Investigating Bidirectional Recurrent Neural Network Language Models for Speech Recognition

Explore this paper's citation graph

Summary

The issues of training bidirectional recurrent neural network language models (bi-RNNLMs) for speech recognition are examined and a probability smoothing technique is proposed, that addresses the very sharp posteriors that are often observed in these models.

Type
article
Published
2017-08-20
Cited by
44
References
23
Access
Open access

Keywords

Computer science, Recurrent neural network, Speech recognition, Time delay neural network, Language model

References

Cited by

Related papers