Author/Authors :
ÇAVDARLI, Murat Uludağ Üniversitesi - Mühendislik Mimarlık Fakültesi, TÜRKİYE , ESKİDERE, Ömer Uludağ Üniversitesi - Teknik Bilimler Meslek Yüksekokulu - Mekatronik Programı, TÜRKİYE , ERTAŞ, Figen Uludağ Üniversitesi - Mühendislik Mimarlık Fakültesi, TÜRKİYE
Title Of Article :
A comparison of training algorithms in speaker identification
Abstract :
In this study, training algorithms are compared in Gaussian mixture model (GMM) based a speaker identification system. The Expectation maximization (EM) algorithm has widely been used to estimation of GMM parameters. In this article, the k-means and LBG are applied to GMM in order to estimate the vector quantization training parameters. The EM, the k-means and LBG training algorithms are tested with TIMIT and NTMIT databases and are compared speaker identification performance. Furthermore, the EM and k-means algorithms which sensitive against model initialization values are found optimum model initialization values for databases.
NaturalLanguageKeyword :
BM , LBG , k , means , Speaker identification , Gaussian mixture model
JournalTitle :
Erciyes University Journal Of The Institute Of Science and Technology
