The Decoding Spectrotemporal Features of Imagined Speech

from Electroencephalography (EEG) and

Magnetoencephalography (MEG).
By

Deepan Waje (202201)

Livin Nadar (202202)

Rutuja Kamble (202203)

Aditya Pacharane (202204)

Om Khairnar (202205)

Under guidance of

Mr. Umesh Mhapankar

Diploma in Industrial Electronics

Agnel Polytechnic, Vashi.

Sector – 9A, Navi Mumbai

1
 Agnel Polytechnic

Sector – 9A, Vashi, Navi Mumbai-400703

Telephone: 41611000 Fax: 27660619

CERTIFICATE
This is to certify the completion of project work assigned. The project was successfully
completed by-

NAME SEAT NUMBER

Deepan Waje (202201)
Livin Nadar (202202)
Rutuja Kamble (202203)
Aditya Pacharane (202204)
Om Khairnar (202205)

The said work has been assessed by us and satisfied that same is up to standard envisaged
for the level of the course.
GUIDE Signature: _____________
Name: Mr. Umesh Mhapankar
Date: _____________
HOD Signature: _____________
Name: Mrs. Raji M.P
Date: _____________
INTERNAL EXAMINER Signature: ____________
Name: ____________
Date: _____________
EXTERNAL EXAMINER Signature: _____________
Name: _____________
Date: _____________

2
 ACKNOWLEDGEMENT
It is a genuine pleasure to express our deep sense of thanks and gratitude to our project guide
Mr. Umesh Mhapankar. His dedication and keen interest above all her overwhelming
attitude to help her students had been solely and mainly responsible for completing our
work. His timely advice, meticulous scrutiny, scholarly advice, and scientific approach have
helped me to a very great extent to accomplish this task.
We owe a deep sense of gratitude to Mrs. Raji M.P (HOD), her prompt inspiration, timely
suggestions with kindness, enthusiasm, and dynamism have enabled us to complete our
project.
We thank profusely all the staff for their kind help and co-operation throughout our study
period.

3
 ABSTRACT

The people suffering from an autism speech disorder, paralysis, or the mute people cannot
produce intelligible speech despite the imagined speech. People who cannot communicate
due to neuropathic disability will use the system which denotes imagined speech directly
from brain signals. Understanding the encoding procedure of the imagined speech signal at
the human cortex is a challenge due to the lack of visible and physiological measures. To
overcome this challenge and develop the system which can be used by physically
challenging with a speech disorder is the task for the researcher. This system has to be
wearable and use in day-to-day life.

This research aims to investigate the exact location of the speech processing area at the brain
scalp. Acquire electrical signal at localized speech area of imagined speech non-invasively
to decode spectrotemporal features of imagined speech. That will use to create speech
prosthetics and also be helpful for BCI in the future. Good spatial and temporal resolution is
required so that feature extraction and recognition will be better. Currently, most of the
research with invasive methods is not a practical solution. It is also risky and can be life-
threatening. The techniques without surgery like fMRI, PET scans are not wearable as well
as portable. It requires a big chamber with a total noise-free environment. The EEG and the
MEG are also techniques without any invasive procedure. Applications of The EEG and the
MEG are in the clinical diagnosis of brain disorders. The EEG and the MEG are wearable as
well as portable methods. The limitation of the EEG and the MEG is a low signal-to-noise
ratio, less bandwidth. The EEG has less spatial resolution same time the MEG has a poor
temporal resolution. To use the EEG and the MEG for decoding spectrotemporal features of
the imagined speech has to improve their limitation. The research objectives are to use the
correct electrodes for the system, locate the exact speech area on the scalp, improve the
signal processing method and analyzing method.

During the last year of research, we have been trying to locate the exact speech area on the
scalp of the brain using a literature study. Identification of the speech area on the scalp had
finished through an intense literature review. The correct portion on the scalp of the speech
area does not reduce only the number of electrodes required for the process but also
minimized the contamination of the bandwidth. We finished the task of selecting the correct
electrodes. An intense study of available literature on EEG electrodes had completed. The
characteristics of commercial dry electrodes are not comparable to the Ag/AgCl wet

4
electrodes. The Ag/AgCl wet electrode is the gold standard in the EEG measurement. The
wet electrodes limit the portability and wearability of the system.

We made the copper-based graphene electrode. Graphene is a two-dimensional nanomaterial

with high conductivity. It is derived from the graphite by standard procedure and then
alloyed with copper. The characterization of this electrode had completed with the typical
method. The results are adequate. Our next goal is to test the system for decoding the
imagined speech, feature extraction, and classification.

5
 PROJECT PLAN
WORK BREAKDOWN STRUCTURE
SR.NO. RESPONSIBILITIES MEMBERS
1
2
3
4
5
6
7
8
9
10
11
12
13
14

6
 SCHEDULING
SR.NO ACTIVITIES WEEKS

7
 TABLE OF CONTENTS

TOPICS

Chapter 1 INTRODUCTION

1.1 INTRODUCTION TO PROJECT

1.2 AIM OF PROJECT

1.2

1.3…………

Chapter 2 LITERATURE SURVEY

2.1

2.2 ….

Chapter 3 SCOPE OF THE PROJECT

Chapter 4 METHODOLOGY

4.1

4.2 ……

Chapter 5 DETAILS OF DESIGN, WORKING AND PROCESSES

5.1

5.2

5.3 ………………..

Chapter 6 RESULTS AND APPLICATION

Chapter 7 CONCLUSIONS AND FUTURE SCOPE

References

8
 LIST Of FIGURE

Name Of Figure Page No.

SR. No.
2.1 Selected electrode 6
2.2 The standard Position of the electrode 6
2.3 Principle of MEG 8
2.4 Relation of electric and Magnetic field 8
2.5 SQUID 9
2.6 Principle of SQUID 9
2.7 SERF 10
2.8 Principle of SERF 10
2.9 SERF 10
3.1 WLG placement 13
3.3 House model 13
3.4 Flow chart 14
3.5 System overview 14
3.6 The actual placement of the electrode 15
3.7 Testing of MEG sensor 16
4.1 Marathi Consonant 'CHa'using Mike 17
4.5 Marathi Consonant 'A' using Mike 18
4.6 Marathi Consonant 'CHA" using EEG 19
4.7 Continuous “A” using EEG 20
4.8 Continuous “Ka” using Mike 20
4.9 ECG using newly developed electrode 36

9
 List of Table

Name Of Table Page No.

SR. No.
2.1 Types of imagined speech 15
4.1 Comparison of result 36

10
 List of Abbreviation

Name Of
SR. Abbreviation Full form
No.
1 AE Active Electrode
2 BCI Brain-Computer Interface
3 BME Biomedical engineering
4 CT Computer Tomography
5 ECG Electrocardiography
6 ECoG Electrocorticography
7 EEG Electroencephalography
8 ERP Event-Related Potential
9 fMRI Functional Magnetic Resonance Imaging
10 fNRI Functional near-infrared
11 IFG Inferior Frontal Gyrus
12 MEG Magnetoencephalography
13 PET Positron Emission Tomography
14 SC Stratum Corneum
15 SNR Signal Noise Ratio
16 SQUID The superconducting quantum interface device
17 STG Superior Temporal

11
 CHAPTER 1
INTRODUCTION
Decoding Spectrotemporal Features of Imagined Speech
From Electroencephalography (EEG) and
Magnetoencephalography (MEG)

1.1 INTRODUCTION TO PROJECT

1.1 Background
The people suffering from an autism speech disorder, paralysis, or the mute people can not
produce intelligible speech despite the imagined speech. People who cannot communicate
due to neuropathic disability will use the system which denotes imagined speech directly
from brain signals (Herff C. et.al .2016). definition of the Imagined speech (S. Martin et.al
.2014) is "the ability to generate internal speech representations in the absence of any
external speech stimulation or self-generated overt speech." Imagined speech is also named
covert speech or verbal thought, silent speech, and inner speech. (Tanja S. et.al .2017).
However, the exact neural activities during imagined speech are still an active research area.
These neural activities produce electrical and magnetic signals that can be acquired and
graphically represented as Electroencephalography (EEG) and Magnetoencephalography
(MEG) wave. Understanding the encoding procedure of the imagined speech signal at the
human cortex is a challenge due to the lack of visible and physiological measures. To
overcome this challenge and develop the system which can be used by physically
challenging with a speech disorder is the task for the researcher. Furthermore, this system
has to be wearable and use in day-to-day life.

The structure and the function of the human brain are very complex. Speech production in
the human brain is the most complex cognitive process. This process starts with various
sensory organs, functions in the brain and is completed at the mouth cavity through the
respiratory tract. There are two reasons to understand speech processing in the brain. The
first reason, several people cannot speak due to various brain disorders, like paralysis,
autism speech disorder, lock-in syndrome, or mute people. They can think but are not able to
articulate and create overt speech. The second reason is that there is many [1] BCI (Brain-

12
Computer Interface) system in which the vocal signals are used to control the automation
system like the chair of the differently able person or patient assistant robot. The speech
production in the brain can be understood by acquiring electrical signals generated during
various brain processes. These electrical signals can be received in two ways, invasively and
noninvasively. The invasive method is called ECoG (Electrocorticography). In this method,
electrodes are placed on the brain cortex by surgery. The output of this method was accurate
and valuable for speech processing and decoding. Since surgery is required for this process,
it is not recommended for practical use. The other way is non-invasive and called the
EEG(Electroencephalography). In the EEG signal generated at the Scalp of the brain. In this
method, electrodes are placed on the Scalp through an adequately established procedure
depending on the electrodes type. The signal acquired through this method is a weaker and
low-frequency signal than the ECoG signal. The high-frequency signal also called the
gamma signal, is significant for decoding the speech signal. The selection of electrodes for
the EEG is more critical than the ECoG. Distinctly there are wet electrodes and dry
electrodes for the EEG acquisition. The wet electrodes are considered to be the gold
standard for EEG acquisition. The wet electrode is made of a solid solution with Ag and
AgCl; because Ag is a less dispersible salt, AgCl quickly soaks and comes to balance.
Chemically it is an adequate metal for skin surface electrodes. To use the wet electrode, the
preparation of the skin surface of the Scalp is necessary. Rub the Scalp surface with the
alcohol content solution to remove all the impurities at the electrode location. Apply the
electrolyte gel with an abrasive paste to reduce impedance between electrode and skin. The
recommended impedance is between 5/15 KΏ. The preparation of the skin and control of the
skin impedance requires an expert technician. Another drawback of these electrodes is
dirtiness due to the gel and paste. The annoying experience for the subject was due to the
time-consuming procedure. The drying of the gel and the loosing of the conductive
properties is one more issue of this electrode.[4] Due to the drawback of wet electrodes, the
dry revolution in the electrode fraternity is good hope for researchers of the biomedical
stream, especially neuropathy or neuro-scientist. For BCI application and speech decoding
purposes, the dry electrodes are more beneficial due to the cleanliness of electrodes and user
comfort. Several dry electrodes are available with different shapes and materials, such as
silicone conducting rubber, comb type, multiple pins, and needle type. Out that comparison
of four electrodes with various parameters is given in table 1. Figure 1 shows different
shapes of electrodes. This research selects the correct and adequate electrode to acquire
brain signals to decode imagined speech features. Specific characteristics of electrodes

13
should be compared, such as skin contact resistance of the electrode, electrode size,
frequency response of electrode, skin preparation, the requirement of electrolyte, firmness
with Scalp, adaptability, and user comfort. Several electrodes were tested and analysed
various kinds of literature on electrodes but found limitations in reaching the quality of the
gold standard wet electrode. While reviewing the literature, graphene is an engineering
material that exhibits excellent quality as a biomedical sensor. Its less affinity to copper
made the scientist create the graphene2022 International Conference for Advancement in
Technology (ICONAT) Goa, India. Jan 21-22, 2022 978-1-6654-2577-3/22/$31.00 ©2022
IEEE 1 copper-composite that keeps the suitable electrical property intact. With this
concept, the graphene-copper composite electrode was made tested with the above
characteristics and specific performance for the said research. This performance is also
tested under ambulatory conditions.

