低资源语言的无监督语音关键词检测技术综述

发布时间：2018-03-20 15:00

本文选题：检测　切入点：低资源　出处：《中国图象图形学报》2015年02期 　论文类型：期刊论文

【摘要】：目的低资源(low-resource)语言的无监督的关键词检测技术近年来引起了广泛的研究兴趣。低资源语言由于缺乏足够的标注数据及相关的专家知识,使得传统的基于大词汇量语音识别系统的关键词检测技术无法使用。近年来,研究者试图寻找一种无监督的技术来完成针对低资源语言的语音关键词检测。方法首先阐述了该技术目前面临的问题与挑战,然后介绍了该技术使用的主流的基于动态时间规整的算法框架,并从特征表示、模板匹配方法、效率提升等几个重要方面介绍了近几年来主要的研究成果,最后介绍了该任务常用的系统评价标准及目前所能达到的水平,讨论了未来可能的研究方向。结果该任务的研究目前取得了很多成果,但仍处于实验室阶段,多系统融合策略导致系统庞大,而且目前还没有好的进行索引的方法,导致检测时间过长,对于低资源语音的关键词检测技术,还有很多研究工作要做。结论期望通过对目前低资源语言的无监督的关键词检测技术做出一个全面的综述,从而给研究者的工作带来便利。
[Abstract]:Objective in recent years, the unsupervised keyword detection technique in low-resource resource language has attracted wide research interest. Due to the lack of sufficient tagging data and related expert knowledge, low-resource language has attracted more and more attention in recent years. In recent years, the traditional keyword detection technology based on large vocabulary speech recognition system can not be used. Researchers are trying to find an unsupervised technique to detect speech keywords in low-resource languages. Then it introduces the mainstream algorithm framework based on dynamic time warping used in this technology, and introduces the main research results in recent years from several important aspects, such as feature representation, template matching method, efficiency improvement and so on. At last, it introduces the system evaluation standard and the level that can be achieved at present, and discusses the possible research direction in the future. Results the research on this task has made a lot of achievements at present, but it is still in the laboratory stage. The multi-system fusion strategy leads to the huge system, and there is no good indexing method, which leads to the detection time is too long, for low-resource voice keyword detection technology, Conclusion A comprehensive review of unsupervised keyword detection techniques in low-resource languages is expected to facilitate the work of researchers.
【作者单位】：西北工业大学计算机学院陕西省语音与图像信息处理重点实验室;
【基金】：国家自然科学基金项目(61175018) 霍英东青年教师基础研究基金项目(131059)
【分类号】：TN912.3

【参考文献】