Advances in Information Science and Applications - Volume II

Endoscopic Procedures Control Using Speech

Recognition
Simão Afonso, Isabel Laranjo, Joel Braga, Victor Alves, José Neves

be accomplished in real-time because the gastroenterologist

Abstract — In this paper it is presented a solution for replacing needs to press an additional programmable button on the
the current endoscopic exams control mechanisms. This kind of endoscope to freeze the image and then press the pedal to
exams require the gastroenterologist to perform a complex capture and save the frame [3]. This approach to the
procedure, using both hands simultaneously, to manipulate the problem is not optimal and raises several new issues, such
endoscope’s buttons and using the foot to press a pedal in order to
perform simple tasks such as capturing frames. The last procedure as limiting the movements of everyone involved and
cannot be accomplished in real-time because the gastroenterologist requiring the gastroenterologist to perform a complex
needs to press an additional programmable button on the procedure, distracting him/her from the task at hand. A new
endoscope to freeze the image and then press the pedal to capture hands-free interface that allows for a richer control scheme
and save the frame. The presented solution replaces the pedal with would solve some of the existing snags.
a hands-free voice control module and it is capable of running on A novel approach to this problem consists of adding a
the background continuously without human physical intervention.
This system was designed to be used seamlessly with the
voice recognition module to the system, providing a hands-
MyEndoscopy system that is being tested in some healthcare free control. This module, called MIVcontrol, will be
institution and uses the PocketSphinx libraries to perform real-time integrated into the device called MIVbox (more details are
recognition of a small vocabulary in two different languages, given in section 3).
namely English and Portuguese. The main goal of the MIVcontrol module is to create a
simple speech recognition system for recognizing a very
Keywords—Automatic Speech Recognition, Hidden Markov small vocabulary of simple pre-determined commands. The
Models, PocketSphinx, SphinxTrain, Endoscopic Procedures
recognized commands are used to control the
MIVacquisition, creating a hands-free control system that
I. INTRODUCTION
should be able to replace the current solution. This system

N OWADAYS it is accepted by most healthcare

professionals that information technologies and
informatics are crucial tools to enable a better healthcare
can perform frame capturing in real-time, without the need
to use any extra buttons.
The system should be speaker-independent and have a
practice. The Pew Health Professions Commission (PHPC) very low error rate, even on noisy environments, and it
recommended that all healthcare professionals should be should be able to capture audio from a microphone
able to use information technologies [1]. The technological continuously, so that it can run in the background without
evolution has led to an enormous increase in the production human intervention. This will require automatic word
of objective diagnostic tests and a decrease on the reliance segmentation, to make recognition possible.
of more subjective problem solving methods, which should The rest of the paper is organized as follows: in section 2
increase the quality of the service provided, and can even be it is presented a review of related work in the area of speech
seen as a consequence of the increased accountability of recognition, from its theoretical foundations to practical
healthcare institutions in relation to the legislation [2]. systems already being used. In section 3 is outlined the
EsophagoGastroDuodenoscopy (EGD) and Colonoscopy overall system architecture and how it integrates with the
occupy relevant positions amongst diagnostic tests, since MyEndoscopy system. In section 4 is presented specific
they combine low cost and good medical results. The details about the implementation of the solution, whereas in
current endoscopic exams require the gastroenterologist to section 5 the methodology used in the study is exhibited.
perform a complex procedure using both hands Finally, the results and their assessment are presented in
simultaneously to manipulate the endoscope’s buttons and section 6 and 7, followed by conclusions in section 8.
using the foot to press the pedal in order to perform such
simple tasks as capturing frames. The last procedure cannot
II. RELATED WORK
This work is funded by ERDF - European Regional Development Fund
through the COMPETE Programme (operational programme for
Automatic Speech Recognition (ASR) is a process by
competitiveness) and by National Funds through the FCT - Fundação para which a computer processes human speech, creating a
a Ciência e a Tecnologia (Portuguese Foundation for Science and textual representation of the spoken words. This process has
Technology) within project PEst-OE/EEI/UI0752/2014. two main areas of study, i.e. discrete speech and continuous
S. Afonso, I. Laranjo, J. Braga and J. Neves are in the Computer Science
and Technology Center (CCTC), University of Minho, Braga, Portugal speech. Discrete speech is useful for the creation of voice
([email protected], [email protected], [email protected], command interfaces, while continuous speech, also known
[email protected]) as dictation, mimics the way two humans communicate.
V. Alves is in the Computer Science and Technology Center (CCTC),
University of Minho, Braga, Portugal (corresponding author to provide e- Though the ultimate objective of having a system capable of
mail: [email protected]).