1.2 Problem Overview

There is a need for developing a system that can acquire imagined speech and convert it to
actual voice that will be helpful for the specific category of patients (autism speech disorder,
paralysis) and mute people. The system should be comfortable and convenient to wear on
the head. Placing the electrodes on the head should be without any surgery and easy. The
processing time for speech production should be minimum. Currently, researchers attempted
to develop similar systems. Some are surgical some are without surgery. In invasive
techniques, there are electrophysiological systems such as ECoG (Electrocardiographic) and
microelectrodes. Hemodynamic mechanism-basis processes are fMRI, fNIR, and
electrophysiological signals working on hemodynamic mechanisms. These methods are non-
surgical methods (Wang K. et. al .2017). The following paragraphs summarize the
advantages and limitations of these systems for decoding spectrotemporal features of
imagined speech.
Invasive systems like the ECoG and microelectrodes give better spatial and temporal
resolutions. These are significant features for different brain signals studies. The size of
electrodes in this technique is approximately 10 mm. There is no spatial contamination of
brain surface due to Dura matter, skull, or scalp in the invasive process. Therefore, they are
less vulnerable to muscle and surrounding artifacts. The spatial bandwidth of the ECoG
signal is above 200Hz. But these techniques require surgery of the brain area. It is possible
only if the patient suffers from any other diseases like the severe case of epilepsy. The

14
ECoG technique can place electrodes inside the brain for a maximum of two weeks, So this
technique is not practical and wearable.
There is the restriction of bandwidth up to 80Hz with poor spatial resolution
in the EEG technique. The skull and scalp with brain fluid act as low pass filters. The
spatial resolution is in centimeters. EEG is also very prone to artifacts due to muscle
movement, environmental conditions, and other motions. Despite the above limitation of the
EEG techniques used in BCI (Brain-Computer Interface) for communication purposes and
minimum speech processing algorithms (Yamaguchi H. Et.al. 2015; U.Ha et.al .2015). Some
researchers tried to extract spectrotemporal features of imagined speech using the EEG
technique. However, they were not able to describe sufficient information about speech
decoding. A selection of appropriate electrodes for decoding spectrotemporal characteristics
of imagined speech from available electrodes is a necessary objective. There are Ag/Ag- Cl
wet electrodes that require conducting jelly with a shaved skull. Other electrodes are dry
electrodes with a throne structure which may trouble the user. There are carbon nanotube
electrodes currently under research state. The exact positioning of the electrode on the skull
for extracting the spectrotemporal feature of imagined speech from the EEG is unknown
(Wang K. et.al .2017) and needs to study.
The MEG has an excellent temporal and good spatial resolution. It is less
contaminated by the scalp compare with the EEG. By muscle artifacts due to facial
movement, the MEG signal gets distracted. Its excellent spatial and temporal resolution
helps to study the speech morphology of the human brain in a different application.
Selection of appropriate electrodes and signal processing technique with a
magnetically shielded enclosure and its effect on the bandwidth and signal-to-noise ratio of
EEG and MEG need to study for 'Decoding Spectrotemporal Features of Imagined
Speech Electroencephalography (EEG) and Magnetoencephalography (MEG).'

1.3 Practical Applications:

The following application of the decoding spectrotemporal features of imagined speech from
EEG and MEG:
i. Thought-to-speech conversions for a speech disable a person.
ii. BCI (Brain-Computer Interface) system for controlling the robotics systems by
thoughts.

15
 iii. For children suffering from autism spectrum disorders (ASD). (In these children
are unable to express thoughts and emotions; and unable to interact in social
situations)
iv. It can implement as a lie detector by investigating officers for a crime.
v. It can implement for the old aged and bedridden persons for assistance.

1.4 Objectives:

1) To further explore the topic 'Decoding Spectrotemporal Features of Imagined

Speech from Electroencephalography (EEG) and Magnetoencephalography
(MEG)' for current state-of-the-art via literature review.
2) To find out research gaps in the topic 'Decoding Spectrotemporal Features of
Imagined Speech from Electroencephalography (EEG) and
Magnetoencephalography (MEG) 'based on literature review and propose a
methodology to solve the problem or gap.
3) To explore various applications of the research topic under consideration
4) To learn how to prepare technical reports and presentations.
5) To learn how to do time management.
6) To localize the speech processing area on the brain scalp by a literature study and
confirm it with the experimental setup.
7) To select appropriate electrodes/sensors to acquire the EEG and to compare
performance parameters.
1.5 Report Outline
The report consists of the chapter outlined below. Chapter one begins with a general
introduction to the research topic with background and problem definition followed by
an overview. Then, the research objectives are mention. Chapter two is the summary of
the literature survey completed in the previous year. This literature survey and review
are needed to select the topic. The study of literature has shown the path of current
research and shortcomings of its. This literature study gave an insight into the feasibility
of the subject. The end of this chapter is the conclusion to find the research gap.
Chapter three is about the proposed system, methodology, and a complete block
diagram of the system. Chapter 4 consists of results and discussion based on
implementation and tabulated performance. Chapter 5 consists of the conclusion of the

16
research from observation and experimentation. In the same chapter, it is detail about
future development

17
 CHAPTER 2
Literature Review
2.1 Summary of previous year literature review
This section presents a summary of the earlier literature review carried out on the following
topics which are relevant to the proposed research area:
• Anatomy and physiology of the brain
• Imagined speech and its features.
• A surgical and non-surgical technique for decoding spectrotemporal features of
imagined speech and current research gaps
• Generation of Electrical Potential in Neuron at the time of speech
• Comparison of dry electrodes and wet electrodes
• Graphene and its characteristics

2.1.1 Anatomy and physiology of speech processing in the brain:

Speech processing is the most complex of the higher cognitive brain function and is unique
in humankind. The anatomical placement described by the researcher regarding the language
is varying considerably. The different literature study gives an insight of speech processing
area. The speech processing area includes the Broca area(Inferior Frontal Gyrus), the
Wernicke area, the STG (superior Temporal Gyrus), the STS (Superior temporal sulcus),
and the primary auditory system(S. Martin et. al .2016; Diego Alfonso Rojas et. al .2016;
Grigorios N. et. al .2018). Researchers thought that the parts of the surrounding gyri with the
ventral parts of the pre-central and post-central gyrus, the super marginal gyrus, the angular
gyrus, and the medial temporal gyrus are part of the core language region. ( Tor-Tora 2011;
Jared M. N. et.al .2011)

Figure 2.1: Showing Broca area and Wernicke area (Tora-tor et. al .2011 )

18
The above survey is not enough to cover all areas associated with language processing in the
brain. In recent research suggested that other regions of the brain also taking part in speech
like the dorsal part of the area within lateral fissures, part of the thalamus, dorsolateral area
of the caudate nucleus and the cerebellum, visual areas, sensory pathways, motor pathways,
and pathways connecting all these regions. ( Grgorios N.2018, S. Martin 2016) The
physiology of language data processing is similar to any neurological processing in the
brain. For example, any stimulation for a speech like listening to somebody or observing
some exciting thing is enough to trigger speech processing in the brain. Geschwind-
Lichtheim-Wernicke Model was the dominant model for language processing brain in the
20th century.
GLW model as follows
1. The STG ( Wernicke area) represents the meaning of a word. When a person listens
to the speech, word sounds has sent through the auditory pathways to the primary auditory
cortex, Heschl gurus. From they have relayed to Wernicke area, where a sense of words has
extracted.
2. For speaking, it is necessary to send word meaning over the arcuate fasciculus to the
Broca area. Instruction about the speech, from the Broca place, will go to the adjacent facial
area of the motor neuron in the brainstem, which relays movement command to facial
muscles.

Cognition

Wernicke Area

Broca Area

The facial area of the

motor cortex

Cranial Nerve

Speak

Figure 2.2: Flow chart of Imagined speech to overt speech as per Geschwind-Lichtheim-Wernicke Model

19
2.1.2 Imagined speech and its features:

Following is a summary of the classification of speaking modes without an acoustic output

(covert speech) ( S.Martin 2017)
Table 2.1 Type of Imagined speech
SR Type of Description Measuring or monitoring
NO speech method
1 Silent Articulators will be move as wordy speech . The motion capture,
speech but suppress their pulmonary airstream so devices, imaging methods,
that sound will not produce or measuring muscle
activity has measured the
monitoring of the
articulatory movement
using motion-capture
devices, imaging methods,
or measuring muscle
activity
.
2 Imagined Imagined word is identical to first-person The imagined speech is
speech motor imagery of speaking in which the non-articulatory movement
speakers should feel as though they are speech. This speaking
producing speech rather than simply talking mode requires observations
to themselves. at the neural level.

3 Inner Though there is a range of descriptions for Inner speech is even more
speech inner speech (e.g., self-talk, verbal thinking, tough to investigate, even
inner voice, inner dialogue). that defines at the neural level
inner speech as an internalized process in of observation
which one thinks in pure meanings. The
inner speech is opposite to imagined and
silent speech. The phonological properties
and turn-taking qualities of an external
dialogue had used.

