高性能汉语数码语音识别芯片系统  

High performance mandarin digit recognition system on a DSP chip

作　　者：董明[1] 刘加[1] 刘润生[1] 

机构地区：[1]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）》2003年第9期1257-1260,共4页Journal of Tsinghua University(Science and Technology)

基　　金：国家自然科学基金资助项目(69975007);国家"八六三"高技术项目(863-306ZD13-04-6)

摘　　要：在嵌入平台上实现高性能的汉语数码语音识别(MDSR),对于电话通讯、工业控制等都具有极高的实用价值。该文描述了一个在16bit定点DSP芯片上实现的高性能汉语数码语音识别系统。识别模型采用连续隐Markov模型(CHMM),识别特征采用Mel频标倒谱系数(MFCC)。在模型的训练中引入MCE区分性训练进一步提高了系统的识别性能。识别过程采用单级识别框架,降低了芯片上系统部分的复杂性,同时保证了很高的识别性能与稳健性。实验证明该系统对11汉语数码发音可以达到98.3%的识别正确率,在58.5MIPS的16bit定点DSP上进行一次识别只需要35ms。Highperformance Mandarin digital speech recognition (MDSR) systems on embedded platforms are needed by many industries such as for telecommunications and automatic control. A highperformance MDSR system was implemented on a 16bit fixedpoint DSP. The system uses the Mel frequency cepstrum coefficient parameter as the main feature parameter and the speech recognition algorithm is based on the continuous density hidden Markov model. MCE training is used for the model training to further improve the recognition accuracy. The onchip speech recognition engine employs singlestage recognition architecture, which reduces the complexity of the onchip program and reduces the sensitivity to the speech parameters. Tests show that the MDSR system provides recognition accuracy rates as high as 98.3% and each numeral recognition requires only 35 milliseconds on a 58.5 MIPS DSP.

关 键 词：汉语数码语音识别芯片系统 DSP 连续隐Markov模型 识别性能 稳健性 

分 类 号：TN912.34[电子电信—通信与信息系统]