Content uploaded by Amir Hossein Poorjam
Author content
All content in this area was uploaded by Amir Hossein Poorjam on May 23, 2017
Content may be subject to copyright.
Dominant Distortion Classification for Pre-Processing of Vowels in Remote
Biomedical Voice Analysis
Amir Hossein Poorjam 1, Jesper Rindom Jensen 1, Max A. Little 2and Mads Græsbøll Christensen 1
1Audio Analysis Lab, AD:MT, Aalborg University, Aalborg, DK
2Engineering and Applied Science, Aston University, Birmingham, UK
2Media Lab, MIT, Cambridge, Massachusetts, USA
1{ahp,jrj,mgc}@create.aau.dk , 2max.little@aston.ac.uk
Abstract
Advances in speech signal analysis facilitate the development
of techniques for remote biomedical voice assessment. How-
ever, the performance of these techniques is affected by noise
and distortion in signals. In this paper, we focus on the vowel
/a/ as the most widely-used voice signal for pathological voice
assessments and investigate the impact of four major types of
distortion that are commonly present during recording or trans-
mission in voice analysis, namely: background noise, reverber-
ation, clipping and compression, on Mel-frequency cepstral co-
efficients (MFCCs) – the most widely-used features in biomed-
ical voice analysis. Then, we propose a new distortion classifi-
cation approach to detect the most dominant distortion in such
voice signals. The proposed method involves MFCCs as frame-
level features and a support vector machine as classifier to de-
tect the presence and type of distortion in frames of a given
voice signal. Experimental results obtained from the healthy
and Parkinson’s voices show the effectiveness of the proposed
approach in distortion detection and classification.
Index Terms: distortion analysis, MFCC, remote biomedical
voice assessment, support vector machine
1. Introduction
Sustained vowels are widely used for evaluation of pathologi-
cal voice caused by a range of medical disorders. Vowels have
two main advantages: first, the complexity of modeling artic-
ulatory movement during running speech is avoided [1], and
second, experimental studies show that most dysphonic speak-
ers cannot produce steady, sustained vowel sounds [2]. Among
vowels, the vowel /a/ is sufficient for many voice analysis ap-
plications [3], [4]. During production of the vowel /a/, the vocal
tract is more open than other vowels resulting in minimal air
pulse reflections between the vocal tract and the vocal folds [1].
Using clean and sustained /a/ vowels, Tsanas et al. [4] achieved
almost 99% overall accuracy in detecting Parkinson’s disease
(PD) from voice recordings, for example.
Due to advances in automatic voice analysis, remote voice
assessment is becoming feasible [5], [6]. For example, re-
cently smartphones are being investigated as tools for measur-
ing pathological voice [7] since smartphones are ubiquitous and
inexpensive devices with built-in, high-quality microphones.
Compared to samples recorded in a clinic or a sound booth,
recordings from smartphones in most environments are subject
to many types of linear and nonlinear distortion. The presence
of distortion in voice signals degrades the performance of al-
gorithms designed to quantify medical symptoms from voice
This work was funded by the Danish Council for Independent Re-
search, grant ID: DFF 4184-00056
recordings [8]. In particular, the performance of different al-
gorithms for PD detection under a variety of acoustic condi-
tions has been evaluated in [9] and it has been demonstrated
that background noise and the use of codecs significantly de-
grade detection performance.
Several approaches to detect different types of noise and
distortion in voice signals have been studied, most of have fo-
cused on detecting a single and specific kind of distortion in
voice [10–14]. In this study, we consider the vowel /a/ and aim
to detect four different types of noise and distortion that are
commonly present during recording or transmission in remote
voice analysis, namely: background noise, room reverberation,
peak clipping and coding (i.e. speech compression). Although
there are an infinite number of possible levels, types and combi-
nations of distortion in real-world scenarios, this study aims to
provide a simplified approach to detect the most dominant dis-
tortion in the signal, which would be useful in practical applica-
tions where it is important to know whether a frame is distortion
free or needs enhancement. We assume that if a given frame is
considered as corrupted, there is a specific type of noise or dis-
tortion which dominates over other distortions. Following this,
we investigate the behavior of Mel-frequency cepstral coeffi-
cients (MFCCs, widely-used features in voice-based biomedi-
cal applications [8], [15]) in the presence of the four kinds of
distortion and noise. Then, a new method is proposed which
uses a support vector machine (SVM) as classifier and MFCCs
as features for that classifier, to detect distortion in each frame.
