=Paper=
{{Paper
|id=Vol-1391/71-CR
|storemode=property
|title=Automatic Clinical Speech Recognition for CLEF 2015 eHealth Challenge
|pdfUrl=https://ceur-ws.org/Vol-1391/71-CR.pdf
|volume=Vol-1391
|dblpUrl=https://dblp.org/rec/conf/clef/LuuPDHC15
}}
==Automatic Clinical Speech Recognition for CLEF 2015 eHealth Challenge==
https://ceur-ws.org/Vol-1391/71-CR.pdf
Automatic Clinical Speech Recognition for CLEF 2015
eHealth Challenge
Thoai Man Luu, Robert Phan, Rachel Davey, Girija Chetty
University of Canberra
ltm128@gmail.com, robertphan10s62@yahoo.com, Rachel.davey@canberra.edu.au,
girija.chetty@canberra.edu.au
Abstract. In this working notes report/paper, we describe the details of two submis-
sions for CLEF 2015 eHealth challenge for Task 1a, with details of methods and tools
developed for automatic speech recognition of NICTA synthetic nursing handover
dataset. The first method involves a novel zero-resource approach based on unsuper-
vised acoustic only modeling of speech involving word discovery, and the second
method is based on combination of acoustic, language, grammar and dictionary models,
using well known open source speech recognition toolkit from CMU, the CMU
Sphinx[7]. The experimental evaluation of the two methods was done on Challenge
dataset (NICTA synthetic nursing handover dataset).
1 Introduction
Fluent information flow is important in any information-intensive area of decision
making, but critical in healthcare. Clinicians are responsible for making decisions
with even life-and-death impact on their patients’ lives. The flow is defined as links,
channels, contact, or communication to a pertinent person or people in the organisa-
tion [1, 2, 3]. In Australian healthcare, failures in this flow are associated with over
one tenth of preventable adverse events [1, 2, 3]. Failures in the flow are tangible in
clinical handover, that is, when a clinician is transferring professional responsibility
and accountability, for example, at shift change [3]. Regardless of verbal handover
being accurate and comprehensive, anything from two-thirds to all of this information
is lost after three to five shifts if no notes are taken or they are taken by hand [1, 2, 3].
Nursing ‘handover’ in the clinical context involves the transfer of information, pro-
fessional responsibility and accountability for patient quality care and safety from one
clinical team to another either temporarily or permanently [4]. With changes in work-
ing hours and shifts of clinical teams (doctors, nurses and registrars in health care
system), and an increasing demand for flexible work practices, the need for mecha-
nisms to support effective and efficient handover processes for transferring infor-
mation, responsibility, accountability and patient safety has become recognised as
increasingly important for the delivery of high quality health care [5]. Clinical hando-
ver has been identified as a high risk scenario for patient safety with dangers of dis-
continuity of care, medical errors, adverse events and the potential for legal claims of
malpractice[5].
In general, implementation of ICTs in health care, to improve quality and safety has
achieved mixed results. While some studies have demonstrated significant benefits
and improvements in patient care, others have either met with mixed success or failed
to generate their forecasted benefits [6]. Given strong advocacy through guidelines
[7] and the vast amount of resources and funding which have been allocated for im-
plementation of electronic solutions to health care, there is an urgent need to generate
a better understanding of the effect of the implementation of ICTs in health care. One
of the reason for such mixed and suboptimal outcomes could be due to complexity of
clinical handover processes, characterized with highly unstructured information
flows(free text from nursing handover notes or those transcribed from speech recog-
nisers, for example), and inability of existing technologies and tools in making sense
of such ill structured or unstructured data. This could be due to limitations of existing
speech recognition technologies for instance, which act as front ends in automatic
transcription of bed side clinical notes to text, and their vulnerability to noisy clinical
environments and sensitivity to accent and dialect variations, leading to errors getting
cascaded in subsequent stages of information extraction. CLEF eHealth Challenge
Task 1a focused on addressing the short comings of existing clinical speech recogni-
tion systems by providing an open source challenge data set developed by authors in
[3], and provided an opportunity for researchers and practitioners by soliciting sub-
missions on suitable approaches and methods to this challenge task.
