Information: Join today and make your research activities more affordable! Technical workshop participation fees and annual registration fees are available at member rates.
Notice: [Important] Announcement of Changes to Registration Fee Payment and Manuscript Upload Procedures for IEICE Technical Meetings
IEICE Technical Committee Submission System
Conference Paper's Information		 Online Proceedings
[Sign in]
Tech. Rep. Archives
 Go Top Page Go Previous   [Japanese] / [English] 

Paper Abstract and Keywords
Presentation 2016-07-30 09:00
Discriminative Training Method of Recurrent Neural Network Language Models for Speech Recognition
Yuuki Tachioka (Mitsubishi Electric), Shinji Watanabe (MERL) SP2016-26
Abstract (in Japanese) (See Japanese page) 
(in English) A recurrent neural network language model (RNN-LM) can consider a longer word context than an n-gram language model, and its effectiveness has recently been shown for automatic speech recognition (ASR) tasks. However, the training criterion of RNN-LM is simply based on cross entropy (CE) between predicted and reference words. On the other hand, on top of models based on maximum likelihood or CE criteria, discriminative training of acoustic models and discriminative language models (DLM) have shown effectiveness because these criteria explicitly consider sequence discriminative criteria calculated from ASR hypotheses and references. This paper proposes a discriminative training method for RNN-LM by considering a log-likelihood ratio of the ASR hypotheses and references as a discriminative criterion in addition to CE. The proposed training criterion emphasizes the improperly recognized words relatively compared to the correct words where the weights for correct words are discounted in training. Experiments on a large vocabulary continuous speech recognition task show that our proposed method improves the CE based RNN-LM baseline and that combining the proposed discriminative RNN-LM and DLM has its additional effectiveness.
Keyword (in Japanese) (See Japanese page) 
(in English) Automatic speech recognition / recurrent neural network / language model / discriminative criterion / log-likelihood ratio / / /  
Reference Info. IEICE Tech. Rep., vol. 116, no. 165, SP2016-26, pp. 33-38, July 2016.
Paper # SP2016-26 
Date of Issue 2016-07-21 (SP) 
ISSN Print edition: ISSN 0913-5685    Online edition: ISSN 2432-6380
Copyright
and
reproduction		 All rights are reserved and no part of this publication may be reproduced or transmitted in any form or by any means, electronic or mechanical, including photocopy, recording, or any information storage and retrieval system, without permission in writing from the publisher. Notwithstanding, instructors are permitted to photocopy isolated articles for noncommercial classroom use without fee. (License No.: 10GA0019/12GB0052/13GB0056/17GB0034/18GB0034)
Download PDF SP2016-26

Conference Information
Committee SP IPSJ-SLP  
Conference Date 2016-07-28 - 2016-07-30 
Place (in Japanese) (See Japanese page) 
Place (in English) Takinoyu Hotel 
Topics (in Japanese) (See Japanese page) 
Topics (in English) Speech recognition and understanding, dialog system, etc. 
Paper Information
Registration To SP 
Conference Code 2016-07-SP-SLP 
Language Japanese 
Title (in Japanese) (See Japanese page) 
Sub Title (in Japanese) (See Japanese page) 
Title (in English) Discriminative Training Method of Recurrent Neural Network Language Models for Speech Recognition 
Sub Title (in English)  
Keyword(1) Automatic speech recognition  
Keyword(2) recurrent neural network  
Keyword(3) language model  
Keyword(4) discriminative criterion  
Keyword(5) log-likelihood ratio  
Keyword(6)  
Keyword(7)  
Keyword(8)  
1st Author's Name Yuuki Tachioka  
1st Author's Affiliation Mitsubishi Electric (Mitsubishi Electric)
2nd Author's Name Shinji Watanabe  
2nd Author's Affiliation Mitsubishi Electric Research Laboratories (MERL)
3rd Author's Name  
3rd Author's Affiliation ()
4th Author's Name  
4th Author's Affiliation ()
5th Author's Name  
5th Author's Affiliation ()
6th Author's Name  
6th Author's Affiliation ()
7th Author's Name  
7th Author's Affiliation ()
8th Author's Name  
8th Author's Affiliation ()
9th Author's Name  
9th Author's Affiliation ()
10th Author's Name  
10th Author's Affiliation ()
11th Author's Name  
11th Author's Affiliation ()
12th Author's Name  
12th Author's Affiliation ()
13th Author's Name  
13th Author's Affiliation ()
14th Author's Name  
14th Author's Affiliation ()
15th Author's Name  
15th Author's Affiliation ()
16th Author's Name  
16th Author's Affiliation ()
17th Author's Name  
17th Author's Affiliation ()
18th Author's Name  
18th Author's Affiliation ()
19th Author's Name  
19th Author's Affiliation ()
20th Author's Name  
20th Author's Affiliation ()
21st Author's Name  
21st Author's Affiliation ()
22nd Author's Name  
22nd Author's Affiliation ()
23rd Author's Name  
23rd Author's Affiliation ()
24th Author's Name  
24th Author's Affiliation ()
25th Author's Name  
25th Author's Affiliation ()
26th Author's Name / /
26th Author's Affiliation ()
()
27th Author's Name / /
27th Author's Affiliation ()
()
28th Author's Name / /
28th Author's Affiliation ()
()
29th Author's Name / /
29th Author's Affiliation ()
()
30th Author's Name / /
30th Author's Affiliation ()
()
31st Author's Name / /
31st Author's Affiliation ()
()
32nd Author's Name / /
32nd Author's Affiliation ()
()
33rd Author's Name / /
33rd Author's Affiliation ()
()
34th Author's Name / /
34th Author's Affiliation ()
()
35th Author's Name / /
35th Author's Affiliation ()
()
36th Author's Name / /
36th Author's Affiliation ()
()
Speaker Author-1 
Date Time 2016-07-30 09:00:00 
Presentation Time 30 minutes 
Registration for SP 
Paper # SP2016-26 
Volume (vol) vol.116 
Number (no) no.165 
Page pp.33-38 
#Pages 6 
Date of Issue 2016-07-21 (SP) 

[Return to Top Page]

[Return to IEICE Web Page]
The Institute of Electronics, Information and Communication Engineers (IEICE), Japan