TY  - INPR
ID  - SisLab3170
UR  - https://eprints.uet.vnu.edu.vn/eprints/id/eprint/3170/
A1  - Ho, Thi Nga
A1  - Can, Duy Cat
A1  - Chng, Eng Siong
Y1  - 2018/11//
N2  - This work investigates the effectiveness of using the word based and sub-word based embedding representations as input for a deep bidirectional Long Short-Term Memory Network for Sentence Unit Detection in Automatic Speech Recognition transcription. Our experimental results show that using sub-word based embedding can significantly improve the SUD performance when a limited text is used to train both the word embedding and the SUD model. The SUD model using the sub-word based embedding gains up to 2.07% absolute improvement in F1-score as compared to the best model trained with the word-based embedding. When tested on a domain-mismatch condition, the SUD model with sub-word based embedding trained from the in-domain data gives an approximate 2% and 1% improvement over the best model using out-of-domain embedding with reference and ASR transcription with 29.5% Word Error Rate respectively.
TI  - An Investigation of Word Embeddings with Deep Bidirectional LSTM for Sentence Unit Detection in Automatic Speech Transcription
M2  - Bandung, Indonesia
AV  - restricted
T2  - International Conference on Asian Language Processing (IALP 2018)
ER  -