论文部分内容阅读
Ｂａｕｍ－Ｗｅｌｃｈ算法在实际应用中存在算法下溢问题，参考文献ｌ～文献３中都介绍了尺度变换（Ｓｃａｌｉｎｇ）算法以解决该问题．然而这３篇文献的算法公式中存在不同程度的错误．实验结果显示原算法会导致模型训练不收敛或收敛性不好而导致识别率不高．本文分析了这些文献算法公式中存在的问题并推导给出正确公式．使用了修正后算法的语音识别系统有良好的收敛性而且可以获得较高的识别率．
Baum-Welch algorithm has an algorithm underflow problem in practical application. In References 1 ~ 3, a scaling algorithm is introduced to solve this problem. However, there are different degrees of error in the algorithms of these three documents. The experimental results show that the original algorithm leads to poor convergence rate or poor convergence of the model training, resulting in low recognition rate. This paper analyzes the problems existing in the algorithms of these documents and derives the correct formulas. The speech recognition system using the modified algorithm has good convergence and can obtain higher recognition rate.