CAJ | 학술논문

针对目前生活中涌现的海量语音数据,人们对语音检索技术准确度的要求越来越高.主要研究了汉语连续语音检索任务中,基于转换音节网格的研究方法.针对语音检索系统中置信度计算的问题,提出了一种基于音节间互信息的置信度计算方法,并将其用于网格结构的语音检索系统中.该方法能够有效地利用上下文之间的互信息量,从而更准确、合理地描述汉语语言模型.实验结果表明,用提出的方法建立转换音节网格来进行语音检索,其检出率(FOM)比后验概率法和N-best法有较大幅度的提高.得到的汉语语音检索系统其FOM最高可以达到83.7%.
침대목전생활중용현적해량어음수거,인문대어음검색기술준학도적요구월래월고.주요연구료한어련속어음검색임무중,기우전환음절망격적연구방법.침대어음검색계통중치신도계산적문제,제출료일충기우음절간호신식적치신도계산방법,병장기용우망격결구적어음검색계통중.해방법능구유효지이용상하문지간적호신식량,종이경준학、합리지묘술한어어언모형.실험결과표명,용제출적방법건립전환음절망격래진행어음검색,기검출솔(FOM)비후험개솔법화N-best법유교대폭도적제고.득도적한어어음검색계통기FOM최고가이체도83.7%.
Nowadays, with the overwhelming amounts of speech data rushing in our life, higher and higher accuracy of speech indexing techniques is required. This paper mainly studied a converted syllable lattice-based approach in a Chinese continuous speech indexing task. Aiming at the computation of confidence measure in a speech indexing system, this paper proposed a confidence measure method based on mutual information between syllables, which was used in a lattice construction system for speech indexing. The method took full advantage of the context mutual information, which could describe Chinese language model more exactly and logically. The experiment results show that using the proposed method to build a converted syllable lattice in a speech indexing system, the FOM of which has great improvement comparing with posterior probability based method and N-best based method. This best system for Chinese speech indexing achieves a FOM of 83.7%.