Initial speech recognition results using the multinet architecture

Multinet is a connectionist architecture designed for certain difficult multi-class pattern classification tasks. These are characterised by very large input feature spaces, rendering a monolithic classifier impractical. The architecture consists of a layer with at least one primary ‘detector’ for each class, followed by a combining net which estimates the posterior probabilities for all classes. Typically primary detectors only input a subset of the input features. Thus the architecture decomposes classification in two ways: by class and by factoring of the input space dimensions. Multinet incorporates the ideas of Modular Neural Networks and Ensembles. In this paper we investigate the use of Multinet on standard speech recognition tasks and present results for phoneme recognition on TIMIT and word recognition on RM. We show Multinet’s performance is comparable with standard HMM and hybrid HMM-NN systems that we run on the same tasks. The value and potential of the Multinet approach is shown by detailing successive improvements to the Multinet system which are easily obtained because of the modularity of the architecture.