7th International Conference on Spoken Language Processing

September 16-20, 2002
Denver, Colorado, USA

Methods to Improve Gaussian Mixture Model Based Language Identification System

Eddie Wong, Sridha Sridharan

Queensland University of Technology, Australia

This paper investigates the use of Vocal Tract Length Normalisation (VTLN) and Output Score Fusion techniques to improve the performance of a Gaussian Mixture Model (GMM) based Language Identi- fication (LID) system. The Universal Background Model (UBM) technique, which has been successfully employed in Speaker Verification, is incorporated into the GMM LID system to reduce the time requirement for both training and testing. The paper also presents a fast approach for selecting the normalisation factor for VTLN during the testing stage of LID which is based on the UBM technique. The output scores generated by the GMM system have been fused with a phonetic based LID system to improve the overall scores. Experimental results show that a reduction in the relative error rate by over 50% is possible for the 45-second test case in the NIST 1994 Evaluation data.

Full Paper

Bibliographic reference.  Wong, Eddie / Sridharan, Sridha (2002): "Methods to improve Gaussian mixture model based language identification system", In ICSLP-2002, 93-96.