2.1.3 Invasive technique for decoding spectrotemporal features of imagined speech:

There are two methods to decode spectrotemporal features of imagined speech.
1)Invasive technique
2)Non-invasive technique
Invasive techniques: In this technique electrode has been placed on the cortex by surgery.
This recording is called ECoG (Electrocorticography). This measurement is possible if the
subject is growing through brain surgery due to some other reason like epilepsy or another
kind of disease where brain surgery is required. ECoG signals measured on the brain surface
do not suffer from spatial blurring from dura matter, skull, and scalp, record electrical
activity from neural tissue directly underneath each electrode and are less susceptible to
muscle and environmental artifacts. ECoG recordings have a spectral bandwidth of over 200

20
Hz, and more emphasis has the high-gamma band (>70 Hz), which is not readily observable
in scalp EEG. The high-gamma range is very spatially localized and highly correlated with
cognitive functions and behavioral output, including speech processes (Herff C. et.al .2015,
Tanja S. et.al .2017, S Martin et.al .2014; S.Martin 2016; Gerorg V.2014). The spatial
resolution of ECoG1mm to 10mm is better compare to EEG. Contamination of signals due
to muscle movement and other surrounding noise was less in the ECoG method. (Deng
S.2013; Wang J.2015; Grigarois N etal 2018; Gerald C etal 2017; Giovanni M.Di. et al
.2015)
2.1.4 Non-surgical technique for decoding spectrotemporal features of imagined
speech:
For non-surgical process does not require any surgery to place the electrodes on the body.
The electrodes are connected externally. In some non-surgical techniques, light rays or
ultrasonic sounds are sent to the body to investigate the disorder. (H.Akbari et. al. 2018; Hill
V. B.2019) Examples are fMRI, MEG, EEG, Sono Scan, Computer Tomography (CT
scan), etc. ( Price C et. al .2012; Hechun X. et. al .2012; S, Balliet 2001;)
2.1.4.1 fMRI:
The fMRI measured concentrations of oxygenated and deoxygenated hemoglobin in
the blood. It measures neural activity is measured. It is related to the increased demand for
oxygen as neurons are active and engaged. Oxygenated and deoxygenated hemoglobin has
different magnetic properties. The strong magnetic fields had produced in the MRI
environment measured the magnetic properties of hemoglobin. Due to the fMRI have high
spatial resolution over the entire brain, fMRI is the de-facto standard in neuroimaging
measures. It is instrumental in a variety of studies investigating speech and language. The
slow hemodynamic response, noisy environment, and the large chamber required for fMRI
significantly limit the utility for practical communication interfaces. (Cathy P 2012; Hill V.
B. et.al .2019)
2.1.4.2.fNIRs:
The fNIRS is a brain imaging technique pioneered by Jobsis (Jobsis P 1977) that also
detects changes in the amount of hemoglobin present in the brain as an indirect marker of
neural activity. Only the hemoglobin absorbed light in the near-infrared spectrum. The
biological tissue (e.g., bones, skin, and muscle) not absorbed the NIR( near-infrared). Thus,
the amount of hemoglobin present is estimated by placing near-infrared light emitters and
detectors around the head and calculating the amount of light absorbed. Similar to fMRI
paradigms, neural activity increases the energy demand. The fresh oxygenated blood

21
fulfilled energy demand that carries hemoglobin to the site of neural processing. The fNIRS
is well suited to investigate speech processes in non-clinical populations. It is less affected
by motion artifacts that plague EEG. fNRI not required laboratory environments. The fNIRS
emitters are simple LEDs that enable low-cost fNIRS devices. Additionally, the light
emitters and detectors do not require additional skin preparation steps common to EEG (e.g.,
skin abrasion and application of conductive gel), which simplifies acquisition. This method
is used as a marker for biosignal but not for spectrotemporal features extraction is not
possible.
2.1.4.3. EEG (Electroencephalography):
The EEG is the method to measure the electrical activity of the brain. For
measurement of electrical activity, the Professional placed the electrodes on the surface of
the scalp. At individual electrode sites due to the simultaneous activation of millions of
neurons whose summed voltage is conducted through the brain volume, skull, and scalp
layers. It is called the EEG. The large number of neurons contributing to the EEG signal,
combined with the properties of the low-pass filter of the skull and the scalp, resulting in a
spatial resolution on the order of centimeters and spectral bandwidth on the order of 80 Hz
maximum. As a non-intrusive measure of electrophysiological activity, EEG has desirable
temporal properties which help to characterize the neural processing of speech production.
EEG is highly susceptible to myoelectrical activity and artifacts. It interferes with EEG
recordings made during overt speech production (e.g., modal, whispered, and silent speech)
(Diego A. et.al .2016). There are methods to cancel this interference (Bagheri A. et al .
2017), validation is still needed to ensure that removed signals are artifacts from the EEG
signal. An alternative is to record EEG during the imagined speech. It restricts the analysis
to the speech motor planning and preparation phases. A comprehensive review is available
in (Xu J. et.al .2015 ) of the EEG components involved in speech and language processing.
EEG is recorded and processed using time-locked averages (i.e., event-related potentials,
ERPs) to overcome its comparatively low signal-to-noise ratio (.Zibrandtsen I.C. et.al
.2017; Durate M.de A. 2015). However, EEG also is analyzed as single-trial ERPs and for
changes in spectral content over time (e.g., event-related (de)synchronization) (Stolze et. al
.2016). Despite the disadvantages such as low bandwidth, low SNR (Signal to Noise Ratio),
contamination of signal due to muscle artifacts difficult for studying speech, EEG remains
the most common technique used in BCIs for communication (C. Herff et. al .2016)
2.1.4.4: MEG (Magnetoencephalography): Magnetoencephalography is the process of
measuring the brain signal due to neurons stimulation using magnetic sensors placed at the

22
vicinity of the head with a magnetic shield for limiting the environmental magnetic field.
MEG has excellent temporal and good spatial resolution and is less contaminated by the
scalp than EEG. However, Facial movement creates muscle artifacts which accounts for its
limited use in overt speech analysis. Its excellent spatial and temporal resolution helps to
study the speech morphology of the human brain in different domains. In the MEG
technique, a technician used the SQUID (Superconducting Quantum Interference Device)
sensors. The SQUID is a sensitive magnetometer to measure the subtle magnetic activity
based on superconducting loops containing Josephson junctions. SQUIDs are very sensitive
to measure fields as low as five aT (5×10−18 T) with a few days of averaged measurements.
Their noise levels are as low as 3 fT·Hz−½. For comparison, a typical refrigerator
magnet produces 0.01 tesla (10−2 T), and in some processes, the animals can generate
magnetic fields between 10−9 T and 10−6 T. Recently invented SERF atomic magnetometers
are potentially more sensitive and do not require cryogenic refrigeration but are orders of
magnitude larger in size (~1 cm3). It has to use in a near-zero magnetic field. The
traditional superconducting materials for SQUIDs are pure niobium or a lead alloy with 10%
gold or indium, as pure lead is unstable when its temperature is repeatedly changed. The
entire device needs to operate within a few degrees of absolute zero, cooled with liquid
helium for maintaining the superconductivity (Ukil A 2010; Puce A et.al .2017; Borna A
2017; Douglas O.C> Et.al.2017; Fernando L. 2013; Jiawei X.et.al.2017; Marlia D.S 2014;
S.Baillet 2001; Farzan F.2016; Ioannides A.A.2006)
Table 2.2: Comparison of different methods for decoding imagined speech (Panachakel et al.2020)
Method Temporal Spatial Type Portability
resolution Resolution
EEG 0.06ms 25mm2 noninvasive portable
MEG 0.1ms 1mm noninvasive Non portable
ECoG 0.02ms 4mm invasive Non portable
fMRI 500ms 0.7mm noninvasive Non portable
fNRI 100ms 100mm noninvasive Portable
ICE 3ms 0.05mm invasive Portable

2.1.5. Generation of Electrical Potential in Neuron at the time of speech.

Various kinds of ions, including positively charged Na+ (sodium) and K+ (Potassium) and
negatively, charged Cl_ (Chloride), had filled intracellular and extracellular fluids of a

23
neuron. These fluids also contain numerous negatively charged protein molecules.
Negatively charged ions are called anions (A_), a term that we will use for negatively
charged protein molecules, too. Three factors influence the movement of ions into and out of
cells ( Brayan Kolb et.al .2011 )
(1)Concentration gradient, (2) voltage gradient, and (3) the structure of the membrane.
2.1.5.1 Stages in Neurotransmission:
In four basic steps, the data has transmitted across a synapse :
(1) the transmitter molecules are synthesized and stored in the axon terminal, (2) the
transmitter is transported to the presynaptic membrane and released in response to an action
potential, (3) the transmitter interacts with the receptors on the membrane of the target cell
located on the other side of the synapse, and (4) the transmitter is inactivated (or it will
continue to work indefinitely).
2.1.5.2 Brain waves classification
For obtaining basic brain patterns of individuals, subjects have to close their eyes and relax.
Brain patterns form a sinusoidal wave shape. Usually, they are measured from peak to peak
and normally range from 0.5 to 100 μV in amplitude, which is about 100 times lower than
ECG signals. Using Fourier transform power spectrum from the raw EEG signal is derived.
In the power spectrum contribution of sine waves with different frequencies are visible. The
spectrum is continuous, ranging from 0 Hz up to one-half of the sampling frequency, the
brain state of the individual may make certain frequencies more dominant. Brain waves have

Figure 2.3 Brain wave samples with dominant frequencies belonging to beta, alpha, theta, and delta band
been categorized into four basic groups (Figure 2.3):
- beta (>13 Hz),
- alpha (8-13 Hz),
- theta (4-8 Hz),
- delta (0.5-4 Hz).