MFCCs are selected because of their sensitivity to changes in
signal characteristics due to noise, distortions or articulatory
movements [16].
2. Effects of distortion on MFCCs
The proposed method is based upon experimental observations
of the effect of distortions on MFCCs reported next. This ex-
perimental analysis reveals that different levels and types of dis-
tortion cause MFCCs to shift to different regions of the space
spanned by the MFCC values, and changes the covariance of
these values. To explore this effect, we take successive frames
from the center of the clean vowel /a/ uttered by 45 healthy
speakers and extract the MFCCs under different types and levels
of distortion and noise. We then evaluate the shift in the sample
mean and covariances of the MFCCs computed on the distorted
signals.
2.1. MFCC Features
MFCCs are based on the source-filter theory of speech produc-
tion [17]. To compute MFCCs, we take the discrete Fourier
transform (DFT) of the speech frames. Then, the power spec-
trum is computed and passed through a set of triangular filter
banks, linearly spaced on the Mel-frequency scale. The log-
energy output of the filter bank, which is sensitive to small
changes in signal characteristics due to noise, distortions or ar-
ticulatory movements [16], is then passed through the discrete
cosine transform (DCT). The MFCCs are the amplitudes of the
DCT coefficients. Specifically, the pth MFCC coefficient of the
kth frame is calculated as [8]:
φk[p] = 1
M+ 1
M
X
q=1
log |˜
Sk(q)|cos πq
M+ 1 p,(1)
where Mis the number of Mel-band filters and ˜
Sk(q)is the
estimate of the spectral energy in the qth band calculated as:
˜
Sk(q) = X
i3fMel
i∈IMel
q
1−|fMel
i−q
M+1 FMel|
∆fMel
2|Sk(i)|,(2)
where IMel
q= [ q−1
M+1 FMel,q+1
M+1 FMel]is the qth filter band in
Mel-frequency scale, fMel
iis the ith Mel frequency, FMel is the
maximum frequency in the Mel domain, ∆fMel/2is the width
of the Mel bands and Skis the short-time DFT of the kth frame.
The transformation from the linear domain to the Mel domain
is performed by [18]:
fMel =1000
log10 2log10 1 + f
1000 .(3)
In this study, 13 MFCC coefficients are extracted for each
frame. In addition, delta and double-delta coefficients, defined
as the first- and second-order time-differences of the MFCC co-
efficients which capture the dynamic changes between frames,
are appended to the MFCCs to form a 39-dimensional vector.
Considering (1) – (3), the effects of distortions on MFCCs
are complex since during the MFCC calculations, a corrupted
signal passes through several nonlinear functions. These effects
can even be more complex when a signal has been subject to a
nonlinear distortion such as clipping or compression. To eval-
uate the behavior of MFCCs in the presence of noise and dis-
tortion, we take successive 30 ms long frames of the vowel /a/
uttered by 45 healthy speakers and compute the change in the
covariance matrix and the mean of the MFCCs under different
types and levels of distortion. Specifically, the mean shift can
be defined as:
ξ(j) = 1
N
N
X
n=1
kµdj
n−µc
nk2,(4)
where Nis the number of speakers, k·k2represents the 2-
norm, and µc
nand µdj
nare the means of the MFCCs computed
respectively from clean signal and distorted signals from the
nth speaker subject to the jth distortion level. ξ= 0 indicates
that the mean of the corrupted MFCCs is unchanged in feature
space. The larger the value of ξ, the farther the MFCC vector
is moved with respect to the clean one. The change in the co-
variance matrix of the MFCC under the jth distortion level is
measured as:
δ(j) = 1
N
N
X
n=1
trΣdj
n
trΣc
n,(5)
where Σc
nand Σdj
nare respectively the covariance matrices of
the MFCCs extracted from the clean and corrupted utterances
of the nth speaker, and tr(·)is the trace operator that maps the
MFCC covariance matrix to a single real number which rep-
resents the sum of variances for individual dimensions of the
MFCC vector. δ= 1 represents no change in covariance. A
value of δ < 1indicates a reduction in covariance with respect
to the covariance of the clean MFCC. That is, the MFCCs be-
come more compact in the feature space.