ISBN: 978-1-61804-237-8 404

 Advances in Information Science and Applications - Volume II

recognizing everything anyone can say in multiple applications [9]. Vertanen [10] tested both the HTK and the
languages has yet to be achieved, research has been focused Sphinx systems with the Wall Street Journal (WSJ) corpus
on smaller-scale approaches [4]. and found no significant differences in error rate and speed.
This conclusion is corroborated by other researchers [11].
A. Theoretical Foundations
Huggins-Daines et al. [12] optimized CMU Sphinx II for
Aymen et al. [5] presented the theoretical foundation of embedded systems, primarily those with ARM architecture.
Hidden Markov Models (HMM) that underpin most modern To balance the loss of precision required in other
implementations of automatic speech recognition. The optimizations, the CMU Sphinx III Gaussian mixture model
authors present the distinction between speech recognition, was back-ported. They managed to have a 1000-word
which aims to recognize almost anyone’s speech, and voice vocabulary running at 0.87 times real-time on a 206 MHz
recognition, which creates systems trained to particular embedded device, with an error rate of 13.95% [12]. “Times
users. The model is constructed based on a large corpus of real-time” is a notation that indicates the amount of time
recorded speech, annotated with the respective transcription. required to process live data. In this case, the system can
The HMM requires three different sub-models: process 1 second of data in 0.87 seconds, which makes it
1) The acoustic model consists of different features for suitable to real-time recognizing. This work has lead to the
each utterance the system recognizes; creation of the PocketSphinx project, an open source
2) The lexical model tries to identify sounds considering initiative to continue this work. This project is in active
the context; development, and it has bindings for C and Python [13].
3) The language model identifies the higher-level
characteristics of speech, such as words and sentences. C. Practical systems
The HMM searches the model for similar patterns that fit Vijay [14] studied the problem of phonetic
into the given audio input, producing probable matches. The decomposition in lesser-studied languages, like Native
HMM’s advantages over previous learning algorithms American and Roma language variants, using the
consists of easy implementation on a computer and PocketSphinx system. While the system does not implement
automated training without human intervention. This stems the complex rules of these languages, it is possible to
from the fact that it is assumed that in short-time ranges the leverage the existing system to recognize unknown
process is stationary, vastly reducing the computational languages, using a relatively simple lookup table that maps
effort [5]. sounds to phones [15]. Varela et al. [15] adapted the system
to the Mexican Spanish language. The authors created a
B. Implementations
language and an acoustic model, based on an auto-attendant
There are several HMM implementations, but the most telephonic system, and achieved an error rate of 6.32% [15].
advanced are the HTK Toolkit [6] and the CMU Sphinx The same process was followed for other languages, like
system [7]. Mandarin [16], Arabic [17], Swedish [18]. These examples
Hidden Markov Model Toolkit (HTK) is a set of libraries show that the PocketSphinx system is flexible enough so
used for research in automatic speech recognition, that it is relatively easy for people with phonetics training to
implemented using HMM. The HTK codebase is owned by extend it to other languages.
Microsoft, but managed by the Cambridge University Harvey et al. [4] researched how ASR systems could be
Engineering Department. Since HTK has been largely integrated with their project aimed at developing a device to
abandoned, since the last release (v3.4.1) was made in 2009, help the elderly, both inside and outside the home. The
the CMU Sphinx system is getting more attention from the authors identified the following challenges associated with
speech recognition community [6]. ASR systems used for voice command interfaces [4]:
The original SPHINX was the first accurate Large 1) Important differences between users;
Vocabulary Continuous Speech Recognition (LVCSR) 2) Similarity between certain sounds;
system, using HMM as its underlying technology, that 3) Short words provide less data for the system to analyze,
managed to be speaker independent [7]. The next version, which may lead to increased error rates;
SPHINX-II, was an improved version that was both faster 4) Different recognition languages lead to variable error
and more accurate, created by most of the same authors, rates using the same system.
using HMM as its underlying technology. It was developed, Specifically to their project, the authors found that
from the beginning, as an open source project, creating a medical conditions, which frequently affect the elderly,
community around it [8]. The next version, SPHINX-III, is create different speech patterns and their tolerance to errors
an offline version of the previous systems, with a different is quite low. With that in mind, the authors leveraged the
internal representation to allow for greater accuracy. The Sphinx library for its maturity and features. Focusing on the
signals go through a much larger amount of pre-processing creation of models and general optimization tasks, the
before they even reach the recognizer [7]. Current hardware authors managed to create a multilingual system that has a
is capable of running the recognizer for SPHINX-III in 2-second processing time on embedded systems, with error
almost real-time, but it is not suitable to processing in such rates above 70% [4].
conditions. SPHINX-4 is a complete rewrite to create a Kirchhoff et al. [19] suggested other methods to improve
more modular and flexible system that can accept multiple the ASR systems’ performance. One proposal consists of
data sources elegantly. It is a joint venture with Mitsubishi replacing the current feature-extraction algorithms with
Electric Research Laboratories and Sun Microsystems, others specially designed to discriminate certain sound
using the Java programming language. As with the third classes depending on the intended use, or through the use of
version, its intended use is offline processing, not real-time noise reduction algorithms, which can improve the data