24
The best-known and most extensively studied rhythm of the human brain is the normal alpha
rhythm. Alpha is usually observed better in the posterior and occipital regions with a typical
amplitude of about 50 μV (peak-peak). According to observations, an alpha wave was
prominent between the rear and central areas compare to other regions. The alpha wave
generates at the time of eye closing and relaxation. The Beta waves are present at the time of
active brain like thinking or playing. Most people are proactive to the phenomenon of eye
closing. It is usually they close their eyes their wave pattern significantly changes from beta
into alpha waves. The precise origin of the alpha rhythm is still not known. Alpha waves are
the outcome of summated dendrite potentials. The ERP (Event-related potential )often
consists of fiber potentials (axonal) and synaptic electrical components is sensitive to a
continuum of states ranging from stress state, alertness to resting state, hypnosis, and sleep.
In relaxation or drowsiness, the alpha activity rises, and if sleep appears, the power of lower
frequency bands increases. Sleep are two types: nonrapid eye movement sleeps (NREM) and
REM sleep. NREM and REM occur in alternating cycles.
The NREM has four parts: stage I, stage II, stage III, and stage IV. The last two stages
correspond to deeper sleep, where slow delta waves show higher proportions. So The
various type of sleep patterns and wakefulness, the different parts of the brain produce
variable EEG signals. An EEG signal between electrodes placed on the scalp consists of
many waves with other characteristics. A large amount of data received from even one
single EEG recording presents a difficulty for interpretation. Every person has unique brain
wave patterns. In some cases, it is possible to distinguish persons only according to their
typical brain activity.
2.1.5.3 Evoked potentials
Evoked potentials or event-related potentials (ERPs) are significant voltage fluctuations
resulting from evoked neural activity. An external or internal stimulus initiates the evoked
potential. ERPs are a suitable methodology for studying the aspects of cognitive processes of
both normal and abnormal conditions (neurological or psychiatric disorders)
Mental operations, such as those involved in perception, selective attention, language
processing, and memory, proceed over time ranges in the order of tens of milliseconds.
Amplitudes of ERP components are often much smaller than spontaneous EEG components,
so they cannot measure from raw EEG traces. Instead, they are extracted from the set of
single recordings by digital averaging of epochs (recording periods) of EEG time-locked to
repeated presence of sensory, cognitive, or motor events. The spontaneous background EEG
fluctuations, which are random relative to a time point when the stimuli occurred, are

25
averaged out, leaving the event-related brain potentials. These electrical signals reflect the
activity that is consistently associated with the stimulus processing in a time-locked way.
The ERP thus reflects, with a high temporal resolution, the patterns of neuronal activity
evoked by a stimulus.
2.1.5.4 EEG Biofeedback So-called mind machines or brain machines are devices for
induction of different mind states (relaxation, top performance) by entrainment of the brain
waves into desired frequency bands by repetitive visual and audio stimuli. For example,
EEG biofeedback or neuro-feedback uses EEG signals for feedback input.
2.1.6 Electrodes / Sensor for EEG
One of the keys to recording good EEG signals is the type of electrodes used. Electrodes that
make the best contact with a user scalp and contain materials that most readily conduct
EEG signals provide the best EEG recordings. ( M. A. Lopez-Gordo et.al .2014)
Some of the types of electrodes available include:
1. Reusable disks.
These electrodes can be placed close to the scalp, even in a region with hair because they are
small. A small amount of conducting gel is necessary under each disk. The electrodes are
held in place by a washable elastic headband. Disks made of tin, silver, and gold are
available. The cost of each disk and lead depend on
1)the type of metal used for conductor
2) the gauge of wire used for lead,
3)the type of insulation on the wire lead.
About applications of the sensor, the cost is low. For example, a set of 5 silver/silver
chloride disks with silicone-coated lead wires, Tac-Gel, and a headband is available from
iWorx for $90.00/set (Part No. A-E-264).

Figure 2.4 Reusable Disk(iWorx)

2. EEG Caps with disks.
Different styles of caps are available with numbers and types of electrodes. The disk
electrode is required to fill the gel through a hole behind the disk. Due to the hair on the

26
head, it is cumbersome to place the electrode at the correct location. A more amount of
conducting gel needs under each electrode. It is a dirty process.

Figure 2.5 EEG electrode with cap (iWorx)

3. Adhesive Gel Electrodes.

These are the same disposable silver/silver chloride electrodes used to record ECGs and
EMGs. These electrodes are an inexpensive solution for recording from regions of the scalp
without hair. They cannot be placed close to the scalp with dense hair since the adhesive pad
around the electrode would attach to hide and not the scalp. When purchased in bulk, their
expense is very low.

Figure 2.6 Adhesive gel electrode (iWorx)

4. Subdermal Needles.
They have placed the sterilized single-use needle under the skin. It is available with
permanently attached wire leads. For this electrode at the time of replacement, they need to
replace the whole assembly included sockets and matching plugs of the needle-type
electrode. Since they are a sterile single-use item, the expense of needle electrodes is
moderate to high. Also, for human subjects and, in some situations, regulatory committees
need to approve the use of these electrodes

27
 .
Figure 2.7 Subdermal needle(iWorx)

5. Dry Electrode:
The electrode can be used without conducting gel and without skin preparation are called
dry electrodes. The electrode surface consists of an array of spikes. The spike directly
comes into contact with the scalp or pierce the SC for better electrical attachment and
mechanical fixation to the scalp. The Spikes are in the scale of nanometers, micrometers
(MEMS), and millimeters. Electrode configurations with different aspects and
characteristics were analyzed, revealing that the most relevant factors that affected the
impedance were the electrode size and the coating material.

Figure 2.8 Dry electrode( Lopez MD 2014)

6 Active electrodes :
Active electrodes (AEs), i.e., electrodes with built-in readout circuitry, are increasingly
being implemented in wearable healthcare and lifestyle applications due to AE robustness to
environmental interference. An AE locally amplifies and buffers μV-level EEG signals
before driving any cabling. In addition, the low output impedance of an AE mitigates cable
motion artifacts, thus enabling the use of high-impedance dry electrodes for greater user
comfort.

28
 Figure 2.9 Active Electrode (iWorx)

2.1.7 The research on an imagined speech to overt speech.

Researchers have developed a mechanism to translate brain activity into simple sentences, a
breakthrough that could lead to new tools for people who cannot communicate through
speech. A state-of-the-art brain-machine interface created by UC San Francisco
neuroscientists can generate natural-sounding synthetic speech using brain activity to control
a virtual vocal tract – an anatomically detailed computer simulation including the lips, jaw,
tongue, and larynx. The new system developed in the laboratory of Edward Chang, MD –
described on dated April 24, 2019, in Nature Journal– demonstrates that it is possible to
create a synthesized version of a human voice, controlling the activity of their brain speech
centers. In the future, this approach could not only restore fluent communication to
individuals with a severe speech disability, the authors say, but could also reproduce some of
the musicality of the human voice that conveys emotions and personality. It builds on a
recent study in which the pair described for the first time how the human brain speech
centers choreograph the movements of the lips, jaw, tongue, and other vocal tract
components to produce fluent speech. This detailed mapping of sound to anatomy allowed
the scientists to create a realistic virtual vocal tract for each participant.
The system comprised two neural network machine learning algorithms: A decoder
transforms brain activity patterns produced during speech into movements of the virtual
vocal tract, and a synthesizer converts these vocal tract movements into a synthetic
approximation of the voice ( Gopala Anumanchipalli et.al .2019). Nima Megharani and his

29
team from the Columbia university San Fransico have done similar research in which
auditory stimulus reconstruction is a technique that finds the best approximation of the
acoustic stimulus from the population of evoked neural activity. Reconstructing speech
from the human auditory cortex creates the possibility of a neuroprosthetic to establish
direct communication with the brain and is possible in both overt and covert conditions.
However, the low quality of the reconstructed speech has severely limited the utility of
this method for brain-computer interface (BCI) applications. They investigated the
dependence of reconstruction accuracy on linear and nonlinear (deep neural network)
regression methods and the acoustic representation. It is for the target of reconstruction,
including auditory spectrogram and speech synthesis parameters. In addition, they
compared the reconstruction accuracy from low and high neural frequency ranges. Their
results show that a deep neural network model that directly estimates the parameters of a
speech synthesizer from all neural frequencies achieves the highest subjective and
objective scores on a digit recognition task, improving the intelligibility by 65% over the
baseline method. They used linear regression to reconstruct the auditory spectrogram.
These results demonstrate the efficacy of deep learning and speech synthesis algorithms
for designing the next generation of speech BCI systems. That is not only can restore
communications for paralyzed patients but also have the potential to transform human -
computer interaction technologies (Nima Megharani et.al .2019.)
2.1.8 Position of EEG electrodes
The 10-20 System of Electrode Placement is a method used to describe the location of scalp
electrodes. In the international, 10-20 system electrodes are placed as per geometry of head.
Each point on this figure to the left indicates possible electrodeposition. Each site has a letter
(to identify the lobe) and a number or another letter to identify the hemisphere location. The
letters F, T, C, P, and O stand for Frontal, Temporal, Central, Parietal and Occipital. Even
numbers (2,4,6,8) refer to the right hemisphere. Odd numbers (1,3,5,7) refer to the left
hemisphere. They used the Z alphabet for the midline electrode. Note that the smaller the
number, the closer to the midline.
Nasion - point between the forehead and nose.
Inion - Bump at the back of the skull.
The "10" and "20" refer to the 10% or 20% inter-electrode distance

30
 Figure2.10 Position of the electrode in 10-20 method
2.1.8 LAB VIEW
National Instruments has designed a sophisticated biomedical engineering (BME)
instrument to integrate state-of-the-art technology with a hands-on learning approach in a
flexible, virtual-based, clinical application setting. The need for biomedical engineers in
research and industry has increased rapidly in recent years. It requires that innovative
methods for training BME students evolve to meet that need. BME students should prepare
with a skill set for approaching practical problems. BME education requires hands-on
learning with cutting-edge technology to produce students ready to solve clinical problems
in research and industry. Exposing students to a wide range of BME applications increases
interest and prepares them to solve real-world problems. It is also better prepares them to
solve real-world problems. National Instrument has designed a wide range of biomedical
instruments to encompass both the basics of physiological signals and utilize them in
practical applications effectively. Using the virtual environment for practical applications
bridges the gap between fundamentals and real-world designs.
LabVIEW is Laboratory Virtual Instrumentation Engineering Workstation[2]. With the
increasing performance of a computer, virtual instrument technology has dramatically
advanced over the years, and then virtual medical instrument technology becomes available.
What is a virtual medical instrument? What is a "virtual" anything? A virtual item exists in
function but not in solid form. Does that mean that a virtual medical device doesn't exist?
No. It exists, but not in the state that we are used to seeing. OK, so, what's a virtual medical
instrument? A virtual medical device is a new style of high-tech product in a company with
computer technology and modern medical measurement technology. It grows up to the
minute trend of the medical instrument of nowadays. The virtual medicine instrument makes
the best of active computer resource, matches with oneness constructive instrument

