A User-centric Design of Permanent Magnetic Articulography based
Assistive Speech Technology
Lam A. Cheah
1
, Jie Bai
1
, Jose A. Gonzalez
2
, Stephen R. Ell
3
, James M. Gilbert
1
,
Roger K. Moore
2
, and Phil D. Green
2
1
School of Engineering, The University of Hull, Kingston upon Hull, U.K.
2
Department of Computer Science, The University of Sheffield, Sheffield, U.K.
3
Hull and East Yorkshire Hospitals Trust, Castle Hill Hospital, Cottingham, U.K.
Keywords: Silent Speech Interface, Assistive Speech Technology, Permanent Magnetic Articulography, User-centric
Design.
Abstract: This paper addresses the design considerations and challenges faced in developing a wearable silent speech
interface (SSI) based on Permanent Magnetic Articulography (PMA). To improve its usability, a new
prototype was developed with the involvement of end users in the design process. Hence, desirable features
such as appearance, portability, ease of use and light weight were incorporated into the prototype. The
device showed a comparable performance with its predecessor, but has a much improved appearance,
portability and hardware in terms of miniaturisation and cost.
1 INTRODUCTION
The ability to communicate through speech is crucial
to humanity and plays a vital role in our social and
work life. Patients whose voice box has to be
removed because of throat cancer, trauma,
destructive throat infections or neurological
problems will inevitably lose their ability to speak.
Hence, they may experience severe impact on their
lives which can lead to social isolation and
depression (Fagan et al., 2008). However,
conventional speech restoration methods (e.g.
oesophageal speech, the electrolarynx and speech
valves) have limitations in terms of quality of speech
and usability (Fagan et al., 2008; Gilbert et al.,
2010). Furthermore, in the case of implanted speech
valves, frequent valve replacement is required within
a time span of 3-4 months, because of the growth of
biofilm coating over time (Ell et al., 1995, 1996;
Heaton and Parker, 1994).
In order to overcome these limitations, a radical
alternative approach has been introduced: silent
speech interfaces (SSIs). SSIs are devices that allow
speech communication in the absence of audible
acoustic signals. Besides their usage as
communication aid for post-laryngectomy patients,
SSIs can also be deployed in a quiet, noisy or
acoustically challenging environment (Denby et al.,
2010). Although SSIs are still in their experimental
stage, there were encouraging reports over recent
years. To date, there are several types of SSIs using
different modalities (Denby et al., 2010). Permanent
Magnetic Articulography (PMA) is a measuring
technique that captures the magnetic field variations
from a set of permanent magnets attached to the
articulators (i.e. lips and tongue) during speech
(Gilbert et al,. 2010). This does not provide explicit
information regarding the position of the attached
magnets. Rather, the measured PMA data is the
combination of the magnetic field patterns
associated to a particular articulatory gesture.
Despite the attractive attributes of SSIs, there are
still challenges in the form of processing software
(e.g. efficiency, robustness and reliable speech
generation) and hardware (e.g. portability, light
weight, unobtrusiveness and wearability).
Preliminary discussion on the influential factors of
the SSIs’ implementation had been reported by
Denby et al. (Denby et al., 2010), based upon criteria
such as ability to operate in silence and noisy
environments, usability by laryngectomee, issue of
invasiveness, market readiness and cost.
The main focus of this paper is to address the
hardware challenges facing the PMA-based SSI
system (as opposed to the speech processing
109
A. Cheah L., Bai J., A. Gonzalez J., R. Ell S., M. Gilbert J., K. Moore R. and D. Green P..
A User-centric Design of Permanent Magnetic Articulography based Assistive Speech Technology.
DOI: 10.5220/0005354601090116
In Proceedings of the International Conference on Bio-inspired Systems and Signal Processing (BIOSIGNALS-2015), pages 109-116
ISBN: 978-989-758-069-7
Copyright
c
2015 SCITEPRESS (Science and Technology Publications, Lda.)
challenges which are addressed elsewhere). A
number of significant steps have been taken in order
to develop a wearable system that is appropriate for
everyday use. A novel embodiment comprising
miniaturised sensing modules and wireless headset
that is compact and comfortable is proposed.
