Proceedings of the International Workshop on Applications of Neural Networks to Telecommunications

Temporal Difference Learning Applied to Continuous Speech Recognition

Laurens R. Leerink, Marwan A. Jabri

Systems Engineering & Design Automation LaboratoryDepartment of Electrical EngineeringThe University of SydneyNSW 2006, Australia [e-mail], [e-mail]

Abstract

This paper evaluates the performance of two reinforcement learning algorithms on four, increasingly complex tasks of continuous speech recognition. Simulations reveal that when errors are examined only at word boundaries, both algorithms provide performance comparable with that of a fully supervised system.

1  Introduction

As a first step towards a continuous speech word recognition system trained only by reward and punishment, this paper examines the performance ...