31
hardware and proprietary software, and achieves all the functions of the traditional tool and
some significant specific tasks that can't carry out on the conventional instrument. A virtual
medical agent is adding software and hardware to the general-purpose computer. At the same
time, the user operates this computer as if operating special traditional medical equipment
that he designed by himself. The appearance of the virtual medical instrument technology
breaks through the mode that the manufacturer defines traditional medicine instrument, and
the user can't change it. Given the adequacy space to exert ability and imagination, user can
design own instrument system to satisfy various application demand.
Personal Computer is becoming ever more popular in the medical community as prices
decrease while performance increases. Many practitioners use PCs for patient records and
information. A PC-based system consists of a few external hardware components for
isolation and amplification of the signals, data acquisition cards (or parallel port), and a
software analysis package. PC-based instrumentation can bypass the need for stand-alone
instruments using the PCs currently available and some inexpensive acquisition equipment.
The virtual medical instrument is the testing platform based on computer software and
hardware. It consists of a computer system and can replace traditional medical devices, such
as electrocardiographs, and Electro-encephalographic data can be displayed, memorizing
and analyzing through the software.
A medical instrument that develops electronic measurement instruments has gone through
the simulated instrument, intelligent instrument, and virtual instrument. The advantage of
virtual medical instrument technology lies in the user-defined system, agile functions, and
easy construction. It is used widely in many fields, especially in scientific research,
exploitation, medical measurement, medical detection, medical signal process, etc. It has
powerful functions to carry out all of the parts of the traditional medical instruments, from
data waveform display, logic analysis, spectral analysis, signal generation to medical image
display and disposal. Matching with the medical sensor and software can measure various
medical parameters such as body temperature, pulse wave, and blood pressure. It is
operation agility and a complete graph interface so that new users can master the operation
rules without any training. It is easy to construct an automatic measurement system with
high-speed medical data acquisition equipment. ( National instrument 2013)

32
2.1.9 Research Gap.
i. The ECoG-based technique can decode the spectrotemporal feature of speech.
Signal bandwidth is 200Hz and above is better than the EEG technique (80Hz
maximum). Excellent temporal (less than 1ms) and spatial resolution (1mm to
10mm). Imagined or covert speech is possible to convert into text with ECoG.
The ECoG is a fully invasive technique. It is not fit to make the system wearable
or portable as well as comfortable to the user. The ECoG technique requires
surgery to put electrodes underneath of skull. It is unpractical. ( Herff C. 2015,
Martin S. 2016, Derix J. 2014)
ii. The fMRI technique is non-surgical. It is the gold standard for the diagnosis of
any physiological impairment. The size of the machine (chamber) is the problem.
It is also prone to environmental noise. In addition, the fMRI has a poor temporal
resolution. So it has limitations for its use as a portable and practical system or
wearable system.
iii. fNIRs technique is an indirect marker for the metabolism of the human body. It
measures the percentage of oxygen in haemoglobin but fails to extract
spectrotemporal features of bio signal-like continuous speech signals.
iv. EEG has recently become the gold standard for BCI (Brain-Computer Interface)
communication. However, the EEG has a limitation of bandwidth. Moreover, it is
poor spatial resolution (around 3cm) compare to ECoG and contamination due to
body artifacts. It is also prone to environmental conditions hence seldom used for
spectrotemporal feature extraction and automatic speech recognition. Therefore,
it is a need for an optimum design of the system with the EEG to improve to
decode of spectrotemporal features of imagined speech.
v. MEG has good spatial resolution and moderate temporal resolution. The
magnetic field generated due brain signal is weak (10-9T) less than the
surrounding magnetic field(10-4T). To sense the small magnetic at the brain
signal requires a different type of magnetic sensor like SQUID. It is costly and
operates at absolute 0 Kelin. It is impractical to use in a non-laboratory
environment.
vi. Most techniques are invasive, i.e., Brain surgery is necessary to do the research.
Therefore, it is not practical, and it may be hazardous to people who are
undergoing surgery.

33
 vii. Creating a prosthetics Device or wearable device is impossible.
viii. In the noninvasive technique, most researchers used regular patterns to achieve
the result, including the selection of electrodes including positioning electrodes
.i.e .10-20 international standard method, which is suitable for clinical diagnosis
and sleep profile but not useful for research
ix. They worked the basis of statistical data in most cases.
x. A deep study of brain physiology and study synaptic data transmission is missing.
xi. EEG has a poor signal-to-noise ratio.
xii. Recording EEG using Ag/AgCl wet electrode for an extensive period is the task
because the conductive gel dried up.
xiii. A skilled technician is necessary to operate the system.
2.2. Current literature Review
2.2.1 Mapping the location of imagined speech electrical field from cortex to scalp
EEG-based systems have many advantages, like cheaper and non-surgical. It has a fair
temporal resolution, although ECoG has a better temporal resolution. Because of distance,
different nonconducting layers, and brain fluid between cortical electrical field source of the
imagined speech and scalp electrodes, temporal resolution is limited (Illman et al .2020;
Ghafoor et al .2019; S. Martin et al .2016; Panacakel J.T. et al .2021). Another issue of the
application of EEG is the time required to make the setup with the dense electrode, and the
wet electrode is large. Designed the customized dry electrode and mapped the exact location
on the scalp to reduce the electrodes. The system will become wearable or user comfortable
(Panacakel J.T. et al .2021).

Figure 2.11: WLG model figure

We reviewed the existing literature on imagined speech decoding for the mapping, the
location on the scalp. Unfortunately, it isn't easy to develop a pictorial language processing
model of the human brain. Nevertheless, the Broca-Wernick-Litchtheim-Geschwind model

34
based on their work model was the standard for the centuries. This model is based on figure
2.11. However, this model is not significant as per new research, as Trembley and Dick
stated since it did not give any distributed framework related to speech connectivity found in
a recent study (Grigorios Nasios et al .2018). A new study about auditory processing said
that speech processing is part of the cortical and subcortical areas. Therefore, there are two
streams, dorsal and ventral pathways.
The article was published in July 2019 in AJNR( AM J Neuroradial). The authors have
reviewed the Images taken by fMRI (figure2.12), showing the activated area during the
speech, which confirmed the WLG model for speech pathway. (Hill V. B. et al. 2019)

Figure 2.12: Sagital 3D image of arcuate fasciculus between inferior parietal lobe (green) and angular gyrus
(red). The frontotemporal segment of the arcuate fasciculus connects the Broca area at the inferior frontal gyrus
with the Wernicke area more posteriorly, at the posterior superior temporal gyrus. (Hill V. B. et.al .2019)

Another important fact is that the mapping from cortex to scalp should be precisely
orthogonal. ( )

35
 Figure.2.13 Stages of speech production according. The color-coded table to the right summarizes the stages
of processing in speech production and time course in response to an image stimulus. (Rabani Q. et al . 2019)

Figure 2.14 Spectral map of STG at time of speech processing Ref. (Hullet et al. 2016)

For mapping speech location on the scalp from the cortex is necessary to study neurophysiology
despite well-established progress in machine learning and automatic speech recognition (ASR).
A previous study with fMRI has mapped the brain areas responsible for processing different
aspects of speech and the general dynamics of how they interact (Rabani Q. et al .2019).
Figure 2.13 shows the model in which six separate speech production signals take the path from
conceptual preparation to phonological code at different location brain auditory pathways. The
literature has cleared the significance of STG and high Gamma (80Hz-150Hz) band signal in

36
speech processing.(Mesgarani et al.2016 and Hullet et al.2016). The posterior STG responds
to short and varying speech sounds with relatively constant energy across the frequency axis
(i.e., low spectral modulation). In other conditions, the anterior STG responds for temporally
slow varying speech sounds with a high degree of spectral variation across the frequency
axis (i.e., high spectral modulation), as illustrated in Figure 2.14 ( Hullet et al. 2016). On the
other hand, (Hamilton et al. 2017) observed that spatially localized, caudal, or posterior regions
of the STG are more sensitive to speech onset, while spatially distributed, rostral or anterior
and medial parts of the STG are more susceptible to sustained speech. Together, these studies
suggest that ECoG adequately captures spectrotemporal tuning of the STG to voice organized
by acoustic features rather than by discrete phoneme categories.

Fig. 9 Shared neural substrate between overt and covert speech. Shared activation between overt and covert
speech is located primarily in the superior temporal gyrus, premotor cortex, and primary motor cortex
Ref( Brumberg J S, 2016)
The literature study is enough to identify the location and then mapped on the scalp to
minimize the number of electrodes.

2.2.2 Electrode Design:

As in previous literature, we had compared various electrodes available commercially. The
wet Ag/AgCl electrode consider as a gold standard for EEG measurement. However, there
are many limitations of Ag/AgCl electrodes when designing the wearable and practical
system. For example, existing dry electrodes did not fulfill the required characteristics for
decoding the spectrotemporal features of imagined speech like bandwidth, skin contact
impedance. So it is necessary to design electrodes. At the thought of creating a new
electrode, we come across graphene material.

37
2.2.2.1 Graphene, the engineering material
It is a two-dimensional nanomaterial with multiple unique characteristics, such as carrier
mobilities ( Weiss et al., 2012). It exhibits superior electrical conductivity, excellent thermal
conductivity (Balandin et al., 2008; Balandin, 2011). Large active area (Zhu et al., 2010). These
properties help us create an excellent biomedical sensor that can easily extract weak brain
responses like structure. The advantages of graphene for sensors are summarized as follows:
the first point is that the high specific surface area and the atomic thickness of graphene
layers render entire carbon atoms directly connected with analytes. As a result, graphene-
based sensors have superior sensitivity compared to silicon (Justino et al., 2017). In addition,
graphene-based sensor/electrode can create close contact with the skin (Ameri et al., 2016),
brain (Park et al., 2017), and eyes (Kim et al., 2017). Du to the mechanical flexibility and
ultrathin thickness of graphene, which is essential in acquiring high-quality signals without
irritation, motion artifacts, or contamination (Ray et al., 2018). An excellent signal-to-noise
ratio will adhere to the graphene base sensor (Ameri et al., 2016). There is multiple
application of health monitoring system of graphene such as the real-time measurement of
body temperature (Trung and Lee, 2016; Wang et al., 2018), heart rate (Karim et al., 2017),
wrist pulse (Yang et al., 2017; Pang et al., 2018), respiration rate (Boland et al., 2014; Xu et
al., 2018), blood pressure (Pang et al., 2016), blood glucose (Pu et al., 2018). It is also used
to extract the ECG, the EEG (Ameri et al .2016; Yun et al .2017), and the EMG (Yun et al.,
2017; Sun et al., 2018), etc.

2.2.3. Data acquisition and processing:

2.2.3.1 Electrode and Sampling:
A researcher has acquired speech data with different methods, and they processed it
primarily offline (refer to Table 2.3)(Khan and Hong 2017; Tayeb et al .2019). For practical
and BCI systems the online data processing is essential (Khan and Hong 2017). For most of
the study, the researcher has used the EEG system with 64 electrodes(Panchakale et al
.2019; Shah and Fels, 2019; and many more ref table 2.3). Researchers used 22 electrodes or
implemented groups of electrodes depending on the portion taking part in speech production
and the remaining area. Sampling used in the system is varied from 2Khz to 200Hz, and
some systems down-sampled afterward. The researcher has used commercial as well as
research type of EEG system.

