Province, China. The data were measured in 232 permanent sample plots. With the data of permanent sample plots, the parameters of transition probability and ingrowth models were estimated, and some models were compared and partly modified. During the simulation of stand structure, four factors such as largest diameter residual tree （LDT）, the ratio of the number of trees in a given diameter class to those in the next larger diameter class （q）, residual basal area （RBA） and selective cutting cycle （C） were considered. The simulation results showed that the optimum stand structure parameters for large diameter trees are as follows： q is 1.2, LDT is 46cm, RBA is larger than 26 m^2 and selective cutting cycle time （C） is between 10 and 20 years.
