Pairwise Learning using Unsupervised Bottleneck Features for Zero-Resource Speech Challenge 2017 (System 3)

The system is for track1 alone. We trained an antoencoder using unsupervised bottleneck features with word-pair information from unsupervised term detection (UTD) on all corpora of five languages. The unsupervised bottleneck features was extracted from an extractor of multi-task learning deep neural networks (MTL-DNN). The word-pair was found by UTD. The UTD process was built on ZRTools. The final features are obtained from the third layer in our pairwise trained autoencoder.

本系统仅适用于赛道1（track1）。我们基于五种语言的全部语料，结合无监督术语检测（unsupervised term detection, UTD）提取的词对信息，使用无监督瓶颈特征完成了自动编码器（autoencoder）的训练。该无监督瓶颈特征由多任务学习深度神经网络（multi-task learning deep neural networks, MTL-DNN）的特征提取器提取得到。上述词对通过无监督术语检测（UTD）获取。该无监督术语检测流程基于ZRTools构建。最终特征取自我们训练的成对式自动编码器的第三层。