Content uploaded by James R. Williamson
Author content
All content in this area was uploaded by James R. Williamson on Sep 14, 2015
Content may be subject to copyright.
Segment-dependent dynamics in predicting Parkinson’s disease
James R. Williamson1, Thomas F. Quatieri 1, Brian S. Helfer1, Joseph Perricone1,
Satrajit S. Ghosh2, Gregory Ciccarelli1, Daryush D. Mehta1
1 MIT Lincoln Laboratory, Lexington, Massachusetts, USA
2 Massachusetts Institute of Technology, Cambridge, Massachusetts, USA
jrw@ll.mit.edu, quatieri@ll.mit.edu, brian.helfer@ll.mit.edu, joey.perricone@ll.mit.edu,
satra@mit.edu, gregory.ciccarelli@ll.mit.edu, daryush.mehta@ll.mit.edu
Abstract
Early, accurate detection of Parkinson’s disease may aid in
possible intervention and rehabilitation. Thus, simple
noninvasive biomarkers are desired for determining severity.
In this study, a novel set of acoustic speech biomarkers are
introduced and fused with conventional features for predicting
clinical assessment of Parkinson’s disease. We introduce
acoustic biomarkers reflecting the segment dependence of
changes in speech production components, motivated by
disturbances in underlying neural motor, articulatory, and
prosodic brain centers of speech. Such changes occur at
phonetic and larger time scales, including multi-scale
perturbations in formant frequency and pitch trajectories, in
phoneme durations and their frequency of occurrence, and in
temporal waveform structure. We also introduce articulatory
features based on a neural computational model of speech
production, the Directions into Velocities of Articulators
(DIVA) model. The database used is from the Interspeech
2015 Computational Paralinguistic Challenge. By fusing
conventional and novel speech features, we obtain Spearman
correlations between predicted scores and clinical assessments
of r = 0.63 on the training set (four-fold cross validation),
r = 0.70 on a held-out development set, and r = 0.97 on a held-
out test set.
Index Terms: Parkinson’s disease, speech biomarkers,
phoneme and pause duration, articulatory coordination, neural
computational models of motor control
1. Introduction
Parkinson’s disease is a neurological disorder with associated
progressive decline in motor precision and sensorimotor
integration stemming presumably from the basal ganglia. In
this disorder, there is a steady loss of cells in the midbrain,
leading to speech impairment in nearly 90% of subjects [1].
Speech and voice characteristics of Parkinson’s disease
include imprecise and incoordinated articulation, monotonous
and reduced pitch and loudness, variable speech rate and
rushes of breath and pause segments, breathy and harsh voice
quality, and changes in intonation and rhythm [2][3][4][5][6].
*This work is sponsored by the Assistant Secretary of Defense for
Research & Engineering under Air Force contract #FA8721-05-C-
0002. Opinions, interpretations, conclusions, and recommendations
are those of the authors and are not necessarily endorsed by the United
States Government. SG was partly supported by a joint collaboration
between the McGovern Institute for Brain Research and MIT LL.
Early and accurate detection of Parkinson’s disease can aid
in possible intervention and rehabilitation. Thus, simple
noninvasive biomarkers are desired for determining severity of
the condition. In this paper, a novel set of speech biomarkers is
introduced for predicting the clinical assessment of the
Parkinson’s disease severity. Specifically, we introduce speech
biomarkers representing essential speech production
mechanisms of phonation, articulation, and prosody, motivated
by changes that are known to occur at different time scales and
differentially across speech segments in the underlying neural
motor and prosodic brain centers of speech production.
With this motivation, our new features are designed based
on two basic tenets: (1) there is a phoneme dependence of the
dynamic speech characteristics that is altered in Parkinson’s
disease, and (2) vocal tract dynamics and stability are altered
in Parkinson’s disease. We exploit the phoneme dependence of
durations and pitch and formant slopes, consistent with
findings that certain speech segments are more prone to
variation than others in Parkinson’s disease [7][8][9]. We also
exploit the decline in the precision of articulatory trajectories
over different time scales and tasks [10][11].
Our paper is organized as follows. In Section 2, we
describe the data collection and preprocessing approaches. In
Section 3, we describe our signal models and corresponding
signal processing methods for speech feature extraction.
Section 4 reports predictor types and prediction results.
Section 5 gives conclusions and projections toward future
work.
2. Parkinson’s database
2.1. Audio recordings
The Parkinson’s disease database used in the Interspeech 2015
Computational Paralinguistic Challenge is described in [1].
