嵌入式语音识别Mahalanobis距离计算模块  

Mahalanobis distance calculation module for speech recognition

作　　者：李鹏[1] 智强[1] 董明[2] 梁维谦[2] 刘润生[2] 

机构地区：[1]清华大学微电子学研究所,北京100084 [2]清华大学电子工程系,北京100084

出　　处：《清华大学学报（自然科学版）》2008年第7期1202-1204,1208,共4页Journal of Tsinghua University(Science and Technology)

基　　金：北京市科技计划项目(Y0105008040111)

摘　　要：为了达到嵌入式语音识别系统低成本、低功耗的目标,提出一种算法硬件映射方法。将基于连续隐含Markov模型语音识别算法中占系统总运算量的50%以上的Mahalanobis距离计算,映射为硬件实现的模块。通过该方法,系统在较低时钟频率下即可完成嵌入式语音识别中实时处理的要求,从而大大降低系统功耗。实验结果表明,该模块在0.18μm和舰工艺库下实现,仅需1.2mm2,包含64kb静态随机存储器。应用该模块可以大大提高嵌入式语音识别系统的性能,达到降低成本,降低功耗的目标。This paper discusses a low cost, low power-consumption hardware implementation of an embedded speech recognition system. The mapped algorithm, called the Mahalanobis distance, handles about 50% of its computational load in the overall speech recognition algorithm using a continuous hidden Markov model （CHMM）. An analyses and tests show that the system can achieve real-time performance at relatively low clock frequencies, which significantly reduces the power consumption. The die area is just 1.2 mm^2based on the HJTC （He Jian Technology Corporation） 0.18 μm technology, including 64 kb SRAM （static random access memory）. As a result, this module significantly improves the performance of the whole speech recognition system to a low cost, low power system.

关 键 词：语音识别 MAHALANOBIS距离 IP(intellectual property)核 

分 类 号：TN912.34[电子电信—通信与信息系统]