2.2. Impact of different distortions on MFCCs
In the first experiment, we investigate the impact of background
noise on MFCCs by corrupting clean vowels /a/ uttered by 45
healthy speakers by three commonly-encountered environmen-
tal noise types, namely “white Gaussian noise”, “quiet office
ambience noise” and “babble noise” under different signal-to-
noise ratio (SNR) conditions (ranging from -20 dB to 60 dB
in 1 dB steps). Babble noise, which consists of multiple speak-
ers talking in the background, has rapid, time-evolving structure
and is considered a challenging type of noise in many speech-
based applications due to its similarity to the target speech [19].
The office environment noise represents a general atmosphere
of a medium size room including the sound of air conditioning
systems and very weak background noise from outside. Fig-
ure 1(a) shows the impact of different types and levels of noise
on the mean and the covariance matrix of MFCCs. The left
vertical axis represents the amount of mean shift as defined in
(4) and the right vertical axis represents the relative change in
the covariance matrix as defined in (5). The plot suggests that
variable noise levels shift the mean of MFCCs to different, but
predictable, regions in the feature space. It can be observed
that the amount of shift monotonically increases as the level of
noise increases. Moreover, it can be noticed that the covariance
of the noisy MFCCs is always smaller than that of the clean one.
However, the covariance does not monotonically reduce. This
is probably due to the fact that as the SNR goes below 0 dB, the
noise dominates the signal and the MFCCs take on a different
profile.
Reverberation in voice recordings is caused by superim-
posed reflections of the original sound wave coming from dif-
ferent surfaces in an acoustic environment and is known to have
a detrimental impact on numerous signal processing tasks. To
study the effect of reverberation on MFCCs, we filtered the
clean signal with synthetic room impulse responses (RIRs) of
reverberation times (RTs) varying from 150 ms to 1 s mea-
sured at a room of dimension 5m×4m×3m. Furthermore, to
evaluate the effect of different source-to-receiver distances on
the MFCCs, the experiments are repeated with three differ-
ent speaker-to-microphone distances, namely 0.5 m, 1 m and
1.5 m. The RIRs are generated using the image method [20]
which is implemented using the RIR Generator toolbox [21]
in MATLAB. Figure 1(b) illustrates similar trends for MFCCs
under different speaker-to-microphone distances. The mean
shift increases as the RT increases. We can observe that when
the microphone records from a close distance, the amount of
shift is always smaller than when the microphone is recording
from a larger distances from the speaker. For large speaker-to-
microphone distances, however, we observe a different trend as
the RT exceeds 250 ms. Reverberation reduces the covariance
of the MFCCs as the RT increases.
Peak clipping and speech coding are two common nonlin-
ear speech signal modifications. Peak clipping occurs when
the amplitude of a speech signal exceeds the dynamic range
of the analogue-to-digital converter which introduces nonlinear
distortion into the signal and affects the subjective quality of
speech [22]. On the other hand, communication channels typ-
Signal SNR
White Gaussian Noise
Babble Noise
Office Ambience Noise
(a) Background noise
Reverberation time (in sec)
Speaker-to-microphone = 0.5m
Speaker-to-microphone = 1 m
Speaker-to-microphone = 1.5m
(b) Reverberation
Clipping level
(c) Peak clipping
Bit rate (kbps)
(d) Compression
Figure 1: Impact of different types and levels of distortion on the mean and covariance matrix of the MFCCs. The left vertical axes
represent ξdefined in (4) which is the amount of mean shift. ξ= 0 indicates that the mean of the corrupted MFCCs is not shifted in
the feature space. The larger the value of ξ, the farther the MFCC vector is positioned with respect to the clean one. The right vertical
axes represent δdefined in (5) which is the relative change in the covariance matrix. δ= 1 indicates no change in covariance of the
corrupted MFCCs, δ > 1represents increase in covariance with respect to the covariance of the clean MFCCs and δ < 1indicates
that the MFCCs become more compact in the feature space.
