摘要
提出一种改进的矢量量化VQ(Vector Quantization)算法,优化了现有的声纹识别技术,并已经得到了应用。对基于LBG(Linde Buzo Gray)算法中现有的倒谱参数MFCC(Mel-Frequency Ceptral Coefficients)在声纹识别中的提取以及声纹模型建立的相关改进,使得矢量量化局部最优的特点在新的声纹模型建立时得到更好的应用。在保证识别率的情况下减少了匹配次数,降低了运行时间。经过测试,语音的平均识别率达到了92%以上,实时识别率达到90%以上。实验结果表明,相对于LBG算法原型,改进的算法的识别精度和速度都有所提高,是一种有效的声纹自动识别的实现方法。
An improved vector quantization has been brought forward to optimize the voice recognition in the current time, which has already been applied. This article is written about the modification of character refinement based on Mel-Frequency Ceptral Coefficients and voice model training by LBG, which takes full advantage of the characteristic of the LBG, local optimization, to establish voice model in a new way. In the guarantee of the quality of the recognition, the operation time is decreased with the time of comparison decreased. The average correct recognition by using this method is above 92 %, 90 % for real time testing. Experiment is also given to prove that the method is an effect way to recognize voice with high precise at a high speed.
出处
《上海第二工业大学学报》
2007年第4期317-322,共6页
Journal of Shanghai Polytechnic University