The remainder of this paper is structured as
follow. The next section briefly overviews the PMA
technique and its development to date. Section 3
describes the system architectural design and the
challenges, followed by the performance evaluation
in section 4. The final section concludes and
provides an outlook for future research.
2 RELATED WORK
The Magnetic Voice Output Communication Aid
(MVOCA) is a PMA-based device developed within
DiSArM (Digital Speech Recovery from Articulator
Movement, www.hull.ac.uk/speech/disarm) project,
aiming to restore speech communication ability for
patients who have undergone surgical removal of the
larynx. In principal, the MVOCA consists of an
array of magnetic sensors mounted onto a
lightweight headset for detection, a set of permanent
magnets, four on the lips (ø1mm x 5mm), one at
tongue tip (ø2mm x 4mm) and one at tongue blade
(ø5mm x 1mm) as illustrated in figure 1.
Information on magnets placement was described in
(Gilbert et al., 2010). The magnets are temporarily
attached using Histoacryl surgical tissue adhesive
(Braun, Melsungen, Germany). For long term usage,
these magnets will be surgically implanted. A data
acquisition/control unit is used to condition the
acquired measurements before transmitting them to
a computer where appropriate processing and
recognition algorithms are then applied.
In previous publications (Gilbert et al., 2010;
Hofe et al., 2013b), the validity of the performance
on isolated-word and connected digits recognition
tasks using the PMA technology were shown. This
was then followed by investigation into the
performance across multiple speakers (Hofe et al.,
2013a). A feasibility study of direct speech synthesis
bypassing the intermediate recognition step was
reported in (Hofe et al., 2011). More recently,
extensive investigation into effectiveness of PMA
data in terms discriminating the voicing, place and
manner of articulation of English phones was
presented in (Gonzalez et al., 2014).
So far, the tests speech experiments were carried
out using the 1
st
generation MVOCA, which
consisted of five tri-axial Honeywell HMC2003
magnetic sensors, mounted on a pair of safety
glasses, as shown in figure 2. The fluctuation in
magnetic field is captured on the 15 PMA channels
and recorded onto a PC via ADLink DAQ-2206
analogue-to-digital converter (ADC), a PCI-based
card with 16-bit linear encoding. The 2
nd
generation
MVOCA was developed and was first used in our
recent study (Gonzalez et al., 2014). Detailed
description of the latest hardware will be discussed
in the next section.
Figure 1: Placement of six magnet pellets onto lips and
tongue. The magnets are cylinders with diameter and
length of 1mm x 5mm for lips (pellets 1-4), 2mm x 4mm
for tongue tip (pellet 5) and 5mm x 1mm for tongue blade
(pellet 6).
Figure 2: MVOCA headset (1
st
generation) - five magnetic
sensors mounted on a frame that attached onto a pair of
safety glasses. Appearance of device when worn by user.
3 SYSTEM ARCHITECTURE
3.1 Design Consideration
Despite encouraging performance, the earlier
MVOCAs (Gilbert et al., 2010; Hofe et al., 2013a,
2013b) were not satisfactory particularly in its
appearance, comfort and ergonomic factors for the
user. Hence, to make the MVOCA more usable and
desirable, the prototype has undergone several
BIOSIGNALS2015-InternationalConferenceonBio-inspiredSystemsandSignalProcessing
110
iterative design cycles over the past 12 months. The
main focus during the development phase was to
consider underexplore user-inclusive requirements
by using a qualitative methodology, including
informal opinion survey, focus group and user
observation. These approaches were commonly used
in other user-centred design studies (Bright and
Conventry, 2013; Hirsch et al., 2000; Martin et al.,
2006).
Based on the discussion with the panels (i.e.
laryngectomees) of the focus group and data from
the survey questionnaires of 50 potentials users and
their families/friends, the appearance of the device is
the major factor affecting the acceptability. In fact,
other researches also indicated that the appearance is
considered a highly desirable feature for any
assistive devices (Cook and Polgar, 2008; Hirsch et
al., 2000). Six possible configurations were
presented in the survey, those resembling a
Bluetooth earpiece or a pair of spectacles were
preferred with the majority of the votes, while the
device resembling a headset microphone was
marginally acceptable by approximately 25% of the
respondents. On the other hand, devices that might
be obstructive to the mouth in anyway (in full or
partially, such as the 1
st
generation MVOCA as
illustrated in figure 2) were deemed unacceptable.
