MULTIVARIATE BAYESIAN CLASSIFICATION OF TONGUE MOVEMENT EAR
PRESSURE SIGNALS BASED ON THE WAVELET PACKET TRANSFORM

Khondaker A. Mamun1, Michael Mace2, Mark E. Lutmen1, Ravi Vaidyanathan2, Lalit Gupta3 and Shouyan Wang1

1Institute of Sound and Vibration Research (ISVR), University of Southampton, UK, 2Department of Mechanical Engineering,
University of Bristol, UK, 3Department of Electrical and Computer Engineering, Southern Illinois University, USA
Emails: {km, mel}@isvr.soton.ac.uk, {mike.mace, r.vaidyanathan}@bristol.ac.uk, lgupta@siu.edu, sy.wang@soton.ac.uk

ABSTRACT

Tongue movement ear pressure signals have been used to
generate controlling commands in human-machine interfaces.
The objective of this study is to classify the controlled
movement relating to an intended action from interfering signals
that can be experienced. These interfering signals include but are
not limited to, speech, coughing and drinking. Thus data was
collected for six types of controlled movement and the various
interfering signals, when subjects spoke, coughed or drank. The
signal processing involves detection, segmentation, feature
extraction and selection, and classification of tongue motions.
The segmented signals were initially transformed into the
wavelet packet domain, allowing for various features to be
extracted based on statistical properties of the wavelet
coefficients. These are then used as input into a Bayesian
classifier under multivariate Gaussian assumptions. The average
classification performance for identifying controlled movements
and interfering tongue signals achieved 98% and 93.5%
respectively. Thus the classification of tongue movement ear
pressure signals based on the wavelet packet transform is robust.
The application of this Bayesian classification strategy
significantly reduces the interference of controlling commands
when considered within a human-machine interface system
operating in a challenging environment.

Index Terms— Tongue movement ear pressure signals,
wavelet packet transform, Bayesian classifier.

1. INTRODUCTION

A wide range of research has been conducted for developing
various human-machine interfaces (HMI) for hands-free
communication based on human physiological signals to assist
the physically impaired patients [1], [2]. Specifically hands-free
communication and control devices are essential for an
individual who has limited upper body mobility or severe motor
dysfunctions due to spinal cord injury, stroke, congenital limb
deformities or arthritis etc. In spite of the significant progress
made to develop techniques and devices for HMI systems,
current products have not fully addressed patient specific
requirements and better interfaces between the patient and
peripheral devices are still greatly expected. To meet a patients
requirements, a novel hands-free communications and control
concept based on tongue movement ear pressure (TMEP)
signals was previously introduced [1]. With this system
preventing issues associated with hygiene and intrusion when
donning the device. Users express their intention by making
impulsive actions of the tongue, which create unique acoustic
pressure signals within the auditory region. The pressure signals
can be easily recorded using a microphone earpiece positioned
non-invasively within the ear canal. The tongue is utilized due to
it’s inherent capability of fine motor control and manipulation
tasks involving many degrees of freedom [2]. Therefore it has
evolved to perform sophisticated motions during speech and
mastication. Individuals who have limited control of their limbs
can use tongue movements to communicate with computers and
control electromechanical or assistive devices through the
sensing of acoustic pressure signals.
Previously a decision fusion classification algorithm was
implemented for differentiating four tongue movements (up,
down, left and right) relating to controlled TMEP signals. This
achieved an average of 97% correct classification accuracy
when utilized in conjunction with a large training set. The
performance of this classifier was also compared with three
different pattern classification strategies, the matched filter
(86%), the Gaussian classifier using autoregressive (AR) model
parameters (85.98%) and the nonlinear alignment classifier
(96.27%) using time domain information [1]. The TMEP data
sets were collected in a controlled environment. Apart from this,
to de-noise the TMEP signal (i.e. the removal of random noise),
an algorithm based on discrete wavelet thresholding was
implemented [3]. There are several challenges when processing
a TMEP signal captured from a real environment. One of the
potential challenges is that only a limited number of signals are
available to train the classifier, while conversely a classification
algorithm with high robustness is necessary to work in noisy
environment. It is also assumed that external interference will
merge with the TMEP signals in real-life such as signals
acquired during conversation or from the street. Pilot analysis
was performed for classifying TMEP movement signals in said
situations [4]. The classification algorithm highly depends on
optimal selection of the signal features in both clean and noisy
environments. To improve the classification in presence of