Paper Abstract and Keywords |
Presentation |
2022-03-02 13:25
[Poster Presentation]
Transformer-based Text Decoding using Electrocorticography Shuji Komeiji, Kai Shigemi (TUAT), Takumi Mitsuhashi, Yasushi Iimura, Hiroharu Suzuki, Hidenori Sugano (Juntendo Univ.), Koichi Shinoda (Tokyo Tech), Toshihisa Tanaka (TUAT) EA2021-87 SIP2021-114 SP2021-72 |
Abstract |
(in Japanese) |
(See Japanese page) |
(in English) |
Invasive brain-machine interfaces (BMIs) are a promising neurotechnology for achieving direct speech communication from a human brain but face many challenges. This paper measured the invasive electrocorticogram (ECoG) signals from seven participating epilepsy patients as they spoke a sentence consisting of multiple phrases. A Transformer encoder was incorporated into a “sequence-to-sequence” model (Transformer Seq2seq) to decode spoken sentences from the ECoG. A Transformer is a successful neural network model for natural language processing and automatic speech recognition. The decoding test revealed that the use of the Transformer model achieved a minimum phrase error rate of 16.4% for one best participant; moreover, the median (±standard deviation) of PER for the Transformer Seq2seq across seven participants was31.3% (±10.0%). This result showed that the Transformer Seq2seq effectively decoded from ECoG. |
Keyword |
(in Japanese) |
(See Japanese page) |
(in English) |
Electrocorticography / Brain--machine interface / Transformer encoder / Sequence to sequence / / / / |
Reference Info. |
IEICE Tech. Rep., vol. 121, no. 384, SIP2021-114, pp. 146-151, March 2022. |
Paper # |
SIP2021-114 |
Date of Issue |
2022-02-22 (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 |
EA2021-87 SIP2021-114 SP2021-72 |
|