In addition, from previous focus group meetings
and observation studies (participants had given their
consent and the studies were approved by The
University of Hull ethics committee), valuable
feedback was gained and has greatly influenced the
creation of a user-centric design prototype. During
the prototype development, critical design questions
were raised, in term of headset appearance,
portability, weight, ease of use and cost.
Besides hardware appearance, the survey
questionnaires also identified other desirable
features, such as processing features (see table 1)
and speech quality (see table 2), by their preferred
ranking. As indicated in table 1, the quality of
reconstructed speech is highly desirable, whereas the
issue of delay between reconstructed sound and lips
movement is least prioritised. In term of speech
quality, this was further subdivided into the
characteristics listed in table 2. Both intelligibility
and naturalness of speech are considered equally
desirable, but the ability to convey emotion into the
reconstructed speech is least preferred. It should be
noted that respondents to the survey may have had
some difficulty interpreting the meaning of some of
these terms since, for instance, they may not be
aware of the extent of emotion present in normal
speech. These non-hardware related features will not
be discussed in this paper but will be addressed
separately in our future work.
3.2 New MVOCA Prototype
Major hardware components of the latest MVOCA
prototype are shown in figure 3. A set of four tri-
axial Anisotropic Magnetoresistive (AMR) magnetic
sensors (Honeywell HMC5883L), a control unit and
a power source (rechargeable 7.4V Lithium Ion
battery) were mounted on a customised headset.
Two headsets design were developed, 1) attached
onto a headband (see figure 4a), and 2) onto a pair of
spectacles (see figure 4b). The headsets (excluding
Table 1: Desirable processing features.
Software feature Description Ranking
Speech quality Measuring the quality of reconstructed speech (see also Table 2) 1
st
Speech mode Ability to communicate in fluent speech (ranging from isolated words to
fluent speech)
2
n
d
Vocabulary Size and range of words available in the database (ranging from a small
context specific vocabulary to unrestricted vocabulary)
3
r
d
Speaking delay Synchronisation between lips movement and synthesised voice (ranging
from speaking a complete phrase before any speech output to no delay)
4
t
h
Table 2: Desirable speech quality.
Speech quality Description Ranking
Intelligibility Ability to communicate intelligibly (i.e. ranging from barely intelligible
to a BBC newsreader)
1
st
Naturalness Ranging from a monotonic electronic voice to natural speech 2
n
d
Personification The choice of using own or preferred voice (ranging from another
appropriate voice to the user’s own voice)
3
r
d
Ability to convey emotion Ability to include emotions (ranging from no emotional content to full
emotion content)
4
t
h
AUser-centricDesignofPermanentMagneticArticulographybasedAssistiveSpeechTechnology
111
the pair of spectacles or headband) were fabricated
using rapid prototyping technology and their
building material were VeroWhitePlus RGD835 and
VeroBlue RGD840. A set of six Neodymium Iron
Boron (NdFeB) permanent magnets are attached
onto the lips and tongue as illustrated in figure 1.
Each magnetic sensor has three orthogonal
sensing elements to measure the three spatial
components of the magnetic field. Sensor1-3 (9
PMA channels) are used to capture magnetic field
variations caused by articulatory movements and
digitize it with 12-bit resolution. Whereas the
sensor4 is used for background cancellation, which
is for compensating the effect of earth’s magnetic
field thus enhancing the signal-to-noise (SNR) of the
desire signals.
