ID 原文 译文
593 接着通过对上一时刻的注意力得分添加约束因子来对前向注意力模型进行优化，达到自适应平滑的目的。 Then, the model is optimized to add constraint factors to the attention score at the previous moment to achieve the purpose of adap-tive smoothing of the above abnormal scores.
594 最后，在优化模型基础上提出多尺度前向注意力模型，其通过引入多尺度模型来对不同等级的语音基元进行建模，进而将所得到的不同等级目标向量进行融合，以达到解决注意力得分异常值的目的。 Then, a multi-scale forward attention model is proposed on the above model. This model introduces a multi-scale method to model the speech primitives of different levels, and then fuses the target vec-tors of different levels to solve the outliers of attention score.
595 采用 SwitchBoard 作为训练集，Hub5'00 作为测试集进行实验。 In the experiment, SwitchBoard is adopted as the training setand Hub5' 00 as the test set.
596 相比于基线系统，多尺度前向注意力模型的词错误率(Word Error Rate，WER)相对降低 14.28% 。 Compared with the baseline system, the Word Error Rate (WER)of the proposed system de-creased by 14. 28% relatively.
597 飞行员疲劳状态识别面临两个重要问题，如何提取表征疲劳的特征以及如何对疲劳特征建模学习。 Pilots' fatigue state recognition faces two important issues: how to extract the characteristics that character-ize fatigue and how to model fatigue characteristics.
598 首先提取脑电信号节律波，计算基于仿射伪平滑 Wigner-Ville 分布的瞬时频域信息，构建疲劳状态指标。 Firstly, the EEG (ElectroEncephaloGram) signal is extracted, and the in-stantaneous frequency domain information based on the affine pseudo-smooth Wigner-Ville distribution is calculated to con-struct the fatigue state index.
599 其次，基于脑电信号各通道的周期性变化提出 Gamma 深度信念网络的疲劳状态分类算法。 Secondly, based on the periodic changes of each channel of EEG signals, the fatigue state classi-fication algorithm of Gamma deep belief network is proposed.
600 与采用卷积与池化运算的学习网络不同，Gamma 深度信念网络没有将图像或信号按尺度分割，但在底部的隐藏层已经可以有效地学习特定区域的特征，且当层数增加时，可有效提取特征的区域增多，学习到的特征更为一般化。 Unlike other learning network using convolution and pooling, the proposed network does not split the image or signal, but the hidden layer at the bottom can effectively learn the features of a specific region, and when the number of layers increases, the number of features increases and the features are more gen-eral.
601 然后改进用于训练深度信念网络的 Gibbs 采样算法，提出向上向下 Gibbs 采样以推断网络参数。 The Gibbs sampling algorithm for training the deep belief network is improved. The up-down Gibbs sampling is pro-posed to infer the network parameters.
602 最后，实验结果显示，本文的 Gamma 深度信念网络在识别准确率、稳定性、迭代用时等方面均达到了令人满意的效果。 Finally, the experimental results show that the Gamma deep belief network in this pa-per has achieved satisfactory results in terms of recognition accuracy, stability and iteration time.