A Brain-Computer Interface (BCI) for Silent Speech Prediction using Machine Learning
Abstract
This work describes the development and implementation of a brain-computer interface (BCI) designed for silent speech recognition using multichannel electroencephalographic (EEG) signals. EEG provides a non-invasive means of capturing cerebral electrical activity through scalp electrodes, allowing neural signal patterns associated with specific cognitive processes to be examined. In this study, EEG patterns related to the imagined words "Yes" and "No" were identified and classified through machine learning methods. The proposed acquisition system records 16 EEG channels by means of two cascaded Texas Instruments ADS1299 integrated circuits, enabling high-resolution and synchronized multichannel measurements. Experimental classification using the K-nearest neighbors (KNN) algorithm yielded an accuracy of 85.9% in distinguishing between the two target words.
Keywords
Download Options
Introduction
Brain-computer interfaces (BCIs) provide an alternative means of interaction in which neural activity is translated into information that can be used to operate or communicate with external systems. The present work addresses the development of a multichannel brain-signal acquisition platform designed as the hardware basis of a BCI. This approach is particularly relevant to assistive technologies because it may allow individuals with severe motor impairments or paralysis to interact with their environment without relying on voluntary muscle activity.
In addition to motor-assistance applications, BCI research has increasingly explored communication alternatives for individuals who are unable to produce conventional speech. One approach involves the analysis of silent speech, in which a person attempts to formulate or articulate words without generating an audible acoustic signal. Although no sound is produced, speech-related neural activity remains present during imagined articulation or during movements of the speech organs. Detecting and characterizing these patterns therefore provides a potential pathway for developing communication systems that do not depend on audible vocalization.
In this work, neural activity was recorded by means of electroencephalography (EEG) [1]. EEG provides a non-invasive method for measuring variations in electrical potentials associated with brain activity using electrodes distributed over the scalp. To standardize electrode positioning across experimental sessions and participants, the recording locations were established according to the international 10–20 electrode placement system [2].
Previous research has demonstrated the potential of neural-signal processing for silent and imagined speech applications. Ramirez [3], for example, developed a machine-learning-based methodology aimed at dual silent communication through silent-speech analysis. Sarmiento [4] examined the principal models associated with human language processing and their relationship to BCI systems, with particular emphasis on approaches for recognizing speech without audible production. More recently, Kunz [5] demonstrated real-time decoding of self-paced imagined sentences selected from a relatively broad vocabulary, illustrating the progress achieved in extracting linguistic information from inner-speech-related neural activity.
The acquisition hardware developed for this study is based on the Texas Instruments ADS1299, an integrated analog front end specifically intended for measuring electrophysiological and other low-amplitude biopotential signals, including EEG [6]. Its combination of high-resolution analog-to-digital conversion, low input-referred noise, and low-power operation makes this device suitable for multichannel neural recording systems and BCI implementations. The acquired EEG data were processed and evaluated using different supervised machine-learning methods to investigate their ability to discriminate the recorded patterns. Among the evaluated classification techniques, the k-nearest neighbors (KNN) method provided the highest performance.
The remainder of this work presents the architecture and implementation of the proposed acquisition platform, including the electronic design and the main hardware components required for EEG recording and subsequent signal analysis.
Conclusion
The proposed platform demonstrates the feasibility of integrating multichannel EEG acquisition, signal visualization, and computational classification within a single BCI-oriented system. The ADS1299-based hardware provided the electrophysiological recordings required for subsequent processing, while the developed software interface enabled signal monitoring and data management. The complete acquisition and processing chain was evaluated through a proof-of-concept silent-speech classification experiment using machine-learning algorithms. Among the evaluated methods, KNN yielded the highest classification performance, reaching an accuracy of 85.9% for the experimental conditions considered in this study.
In a practical sense, this means that the developed system can translate imagined speech-related neural activity into discrete outputs that can be used to control external devices, demonstrating the viability of the proposed architecture for future assistive communication applications.
These findings indicate that the developed architecture can serve as a practical experimental platform for investigating non-invasive EEG-based communication strategies. Rather than representing a complete assistive solution, the present implementation establishes a technological foundation upon which more advanced BCI capabilities can be developed. Further work should include evaluation with larger datasets and participant groups, optimization of signal-processing and classification methods, and assessment under less controlled operating conditions. Such developments would help determine the robustness and generalizability of the system and its suitability for future assistive-technology applications.
References
- Carolina. (2009). Diseño y construcción de un prototipo de interfaz cerebro computador para facilitar la comunicación de personas con discapacidad motora. Revista EIA, (1794-1237), 11.
- Ramos-Argüelles, F. (2009). Técnicas básicas de electroencefalografía: Principios y aplicaciones clínicas. SciELO.
- Ramírez, D. (2015). Desarrollo de un sistema de comunicación silenciosa dual basado en habla subvocal y Raspberry Pi. Redalyc.org.
- Vela, L. C. S. (2019). Interfaces cerebro-computador para el reconocimiento automático del habla silenciosa. Education, Health Sciences.
- Erin, M. K., & Kunz, B. A. K. (2021). Inner speech in motor cortex and implications for speech neuroprostheses. CellPress.
- Texas Instruments. (2017). ADS1299-x low-noise, 4-, 6-, 8-channel, 24-bit, analog-to-digital converter for EEG and biopotential measurements. https://www.ti.com/lit/ds/symlink/ads1299.pdf
- Ameera, A. (2019). Análisis de las bandas del espectro EEG utilizando la densidad espectral de potencia para el estado de placer y displacer. IOP Conference Series: Materials Science and Engineering. https://doi.org/10.1088/1757-899X/557/1/012030
- Niedermeyer, E., & da Silva, F. L. (2005). Electroencephalography: Basic principles, clinical applications, and related fields (5th ed.). Lippincott Williams & Wilkins.