摘要
矢量量化在语音识别中有着重要的作用。经典的K均值算法收敛速度快,但极易收敛于局部最佳点;其它的一系列改进算法在克服其局部收敛问题的同时,又显著增加了运算量。本文提出了用模拟退火算法实现语音识别中的矢量量化过程,能够较好地协调运算量和收敛质量之间的矛盾。文章讨论了具体算法,并给出了实验数据。结果表明该方法的综合性能优于现有算法,具有较高的实用价值。
Vector quantization plays an important role in speech recognition. Traditional K-means algorithm owns the advantage of fast convergence, but it is difficult to get the global optimal result. Some modified algorithms have been proposed to overcome this drawback, but they also increase the computation greatly. In this paper, a new algorithm which is based on annealing algorithm is proposed to compromise the contradiction. In the rest of the paper, the details of the algorithm and related experiments are given. The results demonstrate the algorithm is more effective than other methods.
关键词
矢量量化
语音识别
收敛
退火算法
Vector quantization, Speech recognition, Convergence, Annealing algorithm