ISBN: 978-1-61804-237-8 405

 Advances in Information Science and Applications - Volume II

Patient´s Initial Endoscopy) Report Results

Symptoms Appointment Procedure Writing Appointment

Image Capture
Video Acquisition Control
MIVcontrol

Fig. 1 Workflow for a gastroenterology medical appointment

analysis and increase the models’ accuracy, using the same

data collection routines. A different approach is collecting
more out-of-band information to increase the amount of data
Microphone MIVbox
available to the system. This might include different
processing front-ends for feature extraction (although this MIVcontrol

may be of limited use) or even non-acoustic data, such as

visual information [19]. Word Boundary
Detection
Audio Acquisition

III. ARCHITECTURE
As referred before, the gastroenterologist needs to use a Feature Extraction
pedal to capture and save the frame, and with the proposed
solution the pedal is replaced with a hands-free voice
control module, called MIVcontrol. This module was Model Comparison
developed to tackle the problems that healthcare
professionals face when performing an endoscopic
procedure. This module is part of the MIVbox device, which
is integrated in the MyEndoscopy system.
MyEndoscopy is the name of the global system MIVacquisition Commands

developed by Laranjo et al. [21], which groups several

MIVboxes, which are scattered by various healthcare
institutions. The main goal of MyEndoscopy is to link Fig. 2 MIVcontrol global architecture
different entities and standardize the patient’s clinical
process management, to promote the sharing of information
between different entities [21]. AmCreate
The MIVbox device has a web-based distributed
architecture and it is capable of acquisition, processing, Audio Feature Extraction
archiving and diffusion of endoscopic procedure results
[21]. The main goal of the MIVcontrol module is to replace Text Acoustic Model

the pedal, currently used by gastroenterologists to capture Hidden Markov

interesting frames, by voice commands that interact directly Models

with the MIVacquisition module. The MIVacquisition

module receives the video directly from the endoscopic
tower and provides it to all the MIVbox modules [20]. LmCreate

