摘要
B—W算法是基于隐含马尔可夫模型连续语音建模所采用的技术.探讨了基于B—W算法的连续语音建模的关键技术———计算溢出及训练样本差异问题,提出了解决方法:对计算溢出可采用对数变换的方法扩展计算机所能表示的数值范围给以解决;差异样本训练可在训练中消除将定样句子单元信息予以解决.
Baum-Welch algorithm is a technology based on HMM. This paper discusses the primary technology that trains continuous speech models based on Baum-Welch algorithm. Calculation overflow and different sample training, and provides complete and flexible resolution. Logarithm transform that expands data range in computer is applied to solve the problem of calculation overflow, and different sample training is overcome to eliminate sentence unit information.
出处
《大庆石油学院学报》
CAS
北大核心
2005年第4期121-123,共3页
Journal of Daqing Petroleum Institute
关键词
语音识别
B_W算法
计算溢出
模型参数
speech recognition
B-W algorithm
calculation overflow
model parameter