In this paper, we present two methods we have developed and submitted to this chal-
lenge (CLEF eHealth 2015 evaluation challenge task (Task 1a)). The details of each
method used and outcomes from the experimental trials are described in detail in next
few Sections.
2 Method I : Zero Resource Unsupervised Acoustic Modelling
(Team UC_submission 1)
For this method (Team UC_submission 1), we used a novel approach based on zero
resource unsupervised acoustic modelling technique. Zero resource speech technolo-
gies operate without the expert provided linguistic knowledge that standard recogni-
tion systems rely on—transcribed speech, language models, and pronunciation dic-
tionaries. They are motivated by biologically inspired infant learning modes, and are
suitable for less resourced contexts. As the challenge dataset comprised of non-native
English language speaker recordings, with abbreviations and terms from clinical set-
tings, traditional resources for training in terms of phonetic transcriptions, dictionaries
and grammars for this context are scarce, and speech recognizer cannot perform well
in this scenario. A robust zero-resource system must instead discover this linguistic
knowledge from speech audio automatically.
The system for this approach consists of acoustic feature extraction and seg-
mentation module, clustering module and word discovery module. Here, the acoustic
similarities between multiple acoustic tokens of the same words or word like seg-
ments are exploited to perform recognition. Although, the performance of this method
currently falls short of capabilities of the performance benchmarks provided by the
challenge, the value of this algorithm is its potential to serve as a computational mod-
el in two research directions. First, this method may lead to a speech recognition ap-
proach that is fundamentally liberated from the extensive resources needed to perform
automatic speech recognition, in terms of language models and pronunciation diction-
aries. Second, it can lead to an approach for computational modelling of language
acquisition that takes actual speech signal and is able to discover words as “evolving”
properties from raw input.
The motivation behind using this approach for discovering words from the raw speech
signal is drawn from evolving speech recognition capabilities of babies and young
infants, who can detect words from continuous speech. Psycholinguistic research [10,
11, 12, 13, 14] shows that babies can use the statistical correspondence of sound se-
quences as a cue for word segmentation. Also, the techniques that learn to decode
speech without an upfront specified lexicon and phone models are interesting for
recognizing speech outside of the vocabulary (OOV), such as in clinical domain,
where there are several words with clinical meanings and abbreviations. For these
scenarios - use of existing resources such as language models and pronunciation dic-
tionaries will be a mismatch, and might radically reduce the speech recognizer per-
formance. Hence, it is of considerable interest to investigate recognition approaches
that circumvent the need for a priori defined lexicon.
The focus for this method hence was to discover the words and word-like speech
fragment by combining raw speech signals, and additional abstract representations of
this speech signal that can model statistical co-occurrence information, by extracting
repetitive structure within the speech fragment. For this we exploit two types of
evolving patterns in the speech, the statistical properties of repetitive structure within
the speech modality to hypothesize speech fragments or segments and their labelling,
and cross-modal associations between the speech segments to hypothesize words,
which can evolve when more and more input has been processed to represent the
word correctly. The method consists of three modules, and instead of employing any
phonetic recognizers to transcribe speech fragments in terms of phone sequences, we
do a bootstrap aggregation with abstract representations to improve the speech tran-
scriber performance.
As the speech signal gets transformed from one module to the next, it gets
more symbolic in nature. The first module consists of automatic feature extraction,
followed by a data driven boundary segmentation of speech segments at sentence
level. The output of this module is a set of feature vectors and hypothesized speech
segment boundaries. Module 2 is a clustering module, which reads the sentence fea-
ture vector segments and performs a k-means clustering, and gives label sequence for
each segment. The third and final module implements the word discovery algorithm,
using the sentence and hypothesized labels, and abstract tags representing presence of
a word in that utterance.
2.1 Module 1: Feature Extraction and Boundary Segmentation Module
In this module, audio file is down sampled to 16 kHz, and each speech frame is ob-
tained by windowing the speech signal with 32 milliseconds windows (e.g. 512 sam-
ple points for 16 kHz files) with 25 % percent overlap between consecutive frames.