Assessments of Parkinson’s severity are based on the Unified
Parkinson’s Disease Rating Scale (UPDRS) [12]. The data set
is divided into 42 tasks per speaker, yielding 1470 recordings
in the training set (35 speakers) and 630 recordings in the
development set (15 speakers), both with UPDRS scores
provided. It also contains 462 recordings (11 speakers) in the
test set, without UPDRS scores provided. The duration of
recordings ranges from 0.24 seconds to 154 seconds.
2.2. Audio enhancement
To address noise in the test data, we use an adaptive
Wiener-filter approach that preserves the dynamic components
of a speech signal while reducing noise [13][14]. The
approach uses a measure of spectral change that allows robust
Copyright © 2015 ISCA September 6
-
10, 2015, Dresden, Germany
INTERSPEECH 2015
518
and rapid adaptation of the Wiener filter to speech and
background events. The approach reduces speech distortion by
using time-varying smoothing parameters, with constants
selected to produce less temporal smoothing in rapidly-
changing regions and greater smoothing in more stationary
regions.
2.3. Test set annotation
Identification (ID) labels for 11 test subjects were
manually assigned to the monologue, the read-text, and the ten
sentence recordings by listening. A 128-component Universal
Background Model/Gaussian Mixture Model (UBM/GMM)
classifier was used [15], with a feature vector consisting of 16
Mel-frequency cepstral coefficients (MFCCs) and 16 delta-
MFCCs, to classify the remaining 330 recordings on the test
set. The UBM was trained from all 2,562 recordings in the
data set, and subject ID assignments on the test recordings
were obtained using 11 subject GMMs, adapted using
manually labeled tasks on the test set. Finally, manual
correction of 28 of the test subject assignments was done. The
final subject ID assignments were not perfect, with number of
assigned recordings per subject ranging from 40 to 44.
3. Feature extraction
3.1. Overview of feature sets
Feature development was designed to reflect the three basic
aspects of speech production: Phonation (source), articulation
(vocal tract), and prosody (intonation and timing). The focus
of the features is to characterize variations in dynamics of
pitch (phonation), formants (articulation), and rate and
waveform (prosody) that reflect Parkinson’s severity.
Ten different feature sets are used, designed to
characterize changes in speech as a function of Parkinson’s
severity. Each feature set (FS) comprises a set of raw features,
described in Sections 3.2‒3.4. Dimensionality reduction of all
FSs is done by z-scoring the raw features and then extracting
lower dimensional features using principal components
analysis (PCA). The PCA features provide input into
regression models that map each FS into a Parkinson’s
severity (UPDRS) prediction (see Section 4). The FSs,
summarized in Table 1, are categorized in terms of three
different classes: summary statistics, phoneme dependence,
and correlation structure.
Four of the FSs (1, 3, 5, and 7) are effective on short
duration tasks, and are applied to all the recordings in the
dataset. For these FSs, the data is divided into six different
time bins, based on durations of recordings (see Table 2 for
details). Each statistical model is trained only on tasks of
similar duration. The remaining six FSs are designed to
capture longer duration speech dynamics, and are used to
characterize changes across subjects when speaking the same
sentence. For these FSs, the same three declarative sentences
(sentences 2, 4, and 6) are analyzed.
3.2. Summary statistics
FS 1: Delta-MFCC means. Mel-frequency cepstral
coefficients (MFCCs) are obtained with openSMILE at a 100
Hz frame rate [16]. Delta-MFCC coefficients are computed
using regression over two frames before and after each frame.
Then, the mean values of the 16 delta-MFCCs are computed
across all frames in each recording.
FS
Index
Description
Data Types
# Raw
Features
# PCA
Features
1
Delta-MFCC
means
Time bins
16
15
2
Loudness statistics
Sentences
2
1
3
Phn duration
Time bins
1
1
4
Phn dur. & pitch
slope
Sentences
2
2
5
Phn dur. &
formant slopes
Time bins
3
2
6
Phn freq.
Sentences
7
1
7
Waveform corr.
structure
Time bins
600
6
8
Delta-MFCC corr.
structure
Sentences
864 ‒
912
3
9
Formant corr.
structure
Sentences
162 ‒171
3
10
Articulatory
position corr.
structure
Sentences
324 ‒
342
3
FS 2: Loudness statistics. Loudness is computed from the
Perceived Evaluation of Audio Quality (PEAQ) algorithm
[17], a psychoacoustic measure of audio quality that has been
used for analysis of Lombard speech [18]. Loudness is average
energy across critical auditory bands per frame, FS 2
comprises the mean and standard deviation of this feature.
