前馈序列记忆网络在语音识别中的应用综述

Review of the Applications of Feedforward Sequence Memory Network in Speech Recognition

为了改进传统的语音识别方法,使其达到低时延、高识别率的目的,以前馈序列记忆网络(feed-forward sequential memory network,FSMN)为研究主线,对语音识别的发展历程,FSMN的各种变体及其相应的应用状况进行综述,结果表明FSMN在语言建模任务上,使用wiki 9数据集能使Perplexity值降到90,并且训练周期比RNN快2倍.未来语音识别的发展趋向于更高识别性能和更多语种.

In order to improve the traditional speech recognition method to achieve the purpose of low latency and high recognition rate,a review was made of applications of(feed-forward sequential memory network,FSMN)in terms of the development of speech recognition,the various variants of FSMN with feedforward sequence memory network as the research main line.The results showed that FSMN can reduce the Perplexity to 90 by use of the wiki9 data set on solving language modeling tasks,and the training cycle is 2 times faster than by use of RNN.Future developing trend is directed to a higher recognition performance in speech recognition and multilingual speech recognition.