Cepstral mean substraction is then performed to normalize the frames, different
acoustic features are extracted, including mel frequency cepstral coefficients
(MFFCs), log energy, delta and delta-delta features. A total of 39 features are extract-
ed from each frame comprising 12 MFCCs, 1 log-energy, 12 delta, and 12 delta-delta
features [10, 11]. The distance between two frames, f1 and f2 was obtained by
𝑓𝑡𝑓
𝑑(𝑓1 , 𝑓2 ) = cos −1 ( 1 2⁄ 𝑡 𝑡 ) (1)
√𝑓1 𝑓1 𝑓2 𝑓2
where |𝑡 indicates transpose of the feature vector.
A high similarity or correlation corresponds to a small distance ‘d’ and vice versa.
Next, by using sliding window, we search the segment boundaries, where the bounda-
ry is hypothesized if the distance function that measures the difference between the
average of feature vectors before the boundary and after the boundary attains a local
maximum above a certain threshold. We use a window of 2 frames to either side of
the boundary. And, with log(E) as the weighing factor, the criterion for detecting the
boundary is:
𝑓 +𝑓𝑖−1 𝑓𝑖+2 +𝑓𝑖+1
log(𝐸) . 𝑑( 𝑖−2 , )>𝛿 (2)
2 2
2.2 Module 2: Clustering Module
This module takes as input, the segments from module 1, fits a Gaussian model to
each segment, and clusters different segment models using k-means clustering algo-
rithm . In this clustering module, the distance between two segment models S1 and S2
is defined similar to equation (1). For a better tractability, clustering of segment mod-
els is not applied to complete set, but first 10 utterances were first processed and in
subsequent steps, more segments, in increments of 10 utterances were added and clus-
ters updated until all sentences/segments in the data set were included. The output of
this module is a set of unique labels assigned to each cluster.
Table 1: Abstract Tags indicating the presence of a word in the utterance Each utterance
is associated with abstract information that indicates the presence of a word, but not its
acoustical representation or its position in the utterance. As an illustration, this table
shows eight abstract tags, related to the occurrence of ’forty’, ’eight’, ‘years’, ‘old’, ‘bed’,
‘investigation’, ‘monitoring’, ‘stable’
Sentence/utterance in
the audio file Tag1 Tag2 Tag3 Tag4 Tag5 Tag6 Tag7 Tag8
Mike hanley, 48 years old yes yes yes yes No no no no
under Dr Johnson, bed 3 no no no no Yes no no no
came in for investigation no no no no No yes no no
On regular nitros no no no no No no no no
obs are all stable no no no no No no no yes
monitored accordingly no no no no No no yes no
2.3 Module 3: Word Discovery Module
This module is for word discovery, and works by taking as input the wave files, in
combination with the sequence of labels from the clustering module, and abstract
tags, shown in Table 1. The word discovery algorithm for this module involves a
DTW (Dynamic Time Warping) algorithm, where the likelihood of two utterance
sharing a common word is estimated using a DTW on two label sequences, with the
assumption that the audio segment plus abstract tags are available as a list (Table 1).
The word discovery algorithm works as follows.
1. New utterance is selected
a. Two empty sets Amatch and Ano_match are initialized.
b. The new utterance is compared with all previously observed utter-
ances using DTW algorithm on all corresponding label sequences.
c. On the best path found by DTW, best-matching sub sequence is
found.
d. If both utterances share the same abstract tag, then this best-
matching sub sequence is put in Amatch, otherwise in Ano_match.
2. All items in these sets are sorted according to their occurrence,
3. From the Amatch, N-best utterances are selected, that do not occur in Ano_match.
4. Repeat from Step 1 again.
The advantage of this simple word discovery algorithm is, that it is able to bootstrap
from the speech signal itself without using any predefined lexical knowledge or phone
models. As the word discovery module consists of a cascade of intertwined stages,
the evolution of correct word discovery improves incrementally with more data, better
label sequence information from clustering module and availability of abstract tags
indicating the presence of a word. The same distance measure is used in both module
2 and module 3, and the same DTW principle is used to define distances between
segments and to represent the symbol hypothesis of shared word like speech seg-
ments.