Figure 5 shows an operational block diagram of
the 2
nd
generation MVOCA. Each magnetic sensor
communicates to a low-power ATmega328P
microcontroller (housed inside the control unit)
through an I
2
C interface, samples were acquired at
100 Hz. These samples (total of 12 PMA channels)
are then transmitted to computer/tablet PC
wirelessly via Bluetooth or USB for further
processing. A bespoke graphical user interface
(GUI) had been developed in the MATLAB
environment and used mainly for on-line recognition
testing or demonstration purposes. All necessary
speech processing and recognition algorithms were
embedded into the GUI and running in the
background. If the acquired PMA signal correctly
matched an articulation gesture from the pre-stored
training dataset, thus the corresponded utterance will
be identified. A text-to-speech synthesiser is used to
generate a playback audio as an output for the
identified utterance, via an audio device (e.g.
integrated speaker of a computer).
Figure 3: Overview of the 2
nd
generation MVOCA system,
a) MVOCA headset with b) sensor modules, c) control
unit and battery.
Figure 4: Two MVOCA headset designs were presented,
a) mounted onto headband and b) attached onto a pair
glasses. Appearances of device when worn by user.
For wireless data transmission, a class 2
Bluetooth module BTM411 (housed inside the
control unit) and USB transceiver (attached on
computer) are used. The MVOCA device will
Figure 5: Simplified MVOCA (2
nd
generation) operational block. diagram.
BIOSIGNALS2015-InternationalConferenceonBio-inspiredSystemsandSignalProcessing
112
acquire its power from a battery rather than from the
computer via USB (wired mode). The average
power consumption of the current MVOCA
prototype from a 5V (regulated from 7.4V) supply is
~104 mA, which means that it can run continuously
for ~10 hours on a full charge (total 1080mAh). The
battery can be removed from the headset for
charging using a freestanding charger.
4 PERFORMANCE EVALUATION
4.1 Experimental Design
The data used in evaluating the latest MVOCA
prototype were collected from a male native English
speaker who is proficient in the usage of the
MVOCA interface. As clarified previously (Hofe et
al., 2013b), the MVOCA is a speaker-dependant
device, i.e. all associated headset measurements and
training parameters were calibrated towards
particular individual. Although inter-speaker
performance has proven possible (Hofe et al.,
2013a), the headset design and measurements would
require individually tailored for optimal
performance. In particular, the headset was
specifically designed according to the speaker
anatomy.
To evaluate the performance of the new
MVOCA prototype, a continuous speech recognition
task consisting in the identification of sequences of
English digits is chosen in this paper. This task was
chosen because the limited size of vocabulary
enables whole-word model training from relatively
sparse data and also because of the simplicity of the
language model involved. The algorithm used to
generate the random digits sequences was the one
underlying the TIDigits database of connected digits
(Leonard, 1984). The longest digit sequence consists
of seven individual digits. During the training, both
zero and oh (the two representations of 0) were
denoted as separate items.
The experimental data were collected from six
independent training sessions, i.e. two sessions each
one using different 2
nd
generation MVOCA headsets
(figure 4a and 4b) and the remained two sessions
using the 1
st
generation headset (see figure 3). Each
training session consisted of five sets of data, with 4
sets (three spoken data and one mouthed data) used
as training and remained mouthed data for testing.
The idea is to mimic a scenario where the
laryngectomy patient, where his/her has intact voice
and generate training data before the laryngectomy
operation. A total of 385 utterances (5 subsets)
containing 1265 individual digits were recorded
during each training session.
4.2 Instrumental Setup
For optimal recording performance, the experiments
were conducted inside a sound-proof room, where
the audio signal was recorded with a shock-mounted
AKG C1000S condenser microphone and a
dedicated USB sound card (Lexicon Lambda). A
Matlab-based GUI was created to simultaneously
record the audio signal (sampled at 48 kHz) and
PMA data (sampled at 100 Hz). Since both data
streams were measured from separate modality,
synchronization between the two data streams was
necessary to compensate for any small deviation
from the ideal sampling frequencies of the analog-
to-digital converters (ADC). An automatic timing
alignment mechanism was used to realign both data
streams by generating start-stop markers in addition
to both audio and PMA data streams, in order to
minimised any potential timing error.
The GUI software also provided visual prompt of
the digit sequences to the speaker at regular interval
of 5 seconds during the recording session. The
measured PMA data were transferred to a PC via
USB connection. Since the speaker’s head was not
restrained, large movements could potentially distort
the recorded data and thus degrade the recognition
performance. Hence, background cancellation was
applied to compensate for any movement induced
interference against the desired PMA signals.
