Structure-Optimized Speech Recognition with Deterministic Annealing

点击翻译按钮获取中文摘要

Research in statistical speech recognition will be conducted along two central lines. One research direction is concerned with the development of approaches to determine the optimal structure and complexity of the recognizer given limited training data, where optimality is with respect to expected performance outside the training set. The underlying framework is a structural complexity parameter space where points correspond to types of recognition systems, and optimal trajectories represent judicious trading of increase in system complexity for improved performance on the training set. A number of orthogonal means for performing such tradeoffs will be developed. The optimal point (system) on the trajectory will be determined by development of appropriate cross-validation techniques that estimate performance outside the training set. The complementary research direction involves the optimization technique used in the system design, which determines its ultimate performance. The deterministic annealing approach will be extended and developed as a powerful optimization tool for speech recognition and, in particular, for optimization operations within the above structural parameter space so as to directly minimize the expected rate of misclassification outside the training set. The project is expected to result in powerful speech recognizer design techniques and thereby potentially extend the accessibility of computers to the public, and enhance the feasibility of numerous applications, ranging from voice-activated devices to automated telephony services.