Paper Abstract and Keywords |
Presentation |
2015-12-17 10:15
Silent Speech BCI
-- An investigation for practical problems -- Shun Hirose (KIT), Hiromi Yamaguchi (NEC), Takashi Ito, Toshimasa Yamazaki (KIT), Shinichi Fukuzumi (NEC), Takahiro Yamanoi (Hokkai Gakuen Univ.) |
Abstract |
(in Japanese) |
(See Japanese page) |
(in English) |
We have developed single-trial-EEG-based silent speech BCI (SSBCI) using speech signals. Our algorithm consisted of (1) extraction of independent components (ICs) whose dipole solutions were located at the Broca’s area by applying ICA and equivalent current dipole source localization (ECDL) to EEGs during actual or silent word speeches, (2) description of the relationship between spectrograms of speech signals and the EEGs during the actual speeches by Kalman filters, (3) estimation of SS spectrograms by inputting the EEGs during the silent speeches to the Kalman filters, and (4) recognition of silent words by comparing log-likelihood values of Hidden Markov Models (HMMs) that the estimated spectrograms were inputted to. The accuracy of “haru”- and “natsu”-HMMs was 39 % and 83 %, respectively. Unknown that which silent season EEGs were recorded during, the accuracy was 25 % and 47 %, respectively. Using the speech signal database, the one-hiragana-character-HMM yielded the accuracy of 0 % and 60 % for /ha/ and /ru/, respectively, and that of 88 % and 13 % for /na/ and /tsu/, respectively. Moreover, for the HMM with only gamma bands of the BA-ICs, the accuracy was drastically improved. |
Keyword |
(in Japanese) |
(See Japanese page) |
(in English) |
BCI / Silent Speech / ICA / ECDL / Kalman filter / HMM / / |
Reference Info. |
IEICE Tech. Rep. |
Paper # |
|
Date of Issue |
|
ISSN |
|
Download PDF |
|
|