基于硬件加速模块的嵌入式语音识别系统解决方案  

An embedded speech recognition system solution based on acceleration modules

作　　者：智强[1,2] 李鹏[1,2] 董明[3,2] 梁维谦[3,2] 刘润生[3,2] 

机构地区：[1]清华大学微电子所,北京100084 [2]清华大学信息科学与技术国家实验室,北京100084 [3]清华大学电子工程系,北京100084

出　　处：《电子技术应用》2008年第8期31-33,37,共4页Application of Electronic Technique

摘　　要：在基于 CHMM 模型的语音识别原理的基础上,设计了一个以 MCU 和自行设计的语音识别加速模块(ASIC 模块)为核心的低成本、高性能的嵌入式语音识别系统。该系统配合外围电路,能够独立完成语音识别工作,并且有大幅度的性能提升,从而使嵌入式语音识别更加方便简洁。以 ARM7作为系统的控制内核,语音识别加速模块负责完成隐含马尔可夫模型识别算法中运算量最大的 Ma-halanobis 距离运算部分。该系统具有低成本、高性能、高通用性、可裁剪性强等特点。This paper discusses an embedded speech recognition system Solution with low cost and high performance, which is based on the speech recognition algorithm using a CHMM （continuous hidden Markov model）.This system has acceleration modules designed by ASIC （Application Specific Intergrated Circuits） and a controlling core using ARM7. The acceleration modules finish the calculation of Mahalanobis distance which consume the most part of overall speech recognition algorithm. An FPGA test shows that the system can achieve real-time performance at relatively low clock frequencies, which significantly reduces the power consumption.

关 键 词：嵌入式语音识别 MCU+ASIC MAHALANOBIS距离 

分 类 号：TN912.34[电子电信—通信与信息系统] TN801[电子电信—信息与通信工程]