ID 原文 译文
583 为了充分利用含噪语音特征来提高语音增强网络的性能，基于含噪语音在时间和频率两个维度上的相关性，本文结合卷积神经网络的局部特征提取能力和门控循环单元的长期依赖建模能力，设计了一种适用于语音增强的卷积门控循环网络。 In order to improve the performance of speech enhancement networks by making full use of noisy speech features, based on the correlation of noisy speech in time and frequency, by combining the local feature extraction ability of convolutional neural networks and the long-term dependence modeling ability of gated recurrent unit, a convolutional gated recurrent network suitable for speech enhancement is designed in this paper.
584 该网络采用卷积网络结构代替全连接网络结构来改进门控循环单元中的特征计算过程，从而能够更好地保留含噪语音特征中的时频结构信息。 This network uses a convolutional network structure instead of a fully connected network structure to improve the feature calculation process in the gated recurrent unit, thereby can better retain the time-frequency structure in the noisy speech features.
585 实验结果表明，与其它语音增强网络相比，本文网络在语音成分的保留和噪声成分的抑制上具有明显优势，增强后语音具有更好的语音质量和可懂度。 The experimental results show that com-pared with other speech enhancement networks, the proposed network has obvious advantages in speech component retentionand noise component suppression, and the enhanced speech has better speech quality and intelligibility.
586 图像超分辨率重建(Super-resolution Reconstruction，SR)是由一张或多张低分辨率图像得到高分辨率图像的过程。 Image super-resolution reconstruction (SR)aims to obtain high-resolution images from one or more low-resolution images.
587 近年来，SR 技术不断发展，在许多领域被广泛应用。 Recently, SR has been developing and widely applied in different fields.
588 本文在回顾 SR 技术发展历史的基础上，全面综述了 SR技术在各个时期的代表性方法，重点介绍了基于深度学习的图像超分辨率工作。 This survey retrospects the history of SR technique and provides a comprehensive overview of representative SR methods, with an emphasis on recent deeplearning-based approaches.
589 我们从模型类型、网络结构、信息传递方式等方面对各种算法进行了详细评述，并对比了其优缺点。 We elaborate the details of various deep learning-based SR methods, including their strengths andweakness, in terms of the deep learning model, architecture, and message pass.
590 最后探讨了图像超分辨率技术未来的发展方向。 Finally, we discuss the possible research di-rections on SR technique.
591 注意力模型是当前语音识别中的主流模型，然而其存在一个缺点，即当前时刻的注意力模型可能产生异常得分。 Attention-based model is a popular model in speech recognition, however it has a disadvantage that the at-tention-based model may produce abnormal scores.
592 为此，本文首先提出前向注意力模型，其采用上一时刻正常注意力得分平滑当前时刻异常得分。 To solve this problem, this paper first proposes a forward attention mod-el, which adopts normal attention score at the previous moment to smooth the abnormal score at the current moment.