@techreport{oai:ipsj.ixsq.nii.ac.jp:00057151, author = {板谷, 洋平 and 全, 炳河 and 南角吉彦 and 宮島, 千代美 and 徳田, 恵一 and 北村, 正 and Y., Itaya and H., Zen and Y., Nankaku and C., Miyajima and K., Tokuda and T., Kitamura}, issue = {124(2003-SLP-049)}, month = {Dec}, note = {GMM(Gaussian Mixture Model)やHMM(Hidden Markov Moldel)のパラメータ推定には,広くEM(Expectation Maximization)アルゴリズムが用いられる.しかし,EMアルゴリズムは,推定結果が初期値設定に依存してしまうという,局所最適性の問題を有する.この問題に対処するためにDAEM(Deterministic Annealing Expectation Maximization)アルゴリズムが提案された.本稿では,このDAEMアルゴリズムをGMM,及び音素境界情報が得られない場合のHMMのパラメータ推定(フラットスタート)にそれぞれ適用し,GMMを用いた話者認識,及びHMMを用いた連続音声認識における有効性に関する検討を行う., This paper investigates the effectiveness of a DAEM (Deterministic Annealing Expectation Maximization) algorithm for speaker and speech recognition. The EM (Expectation Maximization) algorithm is widely used for paramete estimation of statistical models. However, the EM algorithm has the problem of initialization dependence. To relax this problem, the DAEM algorithm has been proposed. In this paper, we apply the DAEM algorithm to estimate acoustic models for speaker recognition and continuous speech recognition.}, title = {DAEMアルゴリズムの話者・音声認識における有効性の検討}, year = {2003} }