排序方式: 共有3条查询结果,搜索用时 15 毫秒
1
1.
2.
提出了一种联合深度编解码神经网络和时频掩蔽估计的语音增强方法。该方法利用深度编解码网络估计时频掩蔽表示,并联合带噪语音的幅度谱学习带噪语音与纯净语音幅度谱之间的非线性映射关系。深度编解码网络采用卷积-反卷积网络结构。在编码端,利用卷积网络的局部感知特性,对带噪语音的时频域结构特征进行建模,提取语音特征,同时抑制背景噪声。在解码端,利用编码端提取到的语音特征逐层恢复局部细节信息并重构语音信号。同时,在编解码端对应层之间引入跳跃连接,以减少由于池化和全连接操作导致的低层细节信息丢失的问题。在TIMIT语音库和不完全匹配噪声集下进行仿真实验,实验结果表明,该方法可以有效抑制噪声,且能较好地恢复出语音细节成分。 相似文献
3.
Fabrication of Ge Nano-Dot Heterojunction Phototransistors for Improved Light Detection at 1.55μm 总被引:1,自引:0,他引:1 下载免费PDF全文
Heterojunction phototransistors (HPTs) with several Ge/Si nano-dot layers as the absorption region are fabricated to obtain improved light detectivity at 1.55μm. The HPT detectors are of n-p-n type with ten layers of Ge(8ML ) /Si(45nm) incorporated in the base-collector junction and are grown by an ultrahigh-vacuum chemicalvapor-deposition system. The detectors are operated with normal incidence. Because of the good quality of the grown material and fabrication process, the dark current is only 0.71pA/μm^2 under 5 V bias and the breakdown voltage is over 20 V. Compared to the positive-intrinsic-negative (PIN) reference detector with the same absorption layer, the responsivity is improved over 17 times for normal incidence at 1.55μm. 相似文献
1