4.3 HMM Training and Recognition
PMA data used for speech recognition was first low-
pass filtered (i.e. removal 50 Hz noise) and
normalized as described in (Hofe et al., 2013b).
Next, the first-order time derivatives (i.e. delta
parameters) were computed for each PMA frame,
resulting in a feature vector of size 18. The second-
order derivatives (i.e. delta-delta parameters) were
not included as part of the feature vector since, as
shown in our previous works (Hofe et al., 2013a,
2013b), they did not produced significant
improvement in performance.
Then, the processed PMA data were used for
training the speech recognizer using the HTK
(Young et al., 2009). The acoustic model in the
recognizer uses whole-word Hidden Markov Models
(HMMs) with 25 states and 5 Gaussians per state
(Hofe et al., 2013b). By no means were these
parameters optimal, but the suggested parameters
settings were known for their performances from our
AUser-centricDesignofPermanentMagneticArticulographybasedAssistiveSpeechTechnology
113
previous works (Hofe et al., 2013a, 2013b). For
clarification, audio signals were not used to train the
recognizer, but only the PMA data.
4.4 Recognition Performance between
1
St
and 2
Nd
Generation MVOCAs
Word and sequence accuracy results across multiple
MVOCA devices are presented in figure 6 and 7.
The light bars relates to the condition in which only
the static PMA data is used (i.e. dimension of the
feature vector is 9), whereas the darker bars refers to
using both static and dynamic features (vector size
of 18). We will refer to these two conditions as
Sensor and SensorD features, respectively. The
results reflect the mean of the data collected on two
independent training sessions on each of the 1
st
and
2
nd
generation MVOCA devices.
These data were analysed independently session-
by-session, and an averaged across the sessions was
produced. Merging all the data from different
session for recognition would seem a more attractive
approach, but this might lead to inconsistent
outcomes as very precise repetitive magnets
placement are required on each training session.
Nonetheless this could be overcome, as the magnets
will be surgically implanted in the final MVOCA for
long term usage. Investigations into session-
independent approach on other SSIs technique were
presented in (Maier-Hein et al., 2005) and (Wand
and Schultz 2011).
As seen in both figure 6 and 7, it is clear that
SensorD performed significantly better than using
Sensor data alone. Similar trends were also reported
in (Hofe et al., 2013b). Moreover, the results showed
a comparable performance between the 1
st
and 2
nd
generation MVOCAs and in some cases the newer
MVOCA performed slightly better. Hence, this
suggests that the newer MVOCA can have better
hardware features (i.e. appearance, light weight and
mobility) but without compromising its recognition
performance by using miniaturised components (i.e.
sensors and data acquisition unit).
Figure 8 illustrates that the inclusion of mouthed
data in the training dataset improves the recognition
accuracy, particularly in terms of sequence
recognition. In this figure, the light bars relate to
mixed training data (spoken and mouthed data) and
the darker ones to non-mixed training data (spoken
only data). The results presented in figure 8 were
trained and tested using only SensorD data from the
2
nd
generation MVOCA devices, as they provided
better performance as illustrated in figure 6 and 7.
Although further investigation is needed, we
recognised the importance of mixing both spoken
and mouthed data in any training session.
Figure 6: Comparison of word accuracy of connected
digits between 1
st
and 2
nd
generation MVOCAs.
Figure 7: Comparison of sequence accuracy of connected
digits between 1
st
and 2
nd
generation MVOCAs.
Figure 8: Comparison of training dataset (mixed or non-
mixed data) in connected digits recognition rate.
4.5 Hardware Comparison between 1
St
and 2
Nd
Generation MVOCAs
The greatest challenge here is to satisfy the design
objective to improve the MVOCA’s appearance,
BIOSIGNALS2015-InternationalConferenceonBio-inspiredSystemsandSignalProcessing
114
without compromising the device’s performance. A
summary of the key features of the latest MVOCA
system is presented in table 3.
Table 3: Hardware specifications of the MVOCA.
