73. MATRIX UPDATES FOR PERCEPTRON TRAINING OF CONTINUOUS DENSITY HIDDEN MARKOV MODELS

Department: Computer Science & Engineering
Faculty Advisor(s): Lawrence Saul

Primary Student
Name: Chih-Chieh Cheng
Email: chc028@ucsd.edu
Phone: 858-232-8056
Grad Year: 2011

Abstract
We describe a perceptron-based training in continuous density hidden Markov models (CD-HMMs) for automatic speech recognition (ASR), and study the problem of parameterizations and estimation for the framework. The conventional parameterization of Gaussian mixture HMMs is constituted by covariance matrices and mean vectors. However, these two parameters distinct in both dimensions and scales, and thus make the online training intractable. We present an aggregated reparameterization and factorizations of the parameters, and evaluate on the TIMIT speech corpus. As in perceptron algorithm, we define our criterion function as the errors between target and Viterbi decoded sequence, and with the reparametrization, the optimization is over a parameter space of positive semidefinite matrices. The experiments show that the reparametrizations have led to a significant reduction in phone error rates, and the perceptron training has greatly improved the convergence rate for dscriminative HMMs. The experiments show that the reparametrizations have led to a significant reduction in phone error rates, and the perceptron training has greatly improved the convergence rate for dscriminative HMMs. The experiments show that the reparametrizations have led to a significant reduction in phone error rates, and the perceptron training has greatly improved the convergence rate for dscriminative HMMs. The experiments show that the reparametrizations have led to a significant reduction in phone error rates, and the perceptron training has greatly improved the convergence rate for dscriminative HMMs.

« Back to Posters or Search Results