GMM parameter estimation by means of EM and genetic algorithms

  • Authors:

  • Sergey Zablotskiy;Teerat Pitakrat;Kseniya Zablotskaya;Wolfgang Minker

  • Affiliations:

  • Department of Information Technology, University of Ulm, Germany;Department of Information Technology, University of Ulm, Germany;Department of Information Technology, University of Ulm, Germany;Department of Information Technology, University of Ulm, Germany

  • Venue:
  • HCII'11 Proceedings of the 14th international conference on Human-computer interaction: design and development approaches - Volume Part I
  • Year:
  • 2011

Quantified Score

Hi-index 0.00

Visualization

Abstract

Most of the state-of-the-art speech recognition systems use Hidden Markov Models as an acoustic model, since there is a powerful Expectation-Maximization algorithm for its training. One of the important components of the continuous HMM we focus on is an emission probability which can be approximated by the weighted sum of Gaussians. Although, EM is a very fast iterative algorithm it can only guarantee a convergence to a local result. Therefore, the initialization process determines the final result. We suggested here two modifications of genetic algorithms for the initialization of EM. They are compared to the results of the EM with the same number of local multi-starts.