Specification Parameter
Sensor Module
Type Anisotropic Magnetoresistive
Dimension 12 x 12 x 3 mm
3
Sensitivity 440 LSb/gauss
Sampling rate 100 Hz/sensor
No. channels 12 (3 per sensor)
Control Unit
Microcontroller Low power ATmega328P
Dimension 50 x 60 x 15 mm
3
Operating voltage 5 V
Power source Lithium Ion battery
Headset
Material VeroBlue/VeroWhitePlus resin
Total weight 160g (including battery &
control unit)
Two versions of MVOCA headsets were
designed (see figure 4a and 4b), both headsets aim to
provide the desirable features such as light weight,
comfort and fashionable appearance as suggested by
the survey questionnaires. The current designs
significantly reduced the unattractive appearance of
the previous headset (see figure 2), this would
improve the acceptability by the end user and
ultimately improves its usage.
Significant improvements were also made in
term of the hardware miniaturisations and
portability. The previous MVOCA relied on a PCI-
based data acquisition card, thus restricted it to a
desktop PC/workstation which is highly immobile
and bulky. Although the magnetic sensors
HMC2003 are high precision sensors, they are
significant larger in size (24x45x10 mm
3
) and
required higher operation voltage (i.e. 12V), thus
making them non-power efficient.
In the current prototype, magnetic sensors
HMC5883L were chosen because of their
compactness, low operation voltage, low cost and
wide sensitivity range. As for signal conditioning,
low-powered microcontrollers were used. By
utilising a Bluetooth modules and a tablet PC (i.e.
mobile processing unit), the current MVOCA will be
highly portable and practical for everyday use. In
addition, the cost of the prototyping is relatively
low, as the MVOCA only utilised commercial off-
the-shelf (COTS) components. Moreover, by
shrinking the size of electronics, this inevitably
reduced the overall weight of the headset, and
making it more appealing as a wearable assistive
speech technology.
The downside of the new MVOCA prototype
would be the omission of higher precision
components (i.e. magnetic sensors) used in previous
prototype, reduction in numbers of sensors and the
use of a lower sampling rate. However, from the
results presented in figure 6 and 7, these would
suggest the concerns would be irrelevant as the
performances are comparable and slightly better in
some cases when using the 2
nd
generation MVOCA.
There could be a couple of good explanations for
that. Firstly, the articulation movements during
speech is slow and therefore a lower sampling rate
(i.e. 100 Hz) might be sufficient. Secondly,
reduction in number of sensors was possible because
there were excess of information available from
previous MVOCA, thus some sensors can be made
redundant.
5 CONCLUSIONS
The preliminary evaluation of the new MVOCA
prototype shows comparable recognition
performances to the previous system, but providing
much desirable hardware improvements such as
portability, hardware miniaturisation, desirable
appearance and lower cost. Nevertheless, there are
still many challenges ahead before MVOCA can be
practically operated outside laboratory environments
on a day-to-day basis.
Encouraged by the results obtained so far,
extensive work is needed to create a viable wearable
assistive communication aid. Therefore, future
works include enhancing overall MVOCA
appearance, reducing power consumption and
implementing real-time features (e.g. reducing
processing time delay and utilising on-line
endpointing/segmentation). In addition, investigation
into the work on speech synthesis from PMA data
had started and preliminary results were very
encouraging.
ACKNOWLEDGEMENTS
The report is an independent research funded by the
National Institute for Health Research (NIHR)’s
Invention for Innovation Programme. The views
stated are those of the authors and not necessary
reflecting the thoughts of the sponsor.
AUser-centricDesignofPermanentMagneticArticulographybasedAssistiveSpeechTechnology
115
REFERENCES
Bright, A. K., and Conventry, L., 2013. Assistive
technology for older adults: psychological and socio-
emotional design requirements, Proceedings of the 6
th
International Conference on PErvaesive Technologies
Related to Assistive Environments, pages 1-4, Rhodes,
Greece.
Cook, A. M., and Polgar, J. M., 2008. Assistive
Technologies: Principles and Practice, Mosby-Year
Book, 3
rd
edition.
Denby, B., Schultz, T., Honda, K., Hueber, T., Gilbert, J.
