Putonghua prosody is characterized by its hierarchical structure when influenced by linguistic environments. Based on this, a neural network, with specially weighted factors and optimizing outputs, is described and ap...Putonghua prosody is characterized by its hierarchical structure when influenced by linguistic environments. Based on this, a neural network, with specially weighted factors and optimizing outputs, is described and applied to construct the Putonghua prosodic model in Text-to-Speech (TTS) system. Extensive tests show that the structure of the neural network characterizes the Putonghua prosody more exactly than traditional models. Learning rate is speeded up and computational precision is improved, which makes the whole prosodic model more efficient. Furthermore, the paper also stylizes the Putonghua syllable pitch contours with SPiS parameters (Syllable Pitch Stylized Parameters), and analyzes them in adjusting the syllable pitch. It shows that the SPiS parameters effectively characterize the Putonghua syllable pitch contours, and facilitate the establishment of the network model and the prosodic controlling.展开更多
针对目前BNN(Binarized Neural Network)剪枝方法存在剪枝比例低、识别准确率显著下降以及依赖训练后微调的问题,提出了一种基于三值向二值演化的滤波器级的BNN剪枝方法,命名为ETB(Evolution from Ternary to Binary)。ETB是基于学习的...针对目前BNN(Binarized Neural Network)剪枝方法存在剪枝比例低、识别准确率显著下降以及依赖训练后微调的问题,提出了一种基于三值向二值演化的滤波器级的BNN剪枝方法,命名为ETB(Evolution from Ternary to Binary)。ETB是基于学习的,通过在BNN的量化函数中引入可训练的量化阈值,使权重和激活值逐渐从三值演化到二值或零,旨在使网络在训练期间自动识别不重要的结构。此外,一个剪枝率调节算法也被设计用于调控网络的剪枝率。训练后,全零滤波器和对应的输出通道可被直接裁剪而获得精简的BNN,无需微调。为证明提出方法的可行性和其提升BNN推理效率而不牺牲准确率的潜力,在CIFAR-10上进行实验:在CIFAR-10数据集上,ETB对VGG-Small模型进行了46.3%的剪枝,模型大小压缩至0.34 MByte,准确率为89.97%,并在ResNet-18模型上进行了30.01%的剪枝,模型大小压缩至1.33 MByte,准确率为90.79%。在准确率和参数量方面,对比一些现有的BNN剪枝方法,ETB具有一定的优势。展开更多
基金This work was supported by the National Natural Science Foundation of China (69875008) and 863National High Technology Project
文摘Putonghua prosody is characterized by its hierarchical structure when influenced by linguistic environments. Based on this, a neural network, with specially weighted factors and optimizing outputs, is described and applied to construct the Putonghua prosodic model in Text-to-Speech (TTS) system. Extensive tests show that the structure of the neural network characterizes the Putonghua prosody more exactly than traditional models. Learning rate is speeded up and computational precision is improved, which makes the whole prosodic model more efficient. Furthermore, the paper also stylizes the Putonghua syllable pitch contours with SPiS parameters (Syllable Pitch Stylized Parameters), and analyzes them in adjusting the syllable pitch. It shows that the SPiS parameters effectively characterize the Putonghua syllable pitch contours, and facilitate the establishment of the network model and the prosodic controlling.
文摘针对目前BNN(Binarized Neural Network)剪枝方法存在剪枝比例低、识别准确率显著下降以及依赖训练后微调的问题,提出了一种基于三值向二值演化的滤波器级的BNN剪枝方法,命名为ETB(Evolution from Ternary to Binary)。ETB是基于学习的,通过在BNN的量化函数中引入可训练的量化阈值,使权重和激活值逐渐从三值演化到二值或零,旨在使网络在训练期间自动识别不重要的结构。此外,一个剪枝率调节算法也被设计用于调控网络的剪枝率。训练后,全零滤波器和对应的输出通道可被直接裁剪而获得精简的BNN,无需微调。为证明提出方法的可行性和其提升BNN推理效率而不牺牲准确率的潜力,在CIFAR-10上进行实验:在CIFAR-10数据集上,ETB对VGG-Small模型进行了46.3%的剪枝,模型大小压缩至0.34 MByte,准确率为89.97%,并在ResNet-18模型上进行了30.01%的剪枝,模型大小压缩至1.33 MByte,准确率为90.79%。在准确率和参数量方面,对比一些现有的BNN剪枝方法,ETB具有一定的优势。
