Gaussian Mixture Models (GMMs) are a common approach to modeling class-dependent densities in statistical pattern recognition systems. Since full covariances require a huge number of parameters and are costly to compute, it is typical to assume feature independence, and model the GMMs using diagonal covariance Gaussians. We present a new approach, called Mixtures of Inverse Covariances (MIC), that approximates full covariances while reducing the number of parameters and the cost of computation to a level similar to that of diagonal covariances. In MIC, inverse covariances are represented as a linear combination of a small number of prototype matrices that are shared across Gaussians. We present a maximum-likelihood estimation algorithm for the MIC parameters. Through experimental results on a variety of speech recognition tasks, we show that MIC is superior to diagonal GMM systems in both accuracy and speed.

Speaker Bio:

ANANTH SANKAR works in the Speech Research and Development group at Nuance Communications. Prior to that he has been with the Speech research groups at SRI International, and at AT&T Bell Laboratories. He has made several contributions in the areas of acoustic modeling, adaptation, and pronunciation modeling, both in research, and bringing these technologies to market. In addition to industry, Ananth has taught several times at Stanford and at UCLA and served on Ph.D. committees at Stanford and Rutgers University.

Ananth received the Bachelor of Technology at the Indian Institute of Technology, the MS at the State University of New York, Stony Brook, and the Ph.D. degree at Rutgers University, all in Electrical Engineering.