AGGLOMERATIVE CLUSTERING OF SOUND RECORD SPEECH SEGMENTS BASED ON BAYESIAN INFORMATION CRITERION

Abstract

This paper presents the detailed description of agglomerative clustering system implementation for speech segments based on Bayesian information criterion. Numerical experiment results with different acoustic features, as well as the full and diagonal covariance matrices application are given. The error rate DER equal to 6.4% for audio records of radio «Svoboda» was achieved by means of designed system.

Keywords

• speech segments clustering
• Bayesian variation analysis
• speech technologies