3.3. Phoneme-based features
Changes in phoneme durations and frequencies and in
phoneme-dependent pitch and formant slopes reflect the
phonemic segment dependence of alterations in phonation and
articulation with Parkinson’s severity. Phonemic segments are
used, along with estimated pitch and formant frequency
contours, to generate several phoneme-based feature sets.
Using an automatic phoneme recognition algorithm [19],
phonemic boundaries are detected, with each segment labeled
with one of 40 phoneme classes. The fundamental frequency
(pitch) contour is estimated using an autocorrelation method
over a 40-ms Hanning window every 1 ms [20]. Formant
frequency contours are estimated using a Kalman filter that
smoothly tracks the first three spectral modes while also
smoothly coasting through non-speech regions [21].
For FS 3, 4, and 5, an aggregation step is performed in
which a subset of the 40 phoneme-based measures that are the
most highly correlating with Parkinson’s severity (on the
training set) is linearly combined. In all cases, the top 10 most
highly correlating measures are combined using weights
w=sign(r)/(1-r2) [22]. For measures derived from time-bin
data, the aggregation is done independently in each time bin.
For measures derived from sentences, aggregation is done
independently in each sentence.
FS 3: Average phoneme duration. A linear fit is made of
the logarithm of pitch over time (within each phonemic
segment), yielding a pitch slope (Δlog(Hz)/s) for each
phonemic segment. Phoneme durations are then computed for
those segments where the pitch slope is marked as valid (i.e.,
where the absolute pitch slope is less than eight, indicating that
the slope is likely derived from a continuous pitch contour).
Table 1. Feature Set (FS) summary. Three feature
types (summary statistics, phoneme dependence, and
correlation structure) are applied to time binned and
sentence data.
519
Average phoneme durations were used originally in
classifying depression severity [23].
FS 4: Average phoneme duration and pitch slope. In
addition to the average phoneme duration feature, an average
log-pitch slope is also used where pitch slopes are valid [22].
FS 5: Average phoneme duration and formant slopes.
Linear fits to formant frequencies f1 and f2, along with average
phoneme durations, are computed from all phoneme segments
regardless of pitch slope validity.
FS 6: Phoneme frequencies. The number of occurrences of
each phoneme is computed. These counts are normalized to
sum to one and sorted from highest to lowest. The top seven of
these phoneme frequency measures are used as features.
3.4. Correlation structure features
Measures of the structure of correlations among low-level
speech features have previously been applied in the estimation
of depression [22][24], the estimation of cognitive
performance associated with dementia [25], the detection of
changes in cognitive performance associated with mild
traumatic brain injury [26], and were first introduced for
analysis of EEG signals for epileptic seizure prediction [27].
Channel-delay correlation and covariance matrices are
computed from multiple time series. Each matrix contains
correlation or covariance coefficients between the channels at
multiple time delays. Changes over time in the coupling
strengths among the channel signals cause changes in the
eigenvalue spectra of the channel-delay matrices. The matrices
are computed at four separate time scales, in which successive
time delays correspond to different size frame spacings.
Overall power (logarithm of the trace) and entropy (logarithm
of the determinant) are extracted from the channel-delay
covariance matrices at each scale for FSs 8‒10. A detailed
description of the correlation structure approach can be found
in [27] and its application to speech signals in [22][24].
FS 7: Waveform correlation structure. Correlation
structure features are extracted directly from waveform
segments, thereby characterizing instability in temporal
envelope and phase structure (rhythm and regularity) on
different time scales. This technique was applied to each 0.5 s
frames with 50% overlap. Each frame is divided into five 0.1 s
segments that are treated as separate channels. Utterances ≥
0.75 s in duration resulted in multiple frames. In these cases
the average eigenvalue at each eigenvalue rank is computed
across frames. Features are computed at four scales with delay
spacings of 3, 7, 15, and 31, with 30 delays per scale. These
features reveal an association between Parkinson’s severity
and reduction in dynamical complexity, as illustrated in
Section 3.5.
FS 8: Delta-MFCC correlation structure. Correlation
structure features are derived from all 16 delta-MFCCs using
four delay scales with spacings 1, 3, 7, and 15, and using 15
delays per scale. Fewer delays are used for the largest scale on
sentences 2 and 4 due to their short duration.