ically use lossy codecs such as code-excited linear prediction
(CELP) to compress speech signals to lower bit rates, which in-
evitably degrades the quality of the speech [23]. To study the
effect of peak clipping on MFCCs, we define the clipping level
as a proportion of the unclipped peak absolute signal amplitude
to which samples greater than this threshold are limited. The
clean recordings of the vowel /a/ are clipped with clipping lev-
els varying from 0.1 to 1 in 0.025 steps. Figure 1 (c) illustrates
the impact of peak clipping on the MFCCs. As the clipping
level increases, the mean of the MFCCs is positioned farther
away from that of the clean signal. MFCCs of a clipped signal
possess smaller covariance matrix values compared to that of
the clean MFCCs and become smaller as the clipping level in-
creases. To investigate the behavior of MFCCs when a speech
signal has undergone the distortion of a speech codec, the clean
vowels /a/ are coded by a CELP codec with three different stan-
dard bit rates, namely 6.3, 9.6 and 16 kbps [24]. Figure 1 (d)
shows the impact of speech compression on MFCCs. The plot
is produced by fitting a second order power function to the cal-
culated ξand δas:
˜
ξ(j) = −2.35 ×10−5×j3.88 + 3.59 (6)
˜
δ(j) = 1797 ×j−4.89 + 0.91 (7)
We can observe that speech compression shifts the MFCCs to a
farther position (with respect to the position of the clean ones)
as the compression rate increases. On the other hand, although
MFCCs of a voice signal coded at 16 kbps and 9.6 kbps possess
smaller covariance matrices compared to the covariance of the
clean one, we observe a larger covariance than that of the clean
MFCCs when a signal is coded at 6.3 kbps. The empirical curve
fitted to δalso suggests that MFCCs of a signal compressed at
7.3 kbps are expected to have a comparable covariance matrix
with respect to the covariance of the clean MFCCs.
3. The proposed distortion classification
system
Motivated by the experimental findings above, we introduce a
new method for noise and distortion classification to detect the
presence and type of noise/distortion in /a/ vowels. Although a
recording can be subject to an infinite number of possible types
and levels of noise and distortion in real scenarios, our approach
focuses on detecting the most dominant corruption present in
any frame. We assume the simplifying model that if a given
Estimated
Class
Training Phase
Testing Phase
MFCC Extraction
Framing
VAD
VAD indices
Frame
Dropping
MFCC Extraction
Framing
VAD
VAD indices
Frame
Dropping
SVM
Figure 2: Block diagram of the proposed method for distor-
tion/noise classification, training and testing phases.
frame of a voice recording is corrupted, there is a single type of
noise or distortion which dominates. The block diagram of the
proposed approach in training and testing phases is illustrated
in Figure 2. Using a Hamming window, recordings are seg-
mented into frames of 30 ms. For each frame of a vowel (which
can be clean or corrupted), a 39-dimensional MFCC vector is
computed. Using an energy-based voice activity detection al-
gorithm [25], silent frames at the beginning and the end of the
signals are excluded. Then, a multiclass SVM with a radial
basis function kernel estimated on the training frames is used
to classify distortion in an unseen frame during testing. Intro-
duced by Vapnik et.al [26], SVMs are powerful discriminative
pattern classifiers which find an optimal separating hyperplane
in a high dimensional nonlinear feature space formed using ker-
nels applied to the input feature space.
4. Experimental setup
The proposed system for distortion/noise recognition in /a/ vow-
els was developed and validated using two different databases.
The first database consists speech samples of healthy speak-
ers. This database contains different clean vowels uttered by 45
men, 48 women and 46 childeren, recorded by a dynamic mi-
crophone, sampled at 16 kHz and range from 370 ms to 780 ms
long [27]. There is no dysphonia variability. The only uncon-
trolled parameter is the speaker variability. From this database,
we have chosen 93 samples of /a/ vowels produced by 45 male
and 48 female speakers. Furthermore, to evaluate the proposed
system with more realistic pathological voice signals, we used a
PD voice database since the vast majority of people with PD ex-
hibit some form of vocal disorder [28]. This database was gen-
erated through collaboration between Sage Bionetworks, Pa-
tientsLikeMe and Dr. Max Little as part of the Patient Voice
Table 1: Frame- and recording-level classification performance for the healthy voice and the Parkinson’s voice databases in the form
mean ±STD computed using a 5-fold CV.
