【作者】 黎塔； 蔡尚； 赵庆卫； 潘接林； 颜永红；

【Author】 Li Ta Cai Shang Zhao Qingwei Pan Jielin Yan Yonghong(Think IT Speech Laboratory,Institute of Acoustics,Chinese Academy of Sciences,Beijing 100190,China)

【机构】 中国科学院声学研究所中科信利语音实验室；

【摘要】 为了满足超大词表语法的识别任务在嵌入式语音识别系统上的应用,提出了一种高效的双层图搜索算法.该算法通过分离声学层和词法层来构建2层图搜索空间,其中声学层记录声学模型相关信息,词法层记录词表语法信息.利用这样简洁的搜索空间可以使语音识别的解码过程更加紧凑有效.在对比实验中,传统的基于前缀合并的状态树搜索算法的大词表嵌入式单词拼读系统作为基线系统.实验结果表明,与基线系统相比,所提出的双层图搜索算法在系统解码速度相对提高10%的情况下,系统的动态内存占用仅为基线的8%.通过使用所提出的双层图搜索算法,大大提高了大词表嵌入式单词拼读系统的效率,使其更适用于大多数嵌入式平台.更多还原

【Abstract】 For the application of large vocabulary word lists to embedded speech recognition,an efficient two-layer graph search algorithm is proposed.In this approach,the two-layer graph search space is built by separating the acoustic layer and the lexical layer which contain the acoustic and lexical information respectively,and the compact framework makes the decoding process of speech recognition more efficient.In contrast experiments,the conventional prefix-merged state tree search algorithm is used in the embedded large vocabulary spell word system as baseline system.The experimental results show that compared with the baseline system,the memory footpring is only 8% when the decoding velocity of the proposed algorithm is 10% faster.By using the proposed algorithm,the spell word system is more appropriate and efficient for most embedded platforms.更多还原