FS 9: Formant correlation structure. Correlation structure
features are derived from three formant frequencies using the
same delay and scale parameters as FS 8.
FS 10: Positions of speech articulators correlation
structure. Here, we take advantage of a neurologically
plausible, fMRI-validated computational model of speech
production, the Directions into Velocities of Articulators
(DIVA) model [28]. The DIVA model takes as inputs the first
three formants and the fundamental frequency of a speech
utterance. Then, through an iterative learning process, the
model computes a set of synaptic weights that correspond to
different aspects of the speech production process including
articulatory commands and auditory and somatosensory
feedback errors. We hypothesize that Parkinsonian speech
results from impairments along the speech production
pathway, and therefore, when the model is trained on
Parkinsonian speech, the internal variables will reflect the
severity of the disorder. Correlation structure features are
derived from the DIVA model’s 13 time-varying articulatory
position states, are sampled at 200 Hz. The same delay and
scale parameters are applied as with FS 8 and FS 9.
3.5. Example of discriminative value
In this section we show the discriminative value of the
waveform correlation structure features (FS 7). Figure 1 shows
channel-delay correlation matrices obtained from two women
with low and high Parkinson’s severity (training set files 154
and 303), both speaking the word “crema”. The channels are
five 0.1 s segments of the audio waveform. The matrix
elements are correlation coefficients between channels at
different relative time delays. These matrices are obtained at
the 3rd time scale, and so successive matrix elements
correspond to delays of 15 sub-frames. The matrix from the
low severity speaker exhibits more heterogeneity in the
correlation patterns, indicating higher waveform complexity.
This difference is quantified using matrix eigenvalues
(Figure 2, left panel), ordered largest to smallest. Low UPDRS
speech contains greater power in the small eigenvalues. This
effect is summarized across multiple speakers by plotting the
average eigenvalue for different ranges of Parkinson’s severity
at each rank. Figure 2 (right panel) shows the eigenvalue
averages (in standard units) from all training/development
waveforms < 0.5 s in duration for three UPDRS ranges. These
averages reveal distinct differences related to Parkinson’s
severity, even across multiple different speech tasks.
Figure 1: Correlation matrices from waveform-
segment channel-delay matrices for a low (left) and
high (right) UPDRS from utterance of “crema”.
Figure 2: Left: Eigenspectra for the utterance
“crema” from two female speakers with different
Parkinson’s severity. Right: Average eigenspectra for
low (blue), medium (green), and high (red) UPDRS.
520
4. Parkinson’s score prediction
4.1. Gaussian staircase regression models
There is a separate predictor for each FS, and the predictor
outputs are linearly combined to produce the system’s net
UPDRS prediction. Each predictor uses a Gaussian staircase
(GS) regression model [22][24], which comprises an ensemble
of six Gaussian classifiers, each trained on data with different
UPDRS partitions between Class 1 (lower UPDRS) and Class
2 (higher UPDRS). The Class 1 partitions are 0‒15, 0‒24,
0‒33, 0‒42, 0‒51, and 0‒60, and the Class 2 partitions are
the complement of these partitions. Additional regularization
of the densities is obtained by adding the constant 4 to the
diagonal elements of the data-normalized covariance matrices.
The GS output score is the ratio of the log of the summed
likelihoods from each ensemble of Gaussians. This is followed
by applying a 2nd-order univariate regression model, created
from the GS training scores, to the GS test score to generate a
UPDRS prediction.
4.2. Within-subject prediction averaging
For each FS, subject ID labels are the basis for computing,
across all tasks for that subject, a weighted prediction average
based on the tasks processed by each FS. The provided subject
IDs are used on the training/development sets, and the subject
IDs from our procedure (Section 2.3) are used on the test set.
Time-bin predictions. For FS 1, 3, 5 and 7, there are six
different regression models, one for each time bin. The utility
of each FS in each time bin is assessed based on the Spearman
correlation of its predictions on the training set (using 4-fold
cross validation). Within each FS and subject ID, the predictor
outputs are combined across time bins ID using weights w =
r2/(1-r2) if r > 0; w = 0 otherwise. Table 2 lists the normalized
weights for each FS, with weights summing to one in each
column. Observe that the FSs accumulate most of their
evidence from short duration tasks.
