Maximum a Posteriori Estimation for Multivariate Gaussian Mixture Observations of Markov Chains Jean-Luc Gauvain and Chin-Hui Lee
• 最大後驗機率估計(maximum a posteriori estimation) • 最大似然線性迴歸(maximum likelihood linear regression) 2
Outline • Introduction • Choices of Prior Densities • MAP Estimates for Gaussian Mixture • MAP Estimates for HMM • Prior Density Estimation • Experimental Results • Conclusion 5
Introduction • The choice of the prior distribution family. • The specification of the parameters for the prior densities. • The evaluation of the MAP. •
Choices of Prior Densities •
Choices of Prior Densities •
Choices of Prior Densities •
Choices of Prior Densities •
MAP Estimates for Gaussian Mixture •
MAP Estimates for Gaussian Mixture •
MAP Estimates for Gaussian Mixture •
Forward-Backward MAP Estimate
Forward-Backward MAP Estimate
MAP Estimates for HMM •
Prior Density Estimation •
Prior Density Estimation •
Experimental Results
Conclusion • The forward-backward MAP estimation and the segmental MAP estimation, were formulated(制定). • The proposed Bayesian estimation approach provides a framework to solve various HMM estimation problems posed by sparse training data. • The same framework can also be adopted for the smoothing and adaptation of discrete and tied-mixture hidden Markov models and N-gram stochastic language models.