Some interesting points that should be noted for this method are that the
number of clusters in the k-means module turns out to be approximately equal to the
number of phones that can be identified in the speech material, and acquisition of
phones precedes the acquisition of words. The phone-like units are hypothesized in a
data-driven way, whereas words are hypothesized in an hierarchical manner. With
additional abstract information provided in the word discovery module, including
some paralinguistic cues, such as prosody, accent, gender and culture information,
word detection accuracy can be considerably improved.
Algorithms for different modules for this method (method I) were imple-
mented in Matlab, ported to C++ using mex compiler, and a GUI tool was built to test
different utterances from the challenge data set. A software prototype for this method
was built, and is shown in Figure 1. The experimental evaluation of this method using
challenge dataset, consisting of 100 audio files for training and another 100 files for
testing provided by CLEF challenge task1 is discussed in Section 4.
Fig. 1. Clinical Speech to Text Recognizer Software Tool
3 Method II: Using CMU Sphinx Toolkit (Team
UC_Submission 2)
For this method we used well known existing system based on CMU Sphinx Speech
Recognition toolkit [8], which is an open source repository of tools jointly designed
by Carnegie Mellon University, Sun Microsystems Laboratories and Mitsubishi Elec-
tric Research Laboratories. It is designed differently from earlier versions of Sphinx
systems in terms of modularity, flexibility and algorithmic aspects. Some of the im-
provements from the earlier versions include newer search strategies, wide range of
grammar and language models, and different types of acoustic models and feature
streams. Due to several algorithmic innovations included in the system design it is
possible to incorporate multiple sources in an elegant manner. Further, the system is
modular, and is available in different versions, such as Sphinx4, Sphinx5, Pocket
Sphinx and Pocket Sphinx for Android. While Sphinx4 version is entirely developed
on the Java™ platform and is highly portable, flexible, and easier to use with multi-
threading, the Pocket Sphinx is migrated from legacy C code with appropriate wrap-
pers.
The speech recognition is performed in Sphinx 4 using a combination of HMM-based
acoustic models and appropriate language and grammar models. Due to modularity of
Sphinx architecture, it is possible to change the language model from a statistical N-
gram language model to a context free grammar (CFG) or a stochastic CFG by modi-
fying only one component of the system, namely the linguist. Likewise, it is possible
to run the system using continuous, semi-continuous or discrete state output distribu-
tions by appropriate modification of the acoustic scorer. Further, information from
multiple information streams can be incorporated and combined at any level, i.e.,
state, phoneme, word or grammar, and search module can also be switched between
depth-first and breadth-first search strategies [8]. Figure 2 shows the overall architec-
ture of the CMU Sphinx decoder.
Fig. 2. CMU Sphinx Decoder Architecture [8]
As shown in Figure 2, the front-end module parameterizes the speech signal, and
sends the extracted features to the decoder block. The decoder block consists of
search manager module, linguist module and acoustic scorer module, and decoding is
performed by co-ordination of these three blocks. The details of each module is de-
scribed briefly here.
3.1 Front End Module
Figure 3 shows the detailed representation of the front-end module, which consists of
several communicating blocks, each with an input and an output. The input of each
block is linked to the output of its predecessor, and probes it to find out if the incom-
ing information is speech data or control signal. The purpose of control signal here is
to indicate the beginning or end of speech, or data dropped or some other problem. If
the incoming data is speech, it is processed and the output is buffered, waiting for the
successor block to request it. This design has several advantages, as it allows the out-
put of any of the blocks to be tapped, actual input to the system to be any of the in-
termediate blocks, not just the first block. Due to this arrangement, it is possible to
plugin not only speech signals, but also spectra, cepstra or other kinds of auditory
representations for running the system.
Fig. 3. CMU Sphinx Front End Module [8]
The system is capable of running in different modes, including continuously from a
stream of speech, and fully end pointed, where the system performs explicit end
pointing, determining both beginning and ending end points of a speech segment au-
tomatically. The algorithm for the endpoint detection is based on comparison of ener-
gy level to three threshold levels, where two out of these three are used to determine
start of speech, and one for the end of speech. Also, the starting and/or ending of
speech from the incoming audio is detected by end pointer, and the end pointer en-
sures that the decoder does not waste any time by processing non-speech segments,
by sending only speech segments to the decoder, and discarding any non-speech seg-
ments.
