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 |
6 |
Date of Issue |
2015-07-09 (SP) |
|