結合貝氏估計法的修正最大概似法在語音辨識的應用

Abstract

近代聲音辨識系統普遍建立在Gaussian mixture model 或Hidden Markov Model.最大概似法事最常被用來估計這些模型,而另一種估計法則是採用貝氏估計，在本研究中，我們想設計一個結合貝氏估計與最大概似估計的新方法來提升採用Gaussian mixture model的聲音辨識效果。.

Modern speech recognition systems are generally based on the Gaussian mixture model or Hidden Markov Model. The maximum likelihood method is a popular approach to derive the parameter estimation in these models. An alternative solution is to adopt the Bayes approach for the parameter estimation. In this study, we propose a new method which can combine the results of the maximum likelihood and Bayes approaches to provide more accurate estimators in the Gaussian mixture model for audio signal recognition.