2.2.3.2 Symbol, Feature extraction, and classifier:

38
It is difficult to compare the accuracies reported in different kinds of literature due to the
differences in the data acquisition protocol, including the differences in the number of EEG
channels, number, and nature of imagined speech symbols. Thus, even for the works using
the same dataset, a proper comparison is impossible. Furthermore, the evaluation strategy
(number of folds in cross-validation, classification of individual subjects vs. pooling the data
from the entire set of topics for category, using a subset of the available symbols in the
dataset) varies across these studies. Nevertheless, a comparison of the accuracy reported in
several works reviewed in this manuscript is given in Table 2.4.
Below, we analyze the performance of the systems based on the types of symbols used,
namely:
1. Directional Symbol
2. Polar Symbol
3. Vowel symbols
They mean up, down, left, right, forward, and backward, respectively. Cooney et al. (2020)
have used Spanish words. The six Spanish words used by Cooney et al. (2020) are “arriba”,
“abajo”, “derecha”, “izquierda”, “adelante”, and “atrás” which mean up, down, left, right,
backward, and forward. The EEG data for Pawar and Dhage (2020) is obtained using 64-
channel Neuroscan synapse two research-grade EEG acquisition systems sampled at 1,000
Hz. Koizumi et al. (2018) used a 65- channel EEG-1200, Nihon Kohden Corporation
research-grade EEG acquisition system sampled at 1,000 Hz. In contrast, Cooney et al.
(2020) used the dataset acquired using 18-channel Grass 8-18-36 commercial grade EEG
acquisition system sampled at 1,024 Hz. Accuracy can not be attributed to the system type
(commercial grade/research-grade) because García-Salinas et al. (2019), who also made use
of a commercial-grade system like Cooney et al. (2020), have obtained much better
performance than Cooney et al. (2020). Also, the data sampling rate may not have affected
the accuracy. One key difference between Koizumi et al. (2018) and other works is the use
of the gamma band. Since both Pawar and Dhage (2020) and Koizumi et al. (2018) have
used the gamma band, the higher performance of Koizumi et al. (2018) cannot be attributed
to the use of gamma-band alone.
2.2.3.2.1 Polar Symbol:
Polar symbols are the responses to binary questions or polar questions. Three studies
reviewed in this article have made use. The value of κ theoretically lies in the range [−1, 1].
Values closer to −1 indicate that the system performs poorly, whereas a more relative to 1
means that the system is perfect. A value of 0 indicates that the classifier is only as good as a

39
random guess. In contrast, a value less than 0 means that the classifier's performance is
inferior to a lucky guess.
2.2.3.2.2. Directional symbol
Directional prompts include words used for controlling devices such as wheelchairs and user
interfaces like computer pointing devices. Five studies reviewed in this article make use of
directional prompts. In both García et al. (2012), García-Salinas et al. (2019), five Spanish
words, “arriba”, “abajo”, “izquierda”, “derecha”, and “seleccionar” are used as the prompts.
These words mean up, down, left, ‘right, and select, respectively. The prompts used in
Pawar and Dhage (2020) are “left,” “right,” “up,” and “down.” In Koizumi et al. (2018), six
Japanese words “ue”, “shita,” “hidari,” “mini,” and “mae” are a 64-channel BrainAmp
research-grade EEG acquisition system with a sampling rate of 1 kHz for acquiring the EEG
data. On the other hand, Balaji et al. (2017) used a 32-channel research grade (Electrical
Geodesics, Inc.). EEG acquisition system with a sampling rate of 250 Hz. Unlike Sereshkeh
et al. (2017a) and Sereshkeh et al. (2017b), in Balaji et al. (2017), the binary questions were
posed in two languages, namely Hindi and English. Also, Sereshkeh et al. (2017b) is the
only work that uses an online strategy for decoding imagined speech from EEG.
The following conclusions can be made from the results presented in Balaji et al. (2017):
• Though all the participants were native Hindi speakers who learned English only as
their second language, the classification accuracy is better when the binary questions are
posed in English rather than in Hindi.
• When the responses to all the questions (both Hindi and
English) are pooled together and used for classification, only rarely does the classifier make
a cross-language prediction
2.2.3.2.3 Vowel Symbol
Four studies reviewed in this study have used vowel imagery in their paradigm. Min et al.
(2016) and Cooney et al. (2020) have used the entire set of vowels as their prompts, whereas
Nguyen et al. (2017) and Saha and Fels (2019) have used only three vowels: /a/, /i/, and /u/.
Min et al. (2016) have used a 64-channel, research-grade Electrical Geodesics, Inc. EEG
acquisition system. In contrast, Nguyen et al. (2017) have used a 64-channel, research-grade
BrainProducts ActiCHamp EEG acquisition system, both sampled at 1000 Hz. Min et al.
(2016) and Nguyen et al. (2017) have downsampled the acquired data to 250 Hz and 256 Hz.
Saha and Fels (2019) have used the EEG dataset created by Nguyen et al. (2017). On the other
hand, Cooney et al. (2020) have used an 18-channel, commercial-grade EEG amplifier (Grass
8-18-36) for acquiring the data at 1024 Hz. Data was later downsampled to 128 Hz Saha,

40
and Fels (2019) performs better than those proposed by Min et al. (2016), Nguyen et al.
(2017), and Cooney et al. (2020). Since Nguyen et al. (2017) and Saha and Fels (2019) have
used the same EEG dataset, the improvement can be attributed to the superior classification
technique used by Saha and Fels (2019). Nguyen et al. (2017), Saha and Fels (2019), and
Cooney et al. (2020) have also tested their approach on the EEG data acquired when the
participants were imagining articulating short words (Cooney et al. (2020): “Arriba,”
“Abajo,” “Derecha,” “Izquierda,” “Adelante,” and “atrás”; Nguyen et al. (2017) and Saha
and Fels (2019): “in,” “out,” and “up”). For both Nguyen et al. (2017) and Saha and Fels
(2019), there is a marginal improvement in the κ values when short words are used instead
of vowels, whereas for Cooney et al. (2020), there is a marginal reduction. Therefore, we
cannot concretely claim any advantage for short words over vowels when used as prompts
for imagined speech.

Table2.3: Comparison of the different work EEG systems, number of an electrode, sampling
SR Literature Type of EEG Electrode Sampling Decoding Degree of
No system rate and strategy freedom
resample
1 Min et al. Research 22 wet 1kHz Offline 2
(2016) electrode 250Hz
2 Nguyen et al. Research 22 wet 1KHz Offline 3
(2017) electrode 256Hz
3 Panachakal et Research 64 channel 1KHz Offline 2
al. (2019) wet 256Hz
electrode
4 Cooney et al. Commercial Emotive 1KHz Offline 5
(2018) dry metal 128Hz
electrode
Saha and Fels Research 64 channel 1KHz Offline 3
5 (2019) wet 256Hz
electrode
6 Jahangiri et Commercial Emotive 2KHz Offline 4
al. (2019) dry 256Hz
electrode
7 Pawar and Research 64 channel 1Kz Offline 4

41
 Dhage (2020) wet
8 Koizumi et Research -do- 1KHz Offline 12
al.(2018)
9 Zhang et al. Research -do- 500Hz Offline 4
(2020)
10 Watanabe et Research -do- 1KHz Offline 3
al. (2020)
11 Chegalayan Commercial emotive 256Hz Offline 5
et al. (2020)
12 Jahangiri and Research 64 channel 2KHz Offline 4
Sepulveda Ag/AgCl 256Hz
(2019) wet active

Table 2.4: comparison of accuracy with different symbols, features extraction, and classification algorithms
Sr Paper author Symbols Method for classifier Accuracy
No and year feature %
extraction
1 Min et al Vowels Mean, SD ELM-R 87+-11.4 %
.(2016) variation,
2 Nguyen et al. Vowels, and "in", Tangent vector in mRVM 80+/-7.3%
(2017) "out", "up" Riemannian
manifold
3 Panachakal et “/iy/”, “/ uw/”,“/ DWT DNN 57.1+/-15 %
al. (2019) piy/”, “/tiy/”, (Discrete wave
“/diy/”, “/m/”, Transform)
“/n/”;
4 Cooney et al. ---d0- MFCC, statistical SVM 22.7+/-5.2 %
(2018) features
5 Saha and Fels Vowels, and "in", Channel cross CNN +RNN 80+/-6.9%
(2019) "out", "up" Coariance + DAE
6 Jahangiri et “/ba/,” “/fo/,” DGT LDA 82.5+/-4.1%
al. (2019) “/le/” and ‘/‘ry/.” (Discrete Gabor
Transform)

42
7 Pawar and “left,” “right” DWT ELM-G 47.9+/-6.9%
Dhage (2020) “Up” and “down” (Discrete Wavelet
Transform)
8 Koizumi et “ue”, “shita”, Spectral Power SVM 81.3+/-0.5%
al.(2018) “hidari”, “migi”,
“mae”, “ushiro”
9 Zhang et al. Mandarin lexical Common spatial SVM 80.1+/-1.2%
(2020) tone patterns
10 Watanabe et Constructed using - NN 38.5+/_5.3%
al. (2020) “/ba?”
11 Chegalayan et 50 CVC words Brain DBN 80.0%
al. (2020) connectivity
estimator and the
entropy measure
12 Jahangiri and “/ba/,” “/fo/”, “/ le DGT LDA 96.4+/-2,3%
Sepulveda /”, and “/ry/” (Discrete Gabor One Vs All
(2019) Transform) classification

2.2.4. The summary of current review and gap:

We can conclude through the literature. No existing system can decode imagined speech
from EEG with sufficient accuracy as a wearable system. The algorithms that provide
accuracy are not enough for all subjects, or it is not tested for the different issue with the
various symbol.
1.Type of EEG acquisition method:
Most of the works are based on 64 channel EEG system. It is an accurate system compare to
a fewer electrodes system. However, due to dense electrodes, it isn't easy to deploy as a
wearable system practically. Thus, there is a trade-off between accuracy and deployment.
For most of the implementation, the sampling rate is sufficient, up to 512 samples. Almost
all the paper decoding method is offline which is a problem for real-time application.
Moreover, most papers' vowels are the symbol for testing. It is showing that the bilingual
symbols give more accuracy in classification.
2. Preprocessing:
The most common preprocessing step in the literature is temporal filtering. Most of the

43
researchers have band-pass filtered the EEG signal in the range of 2 to 50 Hz. In addition, a
notch filter is used by most researchers to remove the powerline hum. If ICA is used, a high
pass filter with a cutoff frequency in the range of 1 to 2 Hz is highly recommended. If the
gamma band is also included in feature extraction, algorithms for removing EMG artifacts
should be used. Saha et al. (2019b) noted that it is better to avoid spatial filtering in the
preprocessing pipeline. Most of the popular ICA algorithms currently available are not suited
for real-time applications; hence, other algorithms like those used by Nguyen et al. (2017)
should be used.
3. Features and classifiers used:
Most of the works that use traditional machine learning techniques such as ANN, ELM,
and SVM extract features from each channel independently. In pieces that use deep-
learning methods, features are usually extracted from channel cross-covariance (CCV)
matrices. The use of CCV matrices is preferred since they better capture the information
transfer between different brain regions. Although researchers in other fields such as
speech recognition and computer vision have almost entirely moved to deep learning,
researchers working on decoding imagined speech from EEG still use conventional
machine learning techniques primarily due to the limitation in the amount of data available
for training the classifiers.
Research Direction :

44
 3. Implementation
Through the literature review of different researcher works, we were able to implement the
system to achieve our goal that is "Decoding Spectrotemporal Features of Imagined
Speech from Electroencephalography (EEG) and Magnetoencephalography (MEG)."
Block diagram desire system will be as shown in figure 3.1. It is a general block diagram of
any biomedical system. But while selecting electrodes and other blocks are real challenges.

