GMD-SDDBHMM语音识别模型和分类训练方法

GMD SDDBHMM Models for Speech Recognition and A Classification Based Training Method

本文将混合高斯分布应用于一种非齐次隐含马尔可夫模型——简化的基于段长分布的隐含马尔可夫模型。新模型使语音识别率得到了改善。由于通常的模型训练方法训练时间太长，本文提出了一种分类训练方法，在不降低最终模型性能的前提下，使训练可以分布式完成。

Gaussian Mixture Distribution(GMD) is introduced into one kind of inhomogeneous Hidden Markov Model(HMM)—Simplified Duration Distribution Based HMM(SDDBHMM) in this paper. New model improves the speech recognition rate. Because the training time is too long if using the common model training method, a classification based training method is proposed to carry out training task distributedly under the condition of obtaining the same model performance.