语音识别中的神经网络DTW结构

本文介绍语音识别中一种基于动态规划技术的时间规正算法DTW的神经网络实现方法.DTW是语音识别中最为有效的方法之一,它具有较强的鲁棒性且为语音识别系统提供了可能的最高识别率.但由于其计算量太大,除非用专门的硬件,DTW算法在实现时受到了限制.在本文中,所有的计算是由两个循环神经子网和一记忆层来完成的,该方法展示了算法的硬接线结构,(hard-wiring)的优越性,这为DTW的硬件实现提供了一种新的实施策略.

A Neural Net DTW Architecture in Speech Recognition

This paper reports an implementation of dynamic programming based time-normalized algorithm, called dynamic time warping (DTW), on neural networks. DTW is one of the most successful algorithms for spoken word recognition. It is very robust and usually provides the highest recognition rate possible but it takes a lot of computer time unless it is implemented by special hardware. In this implementation, the computation is governed by two recurrent subnets and one memory layer, demonstrating a hard-wiring mechanism which benefits from existing approaches.