Classification of Silent Speech in English and Bengali Languages Using Stacked Autoencoder

• Published in: SN computer science Vol. 3; no. 5; p. 389
• Main Authors: Ghosh, Rajdeep, Sinha, Nidul, Phadikar, Souvik
• Format: Journal Article
• Language: English
• Published: Singapore Springer Nature Singapore 22.07.2022; Springer Nature B.V
• Subjects: Accuracy; Brain research; Classification; Communication; Computer Imaging; Computer Science; Computer Systems Organization and Communication Networks; Data Structures and Information Theory; Discriminant analysis; Electroencephalography; English language; Fourier transforms; Human-computer interface; Imagination; Information Systems and Communication Service; Languages; Neural networks; Original Research; Pattern Recognition and Graphics; Principal components analysis; Prostheses; Signal processing; Software Engineering/Programming and Operating Systems; Speaking; Speech; Speech recognition; Support vector machines; Vision; Vowels; Deep learning; Stacked autoencoder; Imagined speech; Electroencephalogram; Silent speech; Brain–computer interface
• ISSN: 2661-8907; 2662-995X; 2661-8907
• DOI: 10.1007/s42979-022-01274-y

The purpose of a brain–computer interface (BCI) is to enhance or support the normal functions of disabled people, and as such, BCIs have been utilized for a variety of applications, such as prostheses and identification of mental state. One such application concerned with providing a means of communication for disabled individuals is focused on the recognition of silent speech (also known as imagined speech) in an individual. Silent speech can be defined as the speech originating inside the brain of an individual that has not been vocalized by the individual. The proposed work is concerned with the classification of silent speech from the brain activity of an individual recorded using an electroencephalogram (EEG). EEG data from 45 subjects were collected while they imagined the English vowels /a/, /e/, /i/, /o/, and /u/ without vocalization. EEG data were also recorded from 22 subjects who imagined five Bengali vowels /আ/, /ই/, /উ/, /এ/ and /ও/ without vocalization. The selected Bengali vowels have a similar pronunciation to the English vowels. Various temporal and spectral features were evaluated from the EEG recordings, which were then classified using a stacked autoencoder (SAE). The SAE achieved an accuracy of 75.56% and 73.6% in classifying the silent speech from the English and Bengali languages, respectively. Moreover, it has been observed that the proposed SAE outperforms conventional methods such as common spatial pattern (CSP) and support vector machine (SVM) during classification.