Figure 3.1 Block diagram of system overview

For acquiring desire data from the brain scalp, electrode selection is the most critical task.
As per the literature survey, there are six types of electrodes commercial available. Here we
are using spike dry electrode TDE 202 ¼" diameter with 3mm thick spike from the Flrorida
Research Instrument. It is made Ag/AgCl with silicon-based reusable with TDE207
connector cable. Since dry electrode spike conducting gel is required. Skin preparations are
also not required. It can place on the scalp using a simple elastic hairband. At this time, we
are using three electrodes. One electrode is near the inferior frontal gyrus(IFG), i. e. near the
Broca's area other will be near the superior temporal gyrus (STG), secondary auditory
cortex, and one will be at the center of the forehead as a reference electrode as shown in
figure 3.2. B is at IFG and G at STG

45
 Figure 3.2 Position of electrode
The output of the electrode is connected to instrumentation amplifier INA128 from IC from
the Texas Instrument. It is a high CMRR 120dB instrumentation amplifier with low drift and
high input impedance. The INA128 and INA129 are low-power, general-purpose
instrumentation amplifiers offering excellent accuracy. The versatile 3-op amp design and
small size make these amplifiers ideal for a wide range of applications. Current-feedback
input circuitry provides wide bandwidth even at high gain (200 kHz at G =100). A single
external resistor sets any gain from 1 to10000. The INA128 provides an industry-standard
gain equation; the INA129 gain equation is compatible with the AD620. Feature of INA128
as follows
• Low offset voltage: 50 μV maximum
• Low drift: 0.5 μV/°C maximum
• Low Input Bias Current: 5 nA maximum
• High CMR: 120 dB minimum
• Inputs protected to ±40 V
• Wide supply range: ±2.25 V to ±18 V
• Low quiescent current: 700 μA
• Packages: 8-pin plastic DIP, SO-8
We set the gain up to 100 so that the signal has converted to millivolt. The output of the
amplifier has connected to a fourth-order low pass filter with a 2000 Hz higher cutoff
frequency. There is for the filter circuit, the OPA333 operational amplifier. It is a quad op-
amp with excellent features like INA333. The output of LPF is connected 2nd order high pass
filter .5 Hz lower cutoff frequencies to remove DC offset and drift. The production of HPF is
connected to a notch filter with a 50Hz band stop frequency to reduce power line noise. The
notch filter output is a connected amplifier with 1000 gain. This output is connected to PC
with LABVIEW for further processing of signal and analysis and features extraction.

Electrode Note: We developed our electrodes using copper with graphene composite. We
tested the electrode using Electrocardiography (ECG). ECG signal is a powerful signal
compared to EEG, and it is well defined periodic signal.

46
 4. Result and Discussion

Implementing the system using the said setup, we got the following result.

Figure 4.1 EEG Gamma wave f7 and T3

47
Figure4.2 Marathi consonant "JHA" with EEG electrode placed at STG and IFG

Figure4.3 Marathi consonant "JHA" with EEG

Figure4.4 Marathi consonant "JHA" with mike

48
Figure 4.5 Marathi consonant" CHA" with EEG electrode placed at STG and IFG

\
Figure4.6 Marathi consonant "CHA" with Collar mike

49
Fig 4.7 Continuous CHA with EEG Fig 4.8Continuous CHA with mike C

Figure 4.9 ECG Using Copper-based Graphene electrode

Table4.1
Comparison table using EEG and Mike output at overt speech with Marathi consonant.
Sr. Consent Frequency Frequency Time with Time with mike
No. spectra with spectra with EEG
mike EEG
1 CHA 250,459,750 250, 450,853 4ms 3ms

50
2 CHHA 346, 435,835 346,475,800 4ms 4ms
3 JA 256, 735 320, 653 3ms 3ms
4 JHA 290,570 278, 639 3ms 3ms

4.1 Discussion on EEG pattern

They placed electrodes near the STG and IFG, and specific patterns at overt speech can see
GUI using LABVIEW and some rhythmic noise able to here. We can use a similar kind of
pattern when we speak similar consonants on mike and output measure with the same setup.
Output for "CHA," "CHHA," "JA," JHA" was matching with mike for spectrotemporal
features, but it is not the same for another consonant. We have to investigate this is is pattern
due above consonant or it is the artifact. When we spoke similar words with speed, the
signal has repeated the pattern in the same rhythms. When we try the sound through the
speaker for EEG input, we could hear a sound near to said consonant, but the artifacts
destroyed the signal.
4.2 ECG output With Newly developed Graphene-based electrode:
As shown in Figure 4.9, we can see ECG with a single lead near the thorax. The pattern of
ECG confirm with a medical practitioner and compare with real-time ECG. Rhythms are
matching amplitude and pulse rate. We tested electrodes with ECG before using EEG
because ECG has a stronger signal in all senses and is periodic, which we cannot say about
EEG.

51
 5 Conclusion and Future scope

The result which is observed for certain consonants has required to confirm with many
subjects. Also, we have confirmed that it is a pattern or artifact. A similar investigation has
to do with an imagined speech which is our motto. Similarly, the electrode trying to be
developed should be characterized and authenticated with agency and tested with a standard
EEG signal before using it for actual work.

52
 References

1. Bagheri A., Muhammad T. S., Velazquez J. L. Perez, and Roman Genov, (2017),
"Low-Frequency Noise and Offset Rejection in DC-Coupled Neural Amplifiers: A
Review and Digitally-Assisted Design Tutorial," IEEE TRANSACTIONS ON
BIOMEDICAL CIRCUITS AND SYSTEMS, VOL. 11, NO. 1, FEBRUARY
2017.pp 161-176.
2. Barak O., Khandwala N., and Chopra M., 2014 "Classifying Syllables in Imagined
Speech using EEG" Data https: // www. Semanticscholar.org/…/
3. Borna A., Carter T. R., Colombo A. P., Yuan-Yu Jau, Berry C., McKay J., Stephen
J., Weisend M., Peter D. D., and Schwindt, (2017), "A 20-Channel
Magnetoencehalography System Based on Optically Pumped Magnetometers,
"Medical Physics (physics. med-ph); Instrumentation and Detectors (physics. ins-
det) https://arxiv.org/abs/1706.06158v1
4. Changkyun Im and Jong-Mo S., (2016), A Review of Electrodes for the Electrical
Brain Signal Recording, The Korean Society of Medical & Biological
Engineering and Springer 2016, Biomed Eng Lett (2016) 6:104-112, DOI
10.1007/s13534-016-0235-1
5. Christian M., Martin S., Knight R. T., and Pasley B. N., 2016, "Understanding and
Decoding Thoughts in the Human Brain," Frontiers for young minds, Neuroscience
journal, Published: April 27, 2016, doi: 10.3389/frym.2016.00004
6. Deng, S., Srinivasan, R., and D'Zmura, M. (2013). "Cortical signatures of heard and
imagined speech envelopes." CALIFORNIA UNIV IRVINE DEPT OF
COGNITIVE SCIENCES.
7. Derix J., OlgaIljina, JohannaW., Schulze-Bonhage A, AdAertsen and TonioBall (
2014), "From speech to thought: the neuron alba sis of cognitive units in non-
experimental, real-life communication investigated using ECoG," Frontier in
Human Neuroscience, June 13, 2014, doi: 10.3389/fnhum.2014.00383
8. Diego A. R., Leonardo A. G., and Olga L. R.,( 2016), "EEG Signal Analysis Related
To Speech Process Through Bci Device Emotiv, Fft And Statistical Methods,"
ARPN Journal of Engineering and Applied Sciences, VOL. 11, NO. 5, MARCH
2016
9. Douglas O. C. and Andrew C. P.,( 2017), "Magnetoencephalography and Magnetic
Source Imaging," 10.4.1.57 Med, US, Med_US_Papanicolaou, 311016OUS,
PC_MANUSCRIPT, 12 First proofs, Production_Appln, Book. indb 13,
10. Du W.J., Gook H. K., Na Y. K., Zonghoon L., Sang D. J., and Jeong-O.
LeeA(2017), “High-performance transparent graphene/ vertically aligned carbon
nanotubes (VACNT) hybrid electrode for neural interfacing”, RSC Adv., 2017, 7,
3273, DOI: 10.1039/c6ra26836f
11. Duarte M.de A., (2015) "Using Brain-Computer Interface to understand and decode
neural processes associated with imagined speech," @inproceedings
Almeida2015UsingBI,
12. FabienL., Jonathan. B., PeterBrunner, Aysegul G., Anthony L. R., Cuntai G.and
Gerwin S., (2015), "Electrocorticographic representations of segmental features in
continuous speech," Frontier in HUMAN NEUROSCIENCE, February 2015 |
Volume 9 | Article 97, doi: 10.3389/fnhum.2015.00097

