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 2026-03-03 11:00
[Poster Presentation] Context Dependency in Automatic Speech Recognition: Analyzing Mamba's Effectiveness
Joichiro Sato, Yosuke Higuchi, Tetsunori Kobayashi, Tetsuji Ogawa (Waseda Univ.) EA2025-126 SIP2025-146 SP2025-79
Abstract (in Japanese) (See Japanese page) 
(in English) This paper defines ``context dependency'' in automatic speech recognition (ASR) and investigates whether Mamba maintains recognition performance comparable to Conformer in situations with high dependency on past context.

Recently, Mamba, a sequence modeling architecture based on State Space Models (SSMs), has attracted attention as an alternative to Transformers. Unlike self-attention mechanisms, Mamba processes sequences through recursive state updates, achieving recognition accuracy comparable to Conformer in ASR tasks. However, despite the fundamental differences in how attention mechanisms and SSMs process past context, it remains unclear why Mamba achieves such strong modeling performance.

This study approaches this question by examining whether Mamba can maintain performance comparable to Conformer, even in situations where the dependency on past context is high.

To this end, we define a word-level metric called context dependency, which quantifies how much each word relies on past context by measuring the reduction in output entropy as the accessible context window increases. Using this metric, we classify words into high and low dependency groups and compare the recognition accuracy of Mamba and Conformer for each group.

Experimental results show that Mamba achieves recognition accuracy comparable to or better than Conformer, even for words with high context dependency, with no performance degradation observed. These results suggest that Mamba's recursive state compression mechanism has capabilities comparable to attention mechanisms in utilizing past context.
Keyword (in Japanese) (See Japanese page) 
(in English) automatic speech recognition / Mamba / state space model / Conformer / context dependency / / /  
Reference Info. IEICE Tech. Rep., vol. 125, no. 371, SP2025-79, pp. 318-323, March 2026.
Paper # SP2025-79 
Date of Issue 2026-02-23 (EA, SIP, 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 EA2025-126 SIP2025-146 SP2025-79

Conference Information
Committee SP EA SIP IPSJ-SLP  
Conference Date 2026-03-02 - 2026-03-04 
Place (in Japanese) (See Japanese page) 
Place (in English)  
Topics (in Japanese) (See Japanese page) 
Topics (in English)  
Paper Information
Registration To SP 
Conference Code 2026-03-SP-EA-SIP-SLP 
Language Japanese 
Title (in Japanese) (See Japanese page) 
Sub Title (in Japanese) (See Japanese page) 
Title (in English) Context Dependency in Automatic Speech Recognition: Analyzing Mamba's Effectiveness 
Sub Title (in English)  
Keyword(1) automatic speech recognition  
Keyword(2) Mamba  
Keyword(3) state space model  
Keyword(4) Conformer  
Keyword(5) context dependency  
Keyword(6)  
Keyword(7)  
Keyword(8)  
1st Author's Name Joichiro Sato  
1st Author's Affiliation Waseda University (Waseda Univ.)
2nd Author's Name Yosuke Higuchi  
2nd Author's Affiliation Waseda University (Waseda Univ.)
3rd Author's Name Tetsunori Kobayashi  
3rd Author's Affiliation Waseda University (Waseda Univ.)
4th Author's Name Tetsuji Ogawa  
4th Author's Affiliation Waseda University (Waseda Univ.)
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 2026-03-03 11:00:00 
Presentation Time 80 minutes 
Registration for SP 
Paper # EA2025-126, SIP2025-146, SP2025-79 
Volume (vol) vol.125 
Number (no) no.369(EA), no.370(SIP), no.371(SP) 
Page pp.318-323 
#Pages 6 
Date of Issue 2026-02-23 (EA, SIP, SP) 

[Return to Top Page]

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