Parallel distributed binary mapping models for speech recognition

In this paper, we present a parallel distributed binary mapping models (PDBMM) approach to speech recognition, which is characterized by the following important properties. (1) Based on parallel processing, a PDBMM is constructed for speaker-independent large-vocabulary speech recognition, which optimally combines the techniques of neural networks and HMM. (2) A PDBMM is made up of a certain number of time-sequence binary classifiers (TSBC) which are designed to extract or fire low level acoustic events for basic speech units of modeling. PDBMM deals with speech recognition parallelly and allows to form the arbitary nonlinear decision surfaces more efficiently and rapidly. (3) The Time-Sequence arrangement will enable the TSBC to extract low level acoustic events and the temporal relationships between them. The evaluation results is very inspiring and shows the great potentiality of PDBMM in speech recognition