In Fig. 1 is presented a simple workflow describing the

moments that occur in a gastroenterology medical Language Model

appointment, in the healthcare institution, that results in an

Endoscopy Procedure. The MIVcontrol module can be
Speech Model
seamlessly integrated into the current workflow by allowing Dictionary

the gastroenterologist to control the MIVacquisition module

by using voice commands during the endoscopic procedure.
Grammar
In Fig. 2 is presented the overall system architecture, as a
component of the MIVbox device. The MIVcontrol module
uses the live audio streaming from a microphone to
recognize commands and send them to the MIVacquisition Fig. 3 MIVcontrol model training procedure
module.
The process that leads to the creation of a model is The process is split in two main sub-processes: creation
presented on Fig. 3. It uses a corpus of pre-labeled audio of the textual model, named lmCreate, and the creation of
data to create a speech model that can be used in the the acoustic model, named amCreate. Since the acoustic
MIVcontrol module. This speech model is a combination of model requires parts of the textual model, it must be
acoustic and language models. generated last. The lmCreate process creates the textual
model based on the textual data in the corpus, while the

ISBN: 978-1-61804-237-8 406

 Advances in Information Science and Applications - Volume II

amCreate process analyzes the pre-recorded audio data

using the same feature extraction steps used in the
MIVcontrol module, and uses HMM to learn how to classify
the commands contained in the language model.

IV. IMPLEMENTATION
The creation of the speech model used in the MIVcontrol
module, from higher to lower level comprises three different
phases, namely language model, dictionary and acoustic
model.
A. Language Model
The language model is a high-level description of all
valid phrases (i.e. combination of words) in a certain
language. Statistical language models try to predict all the
valid utterances in a language, by combining all the
recognized words into every possible combination [22].
Context-Free Grammars are restricted forms of a language
model, that restrict the recognized phrases to a
predetermined set, and discard those that do not fit that
model [23].
The decision to adopt a certain language model depends
mostly on its intended application. While statistical Fig. 4 Folder tree required by SphinxTrain.
language models are useful for open-ended applications, like
dictation and general-purpose recognition, context-free The folder directory has two top folders, namely the etc
grammars are suitable for specific applications, like and the wav.
command-and-control systems. The etc folder contains all the metadata and configuration
SphinxBase requires the grammar to be defined in Java parameters needed to train the acoustic model, as well as the
Speech Grammar Format (JSGF), which is a platform- dictionary. It contains both a list of all the phonemes used in
independent standard format to define context-free the model and a list of filler phonemes, such as silences, that
grammars, using a textual representation so that it can be should be ignored. It also has a list of all the files to be used
human-readable [24]. The statistical language model is during both training and testing phases, as well as a
automatically created based on the command list. mapping between each audio file and its corresponding
B. Dictionary transcription. This mapping corresponds to the labeled data
The dictionary is a map between each command and the to be the input to the HMM.
phonemes it contains. A phoneme is defined as the basic The wav folder simply contains all the collected data, as
unit of phonology, which can be combined to form words. audio files, organized in subfolders by speaker
Its internal representation consists of using the ARPAbet to identification, with a subfolder for each set of uttered
represent phonemes as ASCII characters. The ARPAbet commands.
does not allow representing the entire International Phonetic
Alphabet (IPA), but it is sufficient for small vocabularies, The system processes continuous audio in real-time,
such as the one required by this application [25]. splits it in commands and produces a line of text for each
Since the list of required commands is small, all the recognized command. If the spoken command is not
dictionaries used were created manually. recognized, an empty line is produced. The MIVcontrol
module runs on the MIVbox.
C. Acoustic Model The audio picked up by the microphone is stored in a
The acoustic model is trained using SphinxTrain and memory buffer. The first pre-processing stage involves
maps audio features to the phonemes they represent, for splitting the incoming audio into different utterances, or sets
those included in the dictionary. The training performed by of words, by tracking silent periods between them. To
SphinxTrain requires previous knowledge of the dictionary account for noise present during recording, any audio with
and a transcription for each utterance, in order to map each volume below a certain threshold is considered a silence.
utterance to its corresponding phonetic information. It also Each segmented utterance then goes through a similar
requires the data to be in a particular audio format. In order process. The audio is processed creating a set of features,
to minimize clerical errors and cut the time need to analyze and then the Semi-Continuous HMM finds the most likely
the data to a minimum, all the technical considerations and utterance contained in its dictionary. This is the final output,
index building were abstracted away in a script referred to corresponding to a command given to the system.
as amCreate. If there is Internet access and the data to be recognized is
SphinxTrain requires the folder tree presented on Fig. 4, not sensitive, it is possible to use an online speech
where “model” denotes the model name. recognition service, such as the Google Speech API [26], as
a fallback mechanism.

