嵌入式语音识别系统特征参数提取研究

摘要

在噪声环境下能准确有效地提取语音信息是语音识别的重点难点,将其应用于嵌入式系统中,有一定的研究意义.通过比较分析传统的语音特征参数提取的方法:线性预测倒谱系数,Mel频率倒谱系数,提出了一种新的方法,采用Mel频率倒谱系数与一阶差分Mel频率倒谱系数(MFCC+ A MFCC)相结合的方法提取语音特征参数,结合双门限检测法进行端点检测和HMM模型进行模型匹配,并进行了以ARMSX2410为核心硬件与软件的系统设计.该方法较传统方法提高了系统的鲁棒性、识别的准确率和系统效率,适用于噪声环境下的语音识别.%The key points and difficulties of speech recognition is the technology of extracting the voice information accurately and efficiently in noisy environment. Applying this technology in embedded systems has some research significance. Through the comparative analysis of the traditional phonetics characteristic parameters extraction methods which are linear forecast cepstrum coefficients and Mel frequency cepstrum coefficients, proposed a new method in which through combining the method of Mel frequency cepstrum coefficients with first-order differential Mel frequency cepstrum coefficients ( MFCC + Δ MFCC ) to extract phonetic features parameters firstly, then take advantage of double threshold method to detect endpoint and use HMM model to do model matching,and the system design is carried on the ARMS3C2410 as the core of hardware and software. The experiments show that the system robustness, identification accuracy and efficiency of the system in this method got improved compared with the traditional method,and this method is suitable for noise environment of speech recognition.