有噪语音的线性预测HMM参数估计方法

Parameter Estimation of Linear Predictive HMM for Noisy Speech

本文讨论了混有加性高斯白噪声的语音的线性预测隐Markov模型(HMM)参数估计问题.由Baum等人提出来的重估公式没有考虑噪声的影响,因此,对有噪语音的模型参数估计会产生较大的偏差.本文对该公式提出了一种修正方案,用这种方法估计出的模型参数有效地抑制了噪声的影响,较准确地反映了语音真实时变特征.实验证实了这一结论.

The parameter estimation of a Linear Predictive Hidden Markov Model (LPHMM) for additive Gaussian white noisy speech are described. As no reference is made to the influence of noise in the reestimation formula given by Baum and others, fairly great deviation results when it is used in model parameter estimation of noisy speech. A modified reestimation formula is proposed, by which the model parameter estimated suppress the influence of noise effectively and give true time-variable features of speech. This has been proved by experiment.