Database Frame-Level Classification Accuracy (in %±STD) Recording-Level Classification Accuracy (in %±STD)
Clean Noisy Clipped Coded Reverb. Overall Clean Noisy Clipped Coded Reverb. Overall
Healthy voice 61±6 92±3 82±3 71±6 85±4 78±1 77±12 100±0 98±3 82±11 90±7 89±4
Parkinson’s voice 48±5 89±3 74±6 77±8 66±5 72±4 55±11 97±4 82±7 85±9 77±4 79±3
Analysis study (PVA)1. The samples of this database are the
telephone recordings of the sustained vowels /a/ produced by
779 PD patients of both genders, sampled at 8 kHz and range
from 3 s to 30 s long. From this database, we randomly se-
lected 48 female and 26 male samples of 7 s to 15 s duration.
Then, we used 3 s of the middle of the signals, where the speak-
ers produced a steady sustained vowel. This database has both
speaker- and dysphonia- variability. Moreover, the recordings
may have already some types of distortion such as background
acoustic environmental noise and reverberation or may have
been through one or more codec since they are collected over
the telephone network, which makes the noise/distortion detec-
tion more challenging.
To create a database for distortion/noise detection, we en-
larged the databases by adding the distorted versions of all
recordings by applying different types and levels of noise and
distortion which typically present in the recordings of remote
voice analysis. Specifically, for noise, we added “babble”,
“white Gaussian” and “office ambiance” noises at 15 dB, 10
dB and 5 dB. For peak clipping, the clipping level was set
to 0.3, 0.4, 0.5 and 0.6. Signals were compressed using 6.3
kbps, 9.6 kbps and 16 kbps CELP codecs. To provide rever-
berant signals, recordings were filtered by 8 different real RIRs
of the AIR database [29]. The RIRs are measured with mock-
up phone in hand-held and hands-free positions in four realistic
indoor environments, namely an office, a lecture room, a cor-
ridor and a stairway. The measured RTs range from 390 ms
to 1.47 s [30]. Then, using a Hamming window of length 30
ms, we created a database of 30 ms clean and corrupted frames
for both databases. The recordings of each database are then
divided into two subsets: a training subset consisting of 80%
of the speakers, and a testing subset consisting of 20% of the
speakers. The resulting training and testing subsets of the en-
larged healthy vowel database consist of 5105 and 1360 frames,
respectively. The training and the testing subsets of the enlarged
PD voice database consist of 30150 and 7800 frames, respec-
tively. The enlarged databases have the same number of frames
per class of noise/distortion.
To detect different types of distortion in a given frame, a
multiclass SVM classifier implemented in the LIBSVM tool-
box [31] in MATLAB is used. The hyper-parameters of the
SVM, namely the RBF kernel spread and SVM regularization
parameter, were selected by 5-fold cross-validation (CV) on
10% of the training data assigned as the tuning subset.
5. Results and discussion
We used 5-fold CV to evaluate the classification performance
in terms of the number of correctly classified test frames. The
results over all CV repetitions using healthy and the PD voices
are reported in the first and the second rows of Table 1, respec-
tively. Assuming that the most dominant distortion in an utter-
ance usually affects the majority of frames, we also extend the
1They were obtained through Synapse ID [syn2321745]
proposed method to the recording-level by applying a majority
voting algorithm over all frames of a signal. The table reports
the classification accuracy both at frame and recordings levels
in the form mean ±one standard deviation.
The effectiveness of MFCCs in distortion classification
(particularly for noisy frames) can be observed. The results for
healthy voices are consistent with the behavior of MFCCs in the
presence of different types and levels of distortion observed in
Section 2.2. Considering Figure 1, MFCCs of the compressed
signals are, on average, positioned closer to the MFCCs of the
clean signals, while noise, reverberation and peak clipping shift
the MFCCs farther away from the position of clean MFCCs.
