Modular construction of time-delay neural networks for speech recognition

摘要: Several strategies are described that overcome limitations of basic network models as steps towards the design large connectionist speech recognition systems. The two major areas concern problem time and scaling. Speech signals continuously vary over encode transmit enormous amounts human knowledge. To decode these signals, neural networks must be able to use appropriate representations it possible extend nets almost arbitrary sizes complexity within finite resources. is addressed by development a Time-Delay Neural Network; scaling Modularity Incremental Design based on smaller subcomponent nets. It shown small trained perform limited tasks develop invariant, hidden abstractions can subsequently exploited train larger, more complex efficiently. Using techniques, phoneme increasing complexity...