M., and Brumberg, J. S., 2010. Silent speech
interfaces. Speech Communication, 52(4):270-287.
Ell, S. R., Mitchell, A. J., and Parker, A. J., 1995.
Microbial colonisation of the Groningen speaking
valve and its relationship to valve failure. Clinical
Otolaryngol, 20:555-556.
Ell, S. R., Mitchell, A. J., Clegg, R. T., and Parker, A. J.,
1996. Candida: the cancer of silastic. Journal of
Laryngol Otol, 110(3):240-242.
Fagan, M. J., Ell, S. R., Gilbert, J. M., Sarrazin, E., and
Chapman, P. M., 2008. Development of a (silent)
speech recognition system for patients following
laryngectomy. Medical Engineering & Physics,
30(4):419-425.
Gilbert, J. M., Rybchenko, S. I., Hofe, R., Ell, S. R.,
Fagan, M. J., Moore, R. K. and Green, P. D., 2010.
Isolated word recognition of silent speech using
magnetic implants and sensors. Medical Engineering
& Physics, 32(10):1189-1197.
Gonzalez, J. A., Cheah, L. A., Bai, J., Ell, S. R., Gilbert, J.
M., Moore, R. K., and Green, P. D., 2014. Analysis of
phonetic similarity in a silent speech interface based
on permanent magnetic articulography. Proceedings of
15
th
INTERSPEECH, pages 1018-1022, Singapore.
Heaton, J. M., and Parker, A. J., 1994. Indwelling trachea-
oesophageal voice prostheses post-laryngectomy in
Sheffield, UK: a 6-year review. Acta Oto-Laryngol,
114:675-678.
Hirsch, T., Forlizzi, J., Goetz, J., Stoback, J., and Kurtx,
C., 2000. The ELDer project: Social and emotional
factors in the design of eldercare technologies.
Proceedings on the 2000 conference of Universal
Usability, pages 72-79, Arlington, USA.
Hofe, R., Bai, J., Cheah, L. A., Ell, S. R., Gilbert, J. M.,
Moore, R. K., and Green, P. D., 2013a. Performance
of the MVOCA silent speech interface across multiple
speakers. Proceedings of 14
th
INTERSPEECH, pages
1140-1143, Lyon, France.
Hofe, R., Ell, S. R., Fagan, M. J., Gilbert, J. M., Green, P.
D., Moore, R. K., and Rybchenko, S. I., 2013b. Small-
vocabulary speech recognition using silent speech
interface based on magnetic sensing. Speech
Communication, 55(1):22-32.
Hofe, R., Ell, S. R., Fagan, M. J., Gilbert, J. M., Green, P.
D., Moore, R. K., and Rybchenko, S. I., 2011. Speech
synthesis parameters generation for the assistive silent
speech interface MVOCA. Proceedings of 12
th
INTERSPEECH, pages 3009-3012, Florence, Italy.
Leonard, R. G., 1984. A database for speaker-independent
digit recognition. International Conference on
Acoustic, Speech and Signal Processing, pages 328-
331, San Diego, USA.
Maier-Hein, L., Metze, F., Schultz, T., and Waibel, A.,
2005. Session independent non-audible speech
recognition using surface electromyography.
Automatic Speech Recognition and Understanding
Workshop, pages 331-336, Cancun, Mexico.
Martin, J. L., Murphy, E., Crowe, J. A., and Norris, B. J.,
2006. Capturing user requirements in medical devices
development: the role of ergonomics, Physiological
Measurement, 27(8):R49-R62.
Wand, M., and Schultz, T., 2011. Session-independent
EMG-based speech recognition. 4
th
International
Conference on Bio-inspired Systems and Signal
Processing, pages 295-300, Rome, Italy.
Young, S., Everman, G., Gales, M., Hain, T., Kershaw, D.,
Liu, X., Moore, G., Odell, J., Ollason, D., Povery, D.,
Valtchev, V., and Woodland, P., 2009. The HTK Book
(for HTK Version 3.4).
BIOSIGNALS2015-InternationalConferenceonBio-inspiredSystemsandSignalProcessing
116