Moreover, the covariance of MFCCs extracted from coded sig-
nals is comparable to that of the clean signal, while the MFCCs
for noisy, reverberant and peak clipped signals are more com-
pact in the feature space. Taking these two observations into
account, the MFCCs of the coded and clean frames are more
likely to be overlapping in the feature space which results in
misclassification, particularly when there is speaker variability
in the database.
Although the proposed method is effective in distortion
classification for both healthy and pathological voices, we ob-
serve a degradation in overall classification performance (par-
ticularly for clean frame detection) when the system is eval-
uated using the PD voice database. The first factor affecting
the results is the dysphonia variability in the PD voice database
since the presence of pathologies in speech is related to signal
variability. Moreover, bearing in mind that the recordings in the
PD voice database have been collected over the telephone, these
signals may have already been through one or more codecs.
This means that some coded frames have been presented to the
classifier as “clean” ones during the training phase which will
result in some classification performance degradation.
6. Conclusions
In this study, the impact of four major types of distortion,
namely background noise, reverberation, clipping and speech
compression on MFCCs of the frames of the vowel /a/ has been
analyzed. These distortions are commonly present in voice sig-
nals during recording or transmission in remote pathological
voice assessments. It has been demonstrated experimentally
that introducing different types and levels of distortion to the
vowel results in predictable changes in mean and covariance
matrix of the MFCCs. Motivated by this observation, a new
approach for detecting the dominant type of distortion is pro-
posed, which uses MFCCs as frame-level acoustic features and
an SVM as the classifier. Experimental results using record-
ings of healthy speakers and speakers with PD (as an example
of people with voice disorders) show the effectiveness of the
proposed system in distortion classification. Since the presence
of disorders in speech is closely related to signal variability, a
slight degradation in classification performance has been ob-
served when the PD voices were analyzed.
7. References
[1] I. Titze, Principles of voice production, 2nd ed. Iowa City: Na-
tional Center for Voice and Speech, 1999.
[2] J. Schoentgen and R. De Guchteneere, “Time series analysis of
jitter,” J. Phon., vol. 73, pp. 189–201, 1995.
[3] F. Klingholtz, “Acoustic recognition of voice disorders: a compar-
ative study of running speech versus sustained vowels.” J. Acoust.
Soc. Am., vol. 87, no. 5, pp. 2218–24, may 1990.
[4] A. Tsanas, M. A. Little, P. E. McSharry, J. Spielman, and
L. O. Ramig, “Novel speech signal processing algorithms for
high-accuracy classification of Parkinson’s disease,” IEEE Trans.
Biomed. Eng., vol. 59, pp. 1264–1271, 2012.
[5] R. J. Moran, R. B. Reilly, P. De Chazal, and P. D. Lacy,
“Telephony-based voice pathology assessment using automated
speech analysis,” IEEE Trans. Biomed. Eng., vol. 53, no. 3, pp.
468–477, 2006.
[6] P. A. Mashima and C. R. Doarn, “Overview of telehealth activ-
ities in speech-language pathology,” Telemed. e-Health, vol. 14,
no. 10, pp. 1101–1117, dec 2008.
[7] S. Arora, V. Venkataraman, A. Zhan, S. Donohue, K. Biglan,
E. Dorsey, and M. Little, “Detecting and monitoring the symp-
toms of Parkinson’s disease using smartphones: A pilot study,”
Parkinsonism Relat. Disord., vol. 21, no. 6, pp. 650–653, 2015.
[8] R. Fraile, N.S´
aenz-Lech´
on, J. I. Godino-Llorente, and V. Osma-
Ruiz, “Use of Mel frequency cepstral coefficients for automatic
pathology detection on sustained vowel phonations: mathematical
and statistical justification,” in 4th Int. Symp. Image/Video Com-
mun. over Fixed Mob. Networks, no. 3, 2008.
[9] J. Vasquez-Correa, J. Serra, J. F. Orozco-Arroyave, J.R. Vargas-
Bonilla, and E. Noth, “Effect of acoustic conditions on algorithms
to detect Parkinson’s disease from speech,” in ICASSP, 2017, pp.
5065–5069.
[10] W. Yuan and B. Xia, “A speech enhancement approach based on
noise classification,” Appl. Acoust., vol. 96, pp. 11–19, 2015.
