解决Baum-Welch算法下溢问题的参数重估公式中存在的问题及其更正

详细信息

Some mistakes in the algorithm for solving underflow in baum-welch algorithm and their corrections

摘要: 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.

Abstract: Baum-Welch algorithm most likely results in underflow in practice. Some literatures such as 1-3 introduce "Scaling" algorithm for solving the problem. In applications, it is found that there are some mistakes in the formulae presented in these literatures. The practical calculation shows that the original algorithm often results in poor convergence or even divergence and result in high error rate in speech recognition. This paper analyses the mistakes in these literatures and bring forward the right formulae. The speech recognition system using the revised algorithm can converge well and has low error rate.