ISBN: 978-1-61804-237-8 407

 Advances in Information Science and Applications - Volume II

system classified the Portuguese model with 11.22 % WER

and the English model with 4.55 % WER.
V. METHODOLOGIES
Table 2 Effect of the number of Gaussians on the WER
The parameters that have a bigger impact on the model’s
accuracy are the number of tied states used in the HMM and
the number of Gaussian mixture distributions, so testing will
focus on this parameters. Before testing begins, the data is
split randomly between training and testing stacks, with the
testing stack receiving 10% of the data. That same data is
tested varying both the number of tied states in the HMM
and the number of Gaussian distributions. The accuracy of
the model is represented as a Word Error Rate (WER),
which combines both false positives and false negatives into VII. DISCUSSION
a single metric. This is done because this module acts as
As a proof-of-concept, this system managed to create a
middleware, being used by other modules on a global
voice recognizer for a very small vocabulary to be used as a
system. The context where the commands are spoken can
command and control system, leveraging the capabilities of
then be considered, which is not evaluated here.
the CMU Sphinx project. It was created as an alternative to
Furthermore, this methodology is consistent with the
cloud-based solutions, such as Google Speech API. In a
literature on the subject.
medical environment, cloud-based solutions pose certain
The results were obtained on a computer with 2 GB of
challenges that might degrade their performance, such as
RAM and an Intel Celeron CPU, with two cores and a clock
increased communications security, a need to keep recurring
speed of 1.10GHz. The operative system used was Fedora
costs on non-medical equipment to a minimum, and also
19, 32 bits version. The compiler used was gcc v4.8.2, using
privacy and legal reasons on systems that deal with sensitive
PocketSphinx v0.8 and SphinxBase v0.8, both from the
data. Having a system that can be installed inside the
official repositories. The training used CMUcltk v0.7,
healthcare institutions’ network without external
compiled from source, and SphinxTrain v1.0.8, from the
dependencies is a plus for the reasons presented above.
official repositories. The training data was collected with the
PocketSphinx was based on work done for SPHINX II,
built-in laptop microphone, in both noisy and quiet
which was not designed as a real-time recognizer. With all
conditions, to better correspond to the concrete use-case.
the optimizations it has received, it is possible to use it in
real-time with acceptable performance, even in
VI. RESULTS
underpowered computers.
The audio corpus in which the system was tested
contained two languages, Portuguese and English, with a VIII. CONCLUSION AND FUTURE WORK
total of 1405 recordings, totaling 25 minutes of speech,
In summary, this paper presents an automatic speech
recorded by 5 female and 7 male speakers. To test this
recognition system designed specifically to solve a problem
model, SphinxTrain tried to predict the contents of the
that affects gastroenterologists. The system is capable of
testing data using the model created with the training data.
running on the background continuously without human
The main parameters that can be tweaked are the number of
physical intervention, and so it is capable of replacing the
tied states in the HMM and the number of Gaussian
pedal and buttons commonly used in current endoscopic
mixtures distributions.
systems. It was designed to be used seamlessly with the
The effect of the number of tied states in the HMM is
MyEndoscopy system that is being tested in some healthcare
shown on Table 1.
institutions.
The next step will involve improving the integration with
Table 1 Effect of the number of tied states on the WER
the MyEndoscopy system, including a more robust testing
phase, which is facilitated by the fact that PocketSphinx is a
cross-platform library.
To increase the usefulness of the system, it is important
to collect more data, particularly with different voice
features. It is also possible to apply newer training
algorithms to the same data, and test how they affect the
models created. The SphinxTrain suite was created using
CMU Sphinx recognizers, but there are more recent projects
that are able to produce models compatible with
The effect of the number of Gaussian mixtures PocketSphinx-based recognizers. Those newer systems may
distributions on the error rate is shown on Table 2. generate better models with the same data.
Since the trained model is small, the differences in
processing time are negligible. That defined the optimal REFERENCE
conditions for training 8 Gaussian mixture distributions with [1] E. H. O’Neil, “Recreating Health Professional Practice for a New
Century,” San Francisco, CA, 1998.
100 tied states in the HMM. With this configuration, the

