Speech Recognition Using Hidden Control Neural Network Architecture

Multi-layered neural networks have been recently proposed for non-linear signal processing, in particular, non-linear prediction and system modeling [1]. Although proven successful for modeling time invariant nonlinear systems [1]-[3], their inability to capture temporal variability in the system's parameters has so far been the major obstacle in applying neural networks to complicated nonstationary signals, such as speech. In this paper we present a network architecture, called 'Hidden Control Network' (HCNN) which attempts to cope with time varying systems by adding a control signal to the network.