Fourth European Conference on Speech Communication and Technology

Madrid, Spain
September 18-21, 1995

Multi-Variate Mixture Probability Density Modelling Of VQ Codebook Using Gradient Descent Algorithm

S. Dobrisek, R. Mihelic, N. Pavesic

Faculty of Electrical Engineering & Computer Science University of Ljubljana, Ljubljana, Slovenia

A vector quantisation codebook can be modelled as a set of probability density functions. The problem of estimating the parameters determining mixture probability density models can be solved using a log-likelihood based re-estimation procedure. On the other hand, this problem can be also viewed as a conventional optimisation problem. Consequently, gradient descent techniques may be used to obtain values of the model parameters. The main advantage of these techniques over the re-estimation procedure is higher robustness due to an initial estimation of the model parameters. In the paper, we describe a descent algorithm along with a criterion function, we propose. We obtained some promising results by applying this algorithm to one and two-variate pseudo Gaussian mixture probability density functions and further to signal vectors of a continuous speech database.

Full Paper

Bibliographic reference.  Dobrisek, S. / Mihelic, R. / Pavesic, N. (1995): "Multi-variate mixture probability density modelling of VQ codebook using gradient descent algorithm", In EUROSPEECH-1995, 1431-1434.