INITIALIZATION PROBLEMS OF SPEAKERS’ SEGMENTATION SYSTEMS BASED ON VARIATION BAYESIAN ANALYSIS

O. Y. Kydashev, T. . Pekhovsky


Read the full article 

Abstract

 

This paper presents a model description used for the speakers’ segmentation task. The iterative formulas approximating  distribution of a posteriori speakers’ model parameters based on the made assumptions are given and the new original method for initializing model parameters is offered. The scheme of a speakers’ segmentation system based on developed approach is presented. The application of developed system gives relative reduction in error rate up to 26% both at Russian and English speech data bases.


Keywords: Bayesian analysis, variation method, speakers’ segmentation.
Copyright 2001-2017 ©
Scientific and Technical Journal
of Information Technologies, Mechanics and Optics.
All rights reserved.

Яндекс.Метрика