[11] K. El-maleh, A. Samouelian, and P. Kabal, “Frame-level noise
classification in mobile environments,” ICASSP, pp. 237–240,
1999.
[12] S. Aleinik and Y. Matveev, “Detection of clipped fragments in
speech signals,” Int. J. Electr. Comput. Energ. Electron. Commun.
Eng., vol. 8, no. 2, pp. 286–292, 2014.
[13] J. Eaton and P. A. Naylor, “Noise-robust detection of peak-
clipping in decoded speech,” ICASSP, pp. 7019–7023, 2014.
[14] J. M. Desmond, L. M. Collins, and C. S. Throckmorton, “Us-
ing channel-specific statistical models to detect reverberation in
cochlear implant stimuli.” J. Acoust. Soc. Am., vol. 134, no. 2, pp.
1112–20, 2013.
[15] A. Dibazar, S. Narayanan, and T. Berger, “Feature analysis
for automatic detection of pathological speech,” in Second Jt.
EMBS/BMES Conf., vol. 1, 2002, pp. 182–183.
[16] M. Sahidullah and G. Saha, “Design, analysis and experimental
evaluation of block based transformation in MFCC computation
for speaker recognition,” Speech Commun., vol. 54, no. 4, pp.
543–565, 2012.
[17] J. R. Deller, J. H. L. Hansen, and J. G. Proakis, Discrete-time
processing of speech signals, 2nd ed. New York: IEEE Press,
2000.
[18] J. Harrington and S. Cassidy, Techniques in Speech Acoustics.
Kluwer Academic Publishers, 1999.
[19] N. Krishnamurthy and J. Hansen, “Babble noise: modeling, anal-
ysis, and applications,” IEEE Trans. Audio. Speech. Lang. Pro-
cessing, vol. 17, no. 7, pp. 1394–1407, 2009.
[20] J. B. Allen and D. A. Berkley, “Image method for efficiently sim-
ulating small-room acoustics,” J. Acoust. Soc. Am., vol. 65, no. 4,
p. 943, 1979.
[21] E. A. P. Habets, “Room impulse response generator,” International
Audio Laboratories Erlangen, Tech. Rep., 2010.
[22] J. Gruber and L. Strawczynski, “Subjective effects of variable de-
lay and speech clipping in dynamically managed voice systems,”
IEEE Trans. Commun., vol. 33, no. 8, pp. 801–808, 1985.
[23] P. E. Souza, “Effects of compression on speech acoustics, intel-
ligibility, and sound quality.” Trends Amplif., vol. 6, no. 4, pp.
131–65, 2002.
[24] R. Goldberg and L. Riek, A practical handbook of speech coders.
CRC Press, 2000.
[25] T. Kinnunen and H. Li, “An Overview of Text-Independent Speak
er Recognition : from Features to Supervectors,” Speech Com-
mun., vol. 1, 2009.
[26] C. Cortes and V. Vapnik, “Support-vector networks,” Mach.
Learn., vol. 20, no. 3, pp. 273–297, 1995.
[27] J. Hillenbrand, L. A. Getty, M. J. Clark, and
K. Wheeler, “Acoustic characteristics of American English
vowels,” J. Acoust. Soc. Am., 1995. [Online]. Available:
http://homepages.wmich.edu/∼hillenbr/voweldata.html
[28] A. K. Ho, R. Iansek, C. Marigliani, J. L. Bradshaw, and S. Gates,
“Speech impairment in a large sample of patients with Parkinson’s
disease.” Behav. Neurol., vol. 11, no. 3, pp. 131–137, 1998.
[29] M. Jeub, M. Schafer, and P. Vary, “A binaural room impulse re-
sponse database for the evaluation of dereverberation algorithms,”
in 16th Int. Conf. Digit. Signal Process., 2009, pp. 1–5.
[30] M. Jeub, M. Sch¨
afer, and H. Kr¨
uger, “Do we need dereverberation
for hand-held telephony?” in Int. Congr. Acoust., 2010, pp. 1–7.
[31] C.-W. Hsu, C.-C. Chang, and C.-J. Lin, “A practical guide to
support vector classification,” National Taiwan University, Tech.
Rep., 2016.
