A time-derivative neural net architecture-an alternative to the time-delay neural net architecture

Though the time-delay neural net architecture has been recently used in a number of speech recognition applications, it has the problem that it can not use longer temporal contexts because this increases the number of connection weights in the network. This is a serious bottleneck because the use of larger temporal contexts can improve the recognition performance. In this paper, a time-derivative neural net architecture is proposed. This architecture has the advantage that it can utilize information about longer temporal contexts without increasing the number of connection weights in the network. This architecture is studied here for speaker-independent isolated-word recognition and its performance is compared with that of the time-delay neural net architecture. It is shown that the time-derivative neural net architecture, in spite of using less number of connection weights, outperforms the time-delay neural net architecture for speech recognition