Zhang Qian, Hansen John H L
Center for Robust Speech Systems (CRSS), Erik Jonsson School of Engineering and Computer Science, University of Texas at Dallas, Richardson, Texas 75080, USA.
J Acoust Soc Am. 2018 Jan;143(1):418. doi: 10.1121/1.5017608.
Research in open-set language identification (LID) generally focuses on in-set language modeling versus out-of-set (OOS) language rejection. However, unknown/OOS language rejection is essential for effective speech and language pre-processing. To address this, an approach for OOS language selection is proposed. Using probe OOS data, three effective OOS candidate selection methods are developed for universal OOS language coverage. The selected OOS candidates are expected to reflect the entire OOS language space for the state-of-the-art i-vector LID system followed by a Gaussian back-end. Two front-end feature selection strategies are proposed: (i) unsupervised k-means clustering and (ii) complementary candidate selection. Also, (iii) general candidate selection is proposed according to language relationship explored at the score level. All methods are evaluated on a large-scale corpus (LRE-09) containing 40 languages. The proposed selection methods reduce OOS training data diversity by 86% while achieving performance similar to closed-set using all probe OOS for training. The proposed methods also show clear benefits versus random candidate selection (i.e., the proposed solutions achieve sustained performance while employing a minimum number of effective OOS language candidates). To the best of our knowledge, this is the first major effort on effective OOS language selection and enhancement for improved OOS rejection in open-set LID.
开放集语言识别（LID）的研究通常侧重于集合内语言建模与集合外（OOS）语言拒绝。然而，未知/集合外语言拒绝对于有效的语音和语言预处理至关重要。为了解决这个问题，提出了一种集合外语言选择方法。利用探测性集合外数据，开发了三种有效的集合外候选选择方法，以实现通用的集合外语言覆盖。所选的集合外候选语言有望反映采用高斯后端的最先进i-vector LID系统的整个集合外语言空间。提出了两种前端特征选择策略：（i）无监督k均值聚类和（ii）互补候选选择。此外，（iii）根据在分数层面探索的语言关系提出了通用候选选择。所有方法均在包含40种语言的大规模语料库（LRE-09）上进行评估。所提出的选择方法将集合外训练数据的多样性降低了86%，同时在使用所有探测性集合外数据进行训练时实现了与闭集相似的性能。与随机候选选择相比，所提出的方法也显示出明显的优势（即，所提出的解决方案在使用最少数量的有效集合外语言候选的情况下实现了持续的性能）。据我们所知，这是在开放集LID中为改进集合外拒绝而进行的有效集合外语言选择和增强方面的首次重大努力。
