摘要
提出了一种新的基于LBG和DTW结合的模板训练算法,包括模板训练、初始模板设置、空子集处理三个部分,能够完整、有效地解决语音识别中模板训练的问题。该算法实现了语音信号特征矩阵的聚类及其质心的生成,使孤立词语音识别系统更好地适用于非特定人的情况,提高了系统对训练集外说话人语音的正确识别率。设计、实现了一个识别系统,模板训练中较快的收敛速度和系统较高的识别率验证了算法的优良性能。
Based on LBG and DTW algorithm,a novel template-training algorithm is proposed.This algorithm consists of three parts,i.e.template training,template initialization and empty subset processing.h solves the problem of templatetraining in speech recognition integrally and effectively.With this algorithm,the acoustic feature matrixes are clustered to form centroids,which makes the isolated speech recognition system more suitable for non-specific speaker condition.The recognition rate of out-of-training-set speaker is improved.The algorithm is applied to recognition system,the faster convergence speed of template training and higher recognition rate of system validate excellent performance.
出处
《计算机工程与应用》
CSCD
北大核心
2005年第26期85-88,共4页
Computer Engineering and Applications
关键词
语音识别
LBG算法
模板训练
空子集处理
speech recognition,LBG algorithm,template training,empty subset processing