Parallel Neural Networks for Speech Recognition

This paper presents the PNNC (Parallel Neural Networks by Conndence) and PNNS (Parallel Neural Networks by Success/Failure), which generate and integrate parallel neural networks to achieve high performance on the test problem of letter recognition from string of phonemes. Our approach provides a way to create subproblems for a complex problem by partitioning the data, thus each neural network adapts to each subproblem more eeciently. Each neural network is iteratively trained on the training data which the previous neural networks could not guarantee or produce proper results. Each network works by ltering out unsatisfactory instances to pass to the next sub network to handle. This approach provides a way, by exploring diier-ent search spaces, to handle the local minima problem without complex computations via the use of neural networks working in parallel. Experimental results show that our approach achieves improvement over the general multi-layered neural network on the speech recognition problem which converts strings of phonemes to strings of letters taken from the 934 most commonly used English words. Particularly, the fact that error on vowel recognition is reduced signiicantly by the subsequent neural network indicates the usefulness and effectiveness of our approach.