Paper Abstract and Keywords |
Presentation |
2022-06-18 15:00
Unsupervised Training of Sequential Neural Beamformer Using Blindly-separated and Non-separated Signals Kohei Saijo, Tetsuji Ogawa (Waseda Univ.) SP2022-25 |
Abstract |
(in Japanese) |
(See Japanese page) |
(in English) |
We present an unsupervised training method of the sequential neural beamformer (Seq-NBF) using the separated signals from blind source separation (BSS) and observed mixtures as supervisory signals. Recently, separated signals of BSS have been used for training neural separators in an unsupervised manner. However, the performance is limited due to distortions in the supervision. In contrast, unmix-remix-consistent learning (URCL) utilizes distortion-free observed mixtures as the supervision, where we make remixed mixtures obtained by repeatedly separating and remixing two different mixtures closer to the original ones. Still, it is difficult to train separators from scratch with RCCL because it has a trivial solution of not separating signals. The present study provides a novel unsupervised learning algorithm for the Seq-NBF, where we first pre-train Seq-NBF with teacher-student learning with BSS and then fine-tune with URCL. By applying the two methods in stages, we make the most of their strengths and compensate for their weaknesses. We also expect that the configuration of Seq-NBF, which stacks two NBFs, will contribute to outperforming BSS in learning using the BSS outputs and boost the effectiveness of URCL-based fine-tuning. Experiments demonstrated that the proposed method significantly outperformed conventional BSS and achieved performance comparable to supervised learning (0.4 point difference in word error rate). |
Keyword |
(in Japanese) |
(See Japanese page) |
(in English) |
unsupervised speech separation / unmix-remix consistent learning / sequential neural beamformer / blind source separation / / / / |
Reference Info. |
IEICE Tech. Rep., vol. 122, no. 81, SP2022-25, pp. 110-115, June 2022. |
Paper # |
SP2022-25 |
Date of Issue |
2022-06-10 (SP) |
ISSN |
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 |
SP2022-25 |
Conference Information |
Committee |
SP IPSJ-MUS IPSJ-SLP |
Conference Date |
2022-06-17 - 2022-06-18 |
Place (in Japanese) |
(See Japanese page) |
Place (in English) |
Online |
Topics (in Japanese) |
(See Japanese page) |
Topics (in English) |
|
Paper Information |
Registration To |
SP |
Conference Code |
2022-06-SP-MUS-SLP |
Language |
Japanese |
Title (in Japanese) |
(See Japanese page) |
Sub Title (in Japanese) |
(See Japanese page) |
Title (in English) |
Unsupervised Training of Sequential Neural Beamformer Using Blindly-separated and Non-separated Signals |
Sub Title (in English) |
|
Keyword(1) |
unsupervised speech separation |
Keyword(2) |
unmix-remix consistent learning |
Keyword(3) |
sequential neural beamformer |
Keyword(4) |
blind source separation |
Keyword(5) |
|
Keyword(6) |
|
Keyword(7) |
|
Keyword(8) |
|
1st Author's Name |
Kohei Saijo |
1st Author's Affiliation |
Waseda University (Waseda Univ.) |
2nd Author's Name |
Tetsuji Ogawa |
2nd Author's Affiliation |
Waseda University (Waseda Univ.) |
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 |
() |
Speaker |
Author-1 |
Date Time |
2022-06-18 15:00:00 |
Presentation Time |
120 minutes |
Registration for |
SP |
Paper # |
SP2022-25 |
Volume (vol) |
vol.122 |
Number (no) |
no.81 |
Page |
pp.110-115 |
#Pages |
6 |
Date of Issue |
2022-06-10 (SP) |
|