53
13. Farzan F, Vernet M, Shafi MMD, Rotenberg A, Daskalakis Z. J. and Pascual-Leone
A (2016)," Characterizing and Modulating Brain Circuitry through Transcranial
Magnetic Stimulation Combined with Electroencephalography." Front. Neural
Circuits 10:73. doi: 10.3389/fncir.2016.00073
14. Fernando L.(2013).' EEG and MEG, Relevance to Neuroscience", Neuron 80,
December 4, 2013, Elsevier Journals 2013.
15. Gaetano G., Paolo B., Rafael A. C., Mario C., Craig J., Alistair M., and André van
S.,(2010), Noninvasive Electronic Biosensor Circuits and Systems Intelligent and
Biosensors, Book edited by: Vernon S. Somerset, ISBN 978-953-7619-58-9, pp.
386, January 2010, INTECH, Croatia, downloaded from SCIYO.COM.
16. Gerald C. I.,(2017)," Brain Localization and the Integrated Systems Hypothesis:
Evidence from Broca's Region," Journal of Behavioral and Brain Science, 2017, 7,
511-519http://www.scirp.org/journal/jbbs ISSN Online: 2160-587
17. Giovanni M. Di., James A. O'Sullivan, and Edmund C. L. (2015), "Low-Frequency
Cortical Entrainment to Speech Reflects Phoneme-Level Processing," Current
Biology 25, 2457–2465 October 5, 2015, ª2015 Elsevier Ltd All rights reserved
http://dx.doi.org/10.1016/j.cub.2015.08.030.
18. Grigorios N.and Lambros M.,(2018), "Neuroanatomy of Language: New Insights
from Lesioned and Healthy Brains," EC Neurology 10.5 (2018): 343-345.
19. Haryong S., Yunjong P., Hyungseup K., and Hyoungho Ko, (2015), "Fully
Integrated Biopotential Acquisition Analog Front-End IC," Sensors 2015, 15,
25139-25156; doi:10.3390/s151025139
20. Hechun X., Wei Huang, Liang Wu, Hui Ma, Xiaodong W. Xuexin C., Shengyu S.
and Xiaoxiong J.(2012)," Preoperative functional MRI localization of language
areas in Chinese patients brain tumors Validation with intraoperative electrocortical
mapping," NEURAL REGENERATION RESEARCH Volume 7, Issue 20, July
2012, Neural Regen Res. 2012;7(20):1563-1569
21. Herff C. and Tanja S.,( 2016), "Automatic Speech Recognition from Neural
Signals: A Focused Review." Front. Neurosci. 10:429. doi:
10.3389/fnins.2016.00429
22. Hsueh-C., Huei-Mei W., and Hung-Jen Y.,(2012), "Using Simulation to Design and
Implement an EEG Pre-amplifier Circuit, Recent Advances in Communications,
Circuits and Technological Innovation," ISBN: 978-1-61804-138-8 pp 241-246
23. Ioannides A. A., 2006," Magnetoencephalography as a Research Tool in
Neuroscience": State of the Art, Magnetoencehalography in Neuroscience,
NEUROSCIENTIST 12(6):524–544, 2006. DOI: 10.1177/1073858406293696
24. Jared M. N., John C. T. and Sharon L. Thompson-Schill (2011), "Broca's Area and
Language Processing: Evidence for the Cognitive Control Connection Language
and Linguistics Compass, "4/10 (2010): 906–924, 10.1111/j.1749-
818x.2010.00244.x,
25. Jiawei Xu, Mitra S., Chris V. H., Refet F. Y., and Kofi Makina. (2017), "Active
Electrodes for Wearable EEG Acquisition: Review and Electronics Design
Methodology." IEEE REVIEWS IN BIOMEDICAL ENGINEERING, VOL. 10,
2017 pp153-164
26. Jobbagy Á. and Varga S, 2013, "Biomedical Instrumentation," Published by
XMLmind XSL-FO Converter pp152-167
27. Kannan R., Syed S. A. A., Abdulrehman F., Syed H. A. and Khan A. (2016),
"Smart Wearable EEG Sensor," IEEE International Symposium on Robotics and
Intelligent Sensors, IRIS 2016, 17-20 December 2016, Tokyo, Japan Procedia
Computer Science 105 ( 2017 )

54
28. Lopez-Gordo M. A., Sanchez-Morillo D. and Pelayo Valle F. (2014); “Dry EEG Electrodes”;
Sensors 2014, 14, 12847-12870; doi:10.3390/s140712847
29. Manca A. D., Giorgio D. N., and Grimaldi M.,( 2016)," EEG-Based Recognition of
Silent and Imagined Vowels," DOI: 10.17469/O2102AISV000019.
30. Marília D. S.(2014), High-sensitivity MR sensors incorporated in silicon needles for magnetic
neuronal response detection; Thesis to obtain the Master of Science Degree in Biomedical
Engineering, December 2014.
31. Mota A.R., Duarte L., Rodrigues D., Martins A.C., Machado A.V., Vaz F., Fiedler
P., Haueisen J., N'obrega J.M., and Fonseca C., (2013), Development of A Quasi-
DRY Electrode for EEG Recording; Sensors and Actuators: A Physical
(2013).http://dx.doi.org/10.1016/j.sna.2013.06.013.
32. Pengfei S., and Jun Q.,(2016), "Neural EEG-Speech Models" arXiv:1612.05369v1 [cs.SD] December
16, 2016
33. Price C. J., Crinion J. T., and MacSweeney M., (2011), A generative model of
speech production in Broca's and Wernicke's areas, published: September 16, 2011,
doi: 10.3389/fpsyg.2011.00237 www.frontiersin.org,
34. Price C. J., (2012), A review and synthesis of the first 20 years of PET and
fMRI studies of heard speech, spoken language and reading, journal
homepage: www.elsevier.com/ locate/lining, NeuroImage 62 (2012) 816–847
35. Puce A. and Matti S. A,(2017), "Review of Issues Related to Data Acquisition and
Analysis in EEG/MEG Studies," Brain sciences Review, Received: April 3, 2017;
Accepted: May 25, 2017; Published: May 31, 2017.
36. Riaz B., Pfeiffer C. & Schneiderman J. F..(2017), "Evaluation of realistic layouts
for next-generation on-scalp MEG: spatial information density maps," SCIENTIFIC
REPORTS | 7: 6974 | DOI:10.1038/s41598-017-07046-6,
37. S Martin, Millán, J. D. R., Knight, R. T., and Pasley, B. N. (2016)." The use of
intracranial recordings to decode human language: challenges and
opportunities." Brain and language.
38. S, Martin, Brunner, P., Holdgraf, C., Heinze, H. J., Crone, N. E., Rieger, J., and
Pasley, B. N. (2014). "Decoding spectrotemporal features of overt and covert
speech from the human cortex." Frontiers in neuroengineering, 7, 14.
39. S.Baillet, Mosher, J. C., & Leahy, R. M. (2001)." Electromagnetic brain
mapping." IEEE Signal processing magazine, 18(6), 14-30.
40. S.Martin, Brunner, P., Iturrate, I., Millán, J. D. R., Schalk, G., Knight, R. T., &
Pasley, B. N. (2016)." Word pair classification during imagined speech using direct
brain recordings." Scientific reports, 6, 25803.
41. Sander, T. H., Preusser, J., Mhaskar, R., Kitching, J., Trahms, L., & Knappe, S.
(2012). "Magnetoencephalography with a chip-scale atomic
magnetometer." Biomedical optics express, 3(5), 981-990
42. Seyedeh-Rezvan F., Henson R. N. and Olaf H.(2018), "Adaptive cortical
parcellations for source reconstructed EEG/MEG connectomes," journal homepage:
www.elsevier.com/locate/neuroimage, NeuroImage 169 (2018) 23–45
43. Stolze, Sarah J (,2016) "Improving Electroencephalography-Based Imagined
Speech Recognition with a Simultaneous Video Data Stream" (2016). Computer
Science and Computer Engineering Undergraduate Honors Theses. 38. HTTP://
scholarworks .uark.Edu/csceuht/38
44. Tanja S., Krusienski D. J. Herff C, and Brumberg J. S.(2017)," Biosignal-Based
Spoken Communication: A Survey," IEEE/ACM Transactions On Audio, Speech,
And Language Processing, Vol. 25, No. 12, December 2017,2257-2272

55
45. Tang A. C., Pearlmutter B. A., Malaszenko N. A., Phung D. B., and Reeb B. C.,
2002, "Independent Components of Magnetoencephalography: Localization,"
Communicated by Terrence Sejnowski.
46. TORTORA G and Derrickson B 2006. Principle of Anatomy and Physiology,
Willy Publication.
47. U., Ha., Lee, Y., Kim, H., Roh, T., Bae, J., Kim, C., and Yoo, H. J. (2015), "A
wearable EEG-HEG-HRV multimodal system with simultaneous monitoring of tES
for mental health management," IEEE transactions on biomedical circuits and
systems, 9(6), 758-766.
48. Ukil A.,(2010) Denoising and Frequency Analysis of Noninvasive
Magnetoencehalography Sensor Signals for Functional Brain Mapping IEEE
transaction in bioscience and engineering Manuscript received September 29, 2010;
accepted November 25, 2010. Date of publication December 3, 2010; date of
current version February 1, 2012.
49. Wang J., Lingzhong F., Yinyan W., Wenting X., Tao J., Peter T. Fox, Simon B.
Eickhoff, C. Yu, and Tianzi J.,(2015), "Determination of the Posterior Boundary of
Wernicke's Area Based on Multimodal Connectivity Profiles," Hum Brain Mapp.
2015 May; 36(5): 1908–1924. doi:10.1002/hbm.22745.
50. Wang K., Wang X. Wang, and Gang Li., (2017), Simulation Experiment of BCI
Based on Imagined Speech EEG Decoding; arXiv preprint arXiv:1705.07771.
51. Wang, J., Kim, M., Hernandez-Mulero, A. W., Heitzman, D., & Ferrari, P. (2017,).
"Towards decoding speech production from single-trial Magnetoencephalography
(MEG) signals." In Acoustics, Speech and Signal Processing (ICASSP), 2017 IEEE
International Conference (pp. 3036-3040). IEEE.
52. Wang, T. Y., Lai, M. R., Twigg, C. M., and Peng, S. Y. 2014), "A fully
reconfigurable low-noise biopotential sensing amplifier with 1.96 noise efficiency
factor." IEEE Transactions on biomedical circuits and systems, 8(3), 411-422.
53. Wang, X Huang, Y., Liu, A. A., Lafon, B., Friedman, D., Dayan, M, and Parra, L.
C. (2017). "Measurements and models of electric fields in the in vivo human brain
during transcranial electric stimulation." Elife, 6, e18834
54. Xu, J. (2016). “Low-Power Active Electrodes for Wearable EEG Acquisition”,
DOI: 10.4233/uuid:0f80a5f1- d81f-4f1d-b6ae-01289f8dfef
55. Y., Liu, L., Wang, H., Liu, Z., Niu, Z., & Denby, B. (2018)." Updating the Silent
Speech Challenge benchmark with deep learning." Speech Communication, 98, 42-
50.
56. Yamaguchi, H., Yamazaki, T., Yamamoto, K., Ueno, S., Yamaguchi, A., Ito, T and
Fukuzumi, S. (2015)," Decoding silent speech in Japanese from single-trial EEGS:
Preliminary results", Journal of Computer Science & Systems Biology, 8(5), 285.

56