在本文中，我們提出了一種GRU結構的變體，稱為雙序列門控注意單元（DS-GAU），其中計算了x向量基線的每個TDNN層的統計池，並將其通過DS-GAU層傳遞， 在訓練為幀級時從輸入要素的不同時間上下文中聚合更多信息。 我們提出的架構在VoxCeleb2數據集上進行了訓練，其中特徵向量稱為DSGAU-向量。 我們對VoxCeleb1數據集和“野生演說者”（SITW）數據集進行了評估，並將實驗結果與x矢量基線系統進行了比較。 結果表明，相對於VoxCeleb1數據集的x向量基線，我們提出的方法在EER相對改進方面最多可存檔11.6％，7.9％和7.6％.;In this thesis, we present a variant of GRU architecture called Dual-Sequences Gated Attention Unit (DS-GAU), in which the statistics pooling from each TDNN layer of the x-vector baseline are computed and passed through the DS-GAU layer, to aggregate more information from the variant temporal context of input features while training as frame-level. Our proposed architecture was trained on the VoxCeleb2 dataset, where the feature vector is referred to as a DSGAU-vector. We made our evaluation on the VoxCeleb1 dataset and the Speakers in the Wild (SITW) dataset and compared the experimental results with the x-vector baseline system. It showed that our proposed method archived up to 11.6%, 7.9%, and 7.6% in EER relative improvements over the x-vector baseline on the VoxCeleb1 dataset.