3.2 Decoder Block
There are three modules in the decoder block: search manager, linguist, and acoustic
scorer, as described below.
3.2.1 Search Manager
The search manager constructs and searches a tree of possibilities for the best hypoth-
esis, by using the information from the linguist. Also, the communication with the
acoustic scorer to obtain the acoustic scores for incoming data is done by the search
manager. A token tree is used by the search manager [9], to represent the information
about the search and complete history of all active paths a given point. Each token in
the token tree contains the overall acoustic and language scores of the path, the refer-
ence to SentenceHMM reference, an identifier to the input feature frame, and the
previous token reference, facilitating backtracing. The search manager is able to fully
categorize a token to its senone, context-dependent phonetic unit, pronunciation, word
and grammar state with the Sentence HMM reference. A set of active tokens is main-
tained in the active list in the search algorithm, to represent the tips of active search
branches. During search phase, each input feature frame is scored against the acoustic
models associated with each token in the active list, and pruning of low scoring
branches is done. After pruning, the active list is updated by the search manager, us-
ing the successive SentenceHMM states of the tokens. New implementations that can
provide alternate methods of storing and pruning of the active list can be easily creat-
ed. The active list available as part of the final recognition results, can then be used by
applications to inspect the highest scoring paths, and construct N-Best lists.
The next important mechanism in the search manager is searching through the token
tree and the sentenceHMM, which is performed in two different ways: depth-first or
breadth-first. Depth-first search is analogous to conventional stack decoding, where
there is a time-sequential expansion of most promising tokens, and hence the paths
from the root of the token tree to currently active tokens can be of varying lengths.
However, for the breadth-first search, there is a synchronous expansion of all active
tokens, resulting in equally long paths from the root of the tree to the currently active
tokens. Further, breadth-first search is performed using the standard Viterbi algo-
rithm, in which during search process, competing units (phoneme, word, grammar
etch) are each represented by a directed acyclic graph (DAG). As can be seen in Fig-
ure 4, each DAG has a source and a sink, with Figure 4a showing the two-node DAGs
for two competing phonemes AX and AXR, and a more complicated association rep-
resented by DAGs for the competing word units CAT and RAT, as in Figure 4b. For
Viterbi decoding mechanism, the winner is decided by scoring each competing unit
using the probability of the single best path, and the unit with the best-path score
wins. For instance, if the phonemes AX and AXR have probabilities on the edges as
(0.9, 0.02, 0.01) and (0.2, 0.7, 0.6) respectively, then the scores would be 0.9 and 0.7
and the AX would be the winner. However, if sum of the probabilities instead of the
maximum is used for scoring, then the phoneme AXR would be the winner.
.
Fig. 4. Search Manager DAG module [8]
3.2.2 Linguist
The purpose of linguist is to translate the linguistic constraints provided to the sys-
tem into an internal data construct, called the grammar, which search manager uses it
for search. Typical linguistic constraints are provided in the form of context free
grammars, N-gram language models, finite state machines etc. The directed acyclic
graph ( DAG ) representation is also used for grammar, with each node representing a
set of words, that may be spoken at a particular time
Linguistic constraints are typically provided in the form of context free grammars,
N-gram language models, finite state machines etc. The grammar is also represented
with directed graph, with each node representing a set of words that may be spoken at
a particular time. The associated language and acoustic probabilities are shown by
arcs for connecting nodes, which predict the likelihood of transmitting from one node
to another.