ISBN: 978-1-61804-237-8 408

 Advances in Information Science and Applications - Volume II

[2] N. Summerton, “Positive and negative factors in defensive on Speech Communication and Technology, 1997, pp. 2707–
medicine: a questionnaire study of general practitioners.,” BMJ, 2710.
vol. 310, no. 6971, pp. 27–29, Jan. 1995. [23] A. Bundy and L. Wallen, “Context-Free Grammar,” in Catalogue
[3] J. M. Canard, J.-C. Létard, L. Palazzo, I. Penman, and A. M. of Artificial Intelligence Tools, A. Bundy and L. Wallen, Eds.
Lennon, Gastrointestinal Endoscopy in Practice, 1st ed. Berlin, Heidelberg: Springer Berlin Heidelberg, 1984, pp. 22–23.
Churchill Livingstone, 2011, p. 492. [24] A. Hunt, “JSpeech Grammar Format,” 2000.
[4] A. P. Harvey, R. J. McCrindle, K. Lundqvist, and P. Parslow, [25] R. A. Gillman, “Automatic Verification of Hypothesized
“Automatic speech recognition for assistive technology devices,” Phonemic Strings in Continuous Speech,” Arlington, Virginia,
in Proc. 8th Intl Conf. Disability, Virtual Reality & Associated 1974.
Technologies, Valparaíso, 2010, pp. 273–282. [26] B. Ballinger, C. Allauzen, A. Gruenstein, and J. Schalkwyk, “On-
[5] M. Aymen, A. Abdelaziz, S. Halim, and H. Maaref, “Hidden Demand Language Model Interpolation for Mobile Speech
Markov Models for automatic speech recognition,” in 2011 Input,” Elev. Annu. Conf. Int. Speech Commun. Assoc., no.
International Conference on Communications, Computing and September, pp. 1812–1815, 2010.
Control Applications (CCCA), 2011, pp. 1–6.
[6] S. Young, G. Evermann, D. Kershaw, G. Moore, J. Odell, D.
Ollason, V. Valtchev, and P. Woodland, “HTK FAQ.” [Online].
Available: http://htk.eng.cam.ac.uk/docs/faq.shtml. [Accessed:
03-Feb-2014].
[7] K.-F. Lee, H.-W. Hon, and R. Reddy, “An overview of the
SPHINX speech recognition system,” IEEE Trans. Acoust., vol.
38, no. 1, pp. 35–45, 1990.
[8] X. Huang, F. Alleva, H.-W. Hon, M.-Y. Hwang, K.-F. Lee, and
R. Rosenfeld, “The SPHINX-II speech recognition system: an
overview,” Comput. Speech Lang., vol. 7, no. 2, pp. 137–148,
Apr. 1993.
[9] P. Lamere, P. Kwok, E. Gouvea, B. Raj, R. Singh, W. Walker, M.
Warmuth, and P. Wolf, “The CMU SPHINX-4 speech
recognition system,” in IEEE Intl. Conf. on Acoustics, Speech
and Signal Processing (ICASSP 2003), Hong Kong, 2003, vol. 1,
pp. 2–5.
[10] K. Vertanen, “Baseline WSJ Acoustic Models for HTK and
Sphinx: Training recipes and recognition experiments,”
Cavendish Lab. Univ. Cambridge, 2006.
