基于得分归一化和系统融合的语音关键词检测方法

发布时间：2018-06-23 13:39

本文选题：关键词检测 + 得分归一化　；参考：《数据采集与处理》2017年02期

【摘要】：为了有效利用不同关键词检测系统的互补性,解决不同系统检测结果置信度得分不在同一范围的问题,提出了一种基于得分规整和系统融合的语音关键词检测方法。首先,为了克服连续语音识别系统中因剪枝错误而引起的关键词丢失问题,应用了关键词相关的软Beam宽度剪枝策略裁剪词图;其次,在系统融合前采用得分归一化方法,使得不同系统关键词检测结果置信度得分在同一范围;最后,通过系统融合处理将不同系统的关键词输出进行整合,得到最终的关键词检测结果。实验结果表明,经过得分归一化处理后,关键词检测性能的实际查询词权重代价(Actual term-weighted value,ATWV)平均相对提升30%;系统融合后关键词的检测性能,相比于得分归一化处理后的最佳单一系统,得到了10%的提升。
[Abstract]:In order to effectively utilize the complementarities of different keyword detection systems and solve the problem that the confidence score of different detection results is not in the same range, a speech keyword detection method based on score regularity and system fusion is proposed. First of all, in order to overcome the loss of keywords caused by pruning errors in continuous speech recognition system, the keyword related soft Beam width pruning strategy is applied to clipping the word map. Secondly, the score normalization method is adopted before the system fusion. The confidence score of different system keyword detection results is in the same range. Finally, the key word output of different systems is integrated by system fusion processing, and the final keyword detection results are obtained. The experimental results show that after the normalized score processing, the actual query word weight cost (Actual term-weighted value) of keyword detection performance is improved by 30% on average, and the detection performance of keywords after system fusion is improved by 30%. Compared with the best single system after normalized score processing, a 10% improvement was obtained.
【作者单位】：解放军信息工程大学信息系统工程学院;
【基金】：国家自然科学基金(61175017)资助项目
【分类号】：TN912.34

【相似文献】