This paper proposes a novel attention-based recurrent neural network (RNN) to build an end-to-end automatic language identification (LID) system. Inspired by the success of attention mechanism on a range of sequence-to-sequence tasks, this work introduces the attention mechanism with long short term memory (LSTM) encoder to the sequence-to-tag LID task. This unified architecture extends the end-to-end training method to LID system and dramatically boosts the system performance. Firstly, a language category embedding module is used to provide attentional vector which guides the derivation of the utterance level representation. Secondly, two attention approaches are explored: a soft attention which attends all source frames and a hard one that focuses on a subset of the sequential input. Thirdly, a hybrid test method which traverses all gold labels is adopted in the inference phase. Experimental results show that 8.2% relative equal error rate (EER) reduction is obtained compared with the LSTM-based frame level system by the soft approach and 34.33% performance improvement is observed compared to the conventional i-Vector system.
Cite as: Geng, W., Wang, W., Zhao, Y., Cai, X., Xu, B. (2016) End-to-End Language Identification Using Attention-Based Recurrent Neural Networks. Proc. Interspeech 2016, 2944-2948, doi: 10.21437/Interspeech.2016-686
@inproceedings{geng16_interspeech,
author={Wang Geng and Wenfu Wang and Yuanyuan Zhao and Xinyuan Cai and Bo Xu},
title={{End-to-End Language Identification Using Attention-Based Recurrent Neural Networks}},
year=2016,
booktitle={Proc. Interspeech 2016},
pages={2944--2948},
doi={10.21437/Interspeech.2016-686}
}