[11] G. Ma, W. Zhou, J. Zheng, X. You, and W. Ye, “A comparison
between HTK and SPHINX on chinese mandarin,” in IJCAI
International Joint Conference on Artificial Intelligence, 2009,
pp. 394–397.
[12] D. Huggins-Daines, M. Kumar, A. Chan, A. W. Black, M.
Ravishankar, and A. I. Rudnicky, “Pocketsphinx: A Free, Real-
Time Continuous Speech Recognition System for Hand-Held
Devices,” 2006 IEEE Int. Conf. Acoust. Speed Signal Process.
Proc., vol. 1, pp. I–185–I–188, 2006.
[13] D. Huggins-Daines, “PocketSphinx v0.5 API Documentation,”
2008. [Online]. Available:
http://www.speech.cs.cmu.edu/sphinx/doc/doxygen/pocketsphinx
/main.html. [Accessed: 20-Feb-2014].
[14] V. John, “Phonetic decomposition for Speech Recognition of
Lesser-Studied Languages,” in Proceeding of the 2009
international workshop on Intercultural collaboration - IWIC
’09, 2009, p. 253.
[15] A. Varela, H. Cuayáhuitl, and J. A. Nolazco-Flores, “Creating a
Mexican Spanish version of the CMU Sphinx-III speech
recognition system,” in Progress in Pattern Recognition, Speech
and Image Analysis, vol. 2905, A. Sanfeliu and J. Ruiz-
Shulcloper, Eds. Berlin, Heidelberg: Springer Berlin Heidelberg,
2003, pp. 251–258.
[16] Y. Wang and X. Zhang, “Realization of Mandarin continuous
digits speech recognition system using Sphinx,” 2010 Int. Symp.
Comput. Commun. Control Autom., pp. 378–380, May 2010.
[17] H. Hyassat and R. Abu Zitar, “Arabic speech recognition using
SPHINX engine,” Int. J. Speech Technol., vol. 9, no. 3–4, pp.
133–150, Oct. 2008.
[18] G. Salvi, “Developing acoustic models for automatic speech
recognition,” 1998.
[19] K. Kirchhoff, G. A. Fink, and G. Sagerer, “Combining acoustic
and articulatory feature information for robust speech
recognition,” Speech Commun., vol. 37, no. 3–4, pp. 303–319,
Jul. 2002.
[20] J. Braga, I. Laranjo, D. Assunção, C. Rolanda, L. Lopes, J.
Correia-Pinto, and V. Alves, “Endoscopic Imaging Results: Web
based Solution with Video Diffusion,” Procedia Technol., vol. 9,
pp. 1123–1131, 2013.
[21] I. Laranjo, J. Braga, D. Assunção, A. Silva, C. Rolanda, L. Lopes,
J. Correia-Pinto, and V. Alves, “Web-Based Solution for
Acquisition, Processing, Archiving and Diffusion of Endoscopy
Studies,” in Distributed Computing and Artificial Intelligence,
vol. 217, Springer International Publishing, 2013, pp. 317–24.
[22] P. Clarkson and R. Rosenfeld, “Statistical language modeling
using the CMU-cambridge toolkit,” in 5th European Conference

ISBN: 978-1-61804-237-8 409