An Accuracy Optimization of a Dialog ASR System Utilizing Evolutional Strategies

This article deals with an accuracy optimization of a dialog ASR system based on ATK tool, whereas the optimization is performed by the means of evolutional strategies. First an introductory overview of the basic capabilities of ATK system is given followed by the founding principles and variations of the evolutional strategies applied to the nonlinear optimization problems. Then the process of ASR optimization is presented and supported by series of experiments. By applying stochastic optimization methods to the mutual adjustment of Viterbi decoder, speech detection, and grammar related settings, we were able to achieve improved overall performance (WER 18.72%) compared to the manual settings chosen by tests and gained "experience " (generally WER above 23%).