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 2015-07-16 16:20
Sequence Discriminative Training for Low-Rank Deep Neural Networks
Yuuki Tachioka (Mitsubishi Electric), Shinji Watanabe, Jonathan Le Roux, John Hershey (MERL) SP2015-39
Abstract (in Japanese) (See Japanese page) 
(in English) Deep neural network (DNN) acoustic models outperform conventional Gaussian mixture model (GMM) but the number of parameters tends to be larger, leading to higher computational cost than those of GMM. To reduce the size of a DNN model, low-rank approximations of weight matrices, by using singular value decomposition (SVD), have previously been applied. Previous studies only focused on clean speech, whereas under noisy condition model reduction could be difficult. Thus we investigate the effectiveness of this SVD method on noisy reverberated speech recognition task. Furthermore, we combine the low-rank approximation with sequence discriminative training, which further improved the performance of the DNN, which was constructed using a frame-by-frame discriminative criterion. We also investigated the effect of the order of application of the low-rank and sequence discriminative training. Our experiments show that low rank approximation is effective for noisy speech recognition and the most effective combination of discriminative training with model reduction is to apply the low rank approximation to the base model first and then to perform discriminative training on the low-rank model. This low-rank discriminatively trained model outperformed the full discriminatively trained model.
Keyword (in Japanese) (See Japanese page) 
(in English) Automatic speech recognition / Deep neural networks / Singular value decomposition / Model size reduction / Discriminative training / / /  
Reference Info. IEICE Tech. Rep., vol. 115, no. 146, SP2015-39, pp. 19-24, July 2015.
Paper # SP2015-39 
Date of Issue 2015-07-09 (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 SP2015-39

Conference Information
Committee SP IPSJ-SLP  
Conference Date 2015-07-16 - 2015-07-17 
Place (in Japanese) (See Japanese page) 
Place (in English) Katakura Suwako Hotel 
Topics (in Japanese) (See Japanese page) 
Topics (in English) Speech recognition and understanding, dialog system, etc. 
Paper Information
Registration To SP 
Conference Code 2015-07-SP-SLP 
Language Japanese 
Title (in Japanese) (See Japanese page) 
Sub Title (in Japanese) (See Japanese page) 
Title (in English) Sequence Discriminative Training for Low-Rank Deep Neural Networks 
Sub Title (in English)  
Keyword(1) Automatic speech recognition  
Keyword(2) Deep neural networks  
Keyword(3) Singular value decomposition  
Keyword(4) Model size reduction  
Keyword(5) Discriminative training  
Keyword(6)  
Keyword(7)  
Keyword(8)  
1st Author's Name Yuuki Tachioka  
1st Author's Affiliation Mitsubishi Electric Corporation (Mitsubishi Electric)
2nd Author's Name Shinji Watanabe  
2nd Author's Affiliation Mitsubishi Electric Research Laboratories (MERL)
3rd Author's Name Jonathan Le Roux  
3rd Author's Affiliation Mitsubishi Electric Research Laboratories (MERL)
4th Author's Name John Hershey  
4th Author's Affiliation Mitsubishi Electric Research Laboratories (MERL)
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 ()
Speaker Author-1 
Date Time 2015-07-16 16:20:00 
Presentation Time 30 minutes 
Registration for SP 
Paper # SP2015-39 
Volume (vol) vol.115 
Number (no) no.146 
Page pp.19-24 
#Pages
Date of Issue 2015-07-09 (SP) 


[Return to Top Page]

[Return to IEICE Web Page]


The Institute of Electronics, Information and Communication Engineers (IEICE), Japan