Due to pluggable nature of CMU Sphinx, it is possible to load new grammars with
several grammar loaders which can load different external grammar formats and gen-
erate internal grammar structure. This grammar is then compiled into a Sen-
tenceHMM, which is basically a directed state graph, with each state in the graph
represented a unit of speech. Then, a series of word states are extracted by decompo-
sition of grammar nodes, with each node representing a word state. Next, a series of
pronunciation states are obtained by decomposition of word states, with pronuncia-
tions extracted from a dictionary maintained by the linguist. And then, each pronunci-
ation state is decomposed into a set of unit states, where these units may represent
phonemes, diphones, and these could be specific to contexts of arbitrary length. Final-
ly, each phoneme/diphone unit is then further decomposed into a sequence of HMM
states. Each unit is then further decomposed to its sequence of HMM states. The Sen-
tence-HMM construct thus comprises all of these states which are connected by arcs
that have language, acoustic and insertion probabilities associated with them.
The linguist module as such, defines the contents of the SenthenceHMM construct
very well. However, it is possible to improve the search results by altering the topolo-
gy of the SentenceHMM, the memory footprint, the perplexity, speed and the recogni-
tion accuracy. Due to pluggable nature of CMU Sphinx, it is possible to use different
SentenceHMM compilations without changing other aspects of the search.
Although the contents of a SentenceHMM are well defined by the linguist, there
are a number of strategies that can be used in constructing the SentenceHMM that
affect the search. By altering the topology of the SentenceHMM, the memory foot-
print, perplexity, speed and recognition accuracy can be affected. The pluggable na-
ture of CMU Sphinx allows different SentenceHMM compilation methods to be used
without changing other aspects of the search. For large grammars, since Sen-
tenceHMM can grow to be quite large, we use a mechanism that allows dynamic con-
truction of SentenceHMM, where it is possible to discard the SentenceHMM when no
longer needed. These features allow support for very large grammars as is normally
required for general dictation recognition tasks.
3.2.3 Acoustic Scorer Module
The next module is the acoustic scorer module which computes the state output prob-
ability or density values for the various states, for any given input vector using Gauss-
ian scoring procedures. The search module obtains these scores from the acoustic
scorer whenever it needs, and hence the acoustic scorer also communicates with the
front-end module to obtain the features for which the scores need to be computed. All
the information pertaining to the state output densities is retained by the scorer, and
hence the search manager module is ignorant of whether scoring is done with contin-
uous, semi-continuous or discrete HMMs. The speeding up of the scoring procedure
is performed by heuristic algorithms locally within the search module, where such
heuristics can benefit from additional information derived from the search module.
The details of experimental evaluation for method I, is described in Section 4.
4 Experimental Evaluation and Discussion
In this section we report results on CLEF 2015 challenge dataset, which is the NICTA
synthetic nursing handover dataset provided by challenge organizers. We segmented
each audio file to sentence level and down sampled it to 16kHz before feeding it to
both the methods (Method I and Method II). Same approach involving sentence level
segmentation and down sampling was done for both train and test subsets, where the
test subset comprised 100 different audio recordings from the same speaker. As per
the requirements of the challenge for CLEF EHealth Task 1a, the evaluation of per-
formance has to be done with NIST scoring toolkit [9], the submissions involved the
scoring toolkit results, in terms of different performance measures including detection
of correct words, insertions, deletions, substitutions and incorrect words for both
training subset and test subset.
For method 2, we could not finish the evaluation before deadline for submission, and
we submitted partial and incomplete results. However, we completed the experiments
by the due date for working notes submission and Table 2 shows the performance of
system in for both submissions.
Table 2: Evaluation of Method I /Method 2 against the benchmark performance results
Since this is still work in progress, we envisage the performance of the system, par-
ticularly method I, can be improved by appropriate choice of models, model parame-
ters, acoustic features and abstract labels, which is currently being pursued. For meth-
od 2 (UC_2_test/UC_2_train), we could not include the results in the original submis-
sion. However, as can be seen in Table 2, the average word detection accuracy on test
on the test set was 77.7 %, and on training set, it was 74.03%, comparable to bench-
mark results provided by the challenge.
5 Conclusion:
In this working notes/paper, we present the details of methods used for our two sub-
missions to CLEF eHealth Challenge Task 1a, on clinical speech recognition. First
method involve the proposal of novel zero resource word discovery algorithm, where-
as the 2nd method uses well known open source CMU Sphinx speech recognition
toolkit. Further investigations are in progress to improve the performance of each of
these approaches.
6 References:
1. Lorraine Goeuriot, Liadh Kelly, Hanna Suominen, Leif Hanlen, Aurélie Névéol, Cyril
Grouin, Joao Palotti, Guido Zuccon. Overview of the CLEF eHealth Evaluation Lab 2015.
CLEF 2015 - 6th Conference and Labs of the Evaluation Forum, Lecture Notes in Com-
puter Science (LNCS), Springer, September 2015.
2. Hanna Suominen, Leif Hanlen, Lorraine Goeuriot, Liadh Kelly, Gareth J F Jones. Task 1a
of the CLEF eHealth Evaluation Lab 2015: Clinical speech recognition. Working Notes of
the CLEF 2015 - 6th Conference and Labs of the Evaluation Forum, September 2015.
3. Hanna Suominen, Liyuan Zhou, Leif Hanlen, Gabriela Ferraro. Benchmarking clinical
speech recognition and information extraction: New data, methods, and evaluations. JMIR
Medical Informatics 2015 3(2), e19.
4. R. Wilson, B. Harrison, R. Gibberd & J. Hamilton, An analysis of the causes of adverse
events from the quality of Australian Health Care Study. Med J Aust., 170:411-415, 1999.
5. F. Armstrong, How safe are our hospitals? Australian Nursing Journal., 11:9, 18-21. 2004.
6. Charlesworth K, Jamieson M, Butler C, Davey R. The future health care. Australian
Health Review. 2015 http://dx.doi.org/10.1071/AH14243.
7. AUSTRALIAN COMMISSION ON SAFETY AND QUALITY IN HEALTH CARE
(2010) OSSIE guide to clinical handover improvement., Sydney, Australian Commission
on Safety and Quality in Health Care.
8. CMU Sphinx, http://cmusphinx.sourceforge.net/
9. S.J. Young, N.H.Russel, and J.H.S. Russel (1989). “Token passing: A simple conceptual
model for connected speech recognition systems,” Technical Report, Cambridge Universi-
ty Engineering Dept.
10. G. Chetty and M. Wagner "‘Liveness’ verification in audio-video authentication", Proc.
Int. Conf. Spoken Language Processing, pp.2509 -2512, 2004
11. Wagner, M. Chetty G., et al., The Big Australian Speech Corpus (The Big ASC), in 13th
Australasian International Conference on Speech Science and Technology. 2010, ASSTA:
Melbourne. p. 166-170.
12. A. Jansen and K. Church, “Towards unsupervised training of speaker independent acoustic
models,” in Interspeech, 2011.
13. A. Park and J. R. Glass, “Unsupervised pattern discovery in speech,” IEEE T-ASLP, vol.
16, no. 1, pp. 186–197, 2008.
14. A. Jansen, K. Church, and H. Hermansky, “Towards spoken term discovery at scale with
zero resources,” in Interspeech, 2010.
15. N.H. Feldman, T.L. Griffiths, and J.L. Morgan, “Learning phonetic categories by learning
a lexicon,” Proceedings of the 31st Annual Conference of the Cognitive Science Society,
pp. 2208–2213, 2009.
16. A. Martin, S. Peperkamp, and E. Dupoux, “Learning phonemes with a proto-lexicon,”
Cognitive Science, 2012.
17. J.F. Werker and R.C. Tees, “Influences on infant speech processing: Toward a new synthe-
sis,” Annual review of psychology, vol. 50, no. 1, pp. 509–535, 1999.
18. S. Goldwater, T. L. Griffiths, and M. Johnson, “A Bayesian framework for word segmen-
tation: Exploring the effects of context,” Cognition, vol. 112, no. 1, pp. 21–54, 2009.
19. M. Johnson, “Using adaptor grammars to identify synergies in the unsupervised acquisi-
tion of linguistic structure,” in ACL, Columbus, Ohio, 2008.
20. M. Johnson and S. Goldwater, “Improving nonparametric Bayesian inference: Experi-
ments on unsupervised word segmentation with adaptor grammars,” in NAACL, Boulder,
Colorado, 2009.
21. http://www.nuance.com/products/dragon-medical-practice-edition/index.htm