Time bin range (s)
Time bin weights
FS 1
FS 3
FS 5
FS 7
0.0 – 0.5
0.54
0.00
0.00
0.32
0.5 – 0.75
0.03
0.44
0.56
0.00
0.75 – 2.0
0.23
0.56
0.11
0.03
2.0 – 4.0
0.20
0.00
0.01
0.58
4.0 – 7.0
0.00
0.00
0.00
0.01
7.0 –
0.00
0.00
0.32
0.06
Sentence predictions. The remaining FSs use data from
matched sentences (sentences 2, 4 and 6), which range in
duration from 1.56 s to 9.63 s. The sentence tasks allow for
phonemic timing and correlation structure biomarkers that
have been used to predict various neurological states
[11][12][21][22][23]. Within each FS and test subject ID, the
GS training and test scores are averaged across the three
sentences. A 2nd-order univariate regression model obtains a
sentence-based prediction for each FS and test subject.
4.3. Fusing predictors
The ten FS predictors are applied individually to the
training set (four-fold cross-validation with held out speakers)
and to the development set (Table 3). We used linear
combinations of the predictors with three different weight
vectors w1, w2, and w3. With w1, each predictor is weighted
equally. To improve performance, we also applied differential
weighting of the ten predictors. To choose the weights, we
conducted a grid search over all 1,024 possible weight
combinations in which each weight can have a value of 1 or 2.
From this search, we obtained two different weight vectors
based on different constraints. The first weight vector is w2,
the weights that yield the highest average of 1) Spearman
correlation on the training set, 2) Spearman correlation on the
development set, and 3) a test metric score. The second weight
vector is w3, the weights that yield the highest test metric
score. The test metric is the Pearson correlation between
previous submission scores and the Spearman correlations
between the prediction vectors that had produced the previous
scores and a candidate prediction vector. The test metric
equals one if a candidate prediction vector has Spearman
correlation of one with the true UPDRS scores. To compute
this metric, we used ten previously obtained scores, which
range between 0.12 and 0.89. These scores are our previous
eight submissions and two Baseline results (Table 3, rows 1, 3
of [29]). After training on the combined train and development
sets, we submitted test results using w2 and w3, obtaining
Spearman correlations of r = 0.96 and r = 0.97, respectively.
Predictor
Train r
Devel. r
w1
w2
w3
1
0.59
0.39
1
2
1
2
0.29
0.35
1
2
2
3
0.56
0.67
1
1
1
4
0.24
0.37
1
1
1
5
0.46
0.10
1
2
1
6
0.10
0.09
1
1
1
7
0.56
0.53
1
1
2
8
0.43
0.49
1
2
1
9
0.46
0.21
1
1
2
10
0.39
0.25
1
1
1
w1
0.61
0.75
‒
‒
‒
w2
0.62
0.78
‒
‒
‒
w3
0.63
0.70
‒
‒
‒
5. Conclusions and discussion
We applied standard and novel speech features predicting
levels of Parkinson’s disease. Our features capture segment-
based dynamics across phonemes, formant frequencies and
articulatory positions, based on an understanding of the effect
of Parkinson’s disease on speech production components. Our
ongoing work involves the further enhancement of current
features and establishment of new features, with emphasis on
approaches that provide a neurological basis for understanding
the effect of Parkinson’s disease on speech.
6. Acknowledgements
The authors thank Dr. Elizabeth Godoy and Dr. Christopher
Smalt for providing code and guidance on the extraction of
loudness features, and Dr. Douglas Sturim for helpful
discussions on automatic speaker identification.
Table 2. Weights used for combining predictions
across recording duration bins for feature sets (FS).
Table 3. Performance for each feature set predictor
and for linear combinations of predictors.
521
7. References
[1] J. Orozco-Arroyave, J. Arias-Londono, J. Vargas-Bonilla, M.
González-Rátiva, and E. Nöth, “New Spanish speech corpus
database for the analysis of people suffering from Parkinson’s
disease,” in Proceedings of the 9th Language Resources and
Evaluation Conference (LREC), 2014, pp. 342–347.
[2] G. J. Canter, “Speech characteristics of patients with Parkinson’s
disease: I. Intensity, pitch, and duration,” Journal of Speech and
Hearing Disorders, vol. 28, no. 3, pp. 221–229, 1963.
[3] G. J. Canter, “Speech characteristics of patients with Parkinson’s
disease: II. Physiological support for speech.” Journal of Speech
and Hearing Disorders, vol. 30, no. 1, pp. 44–49, 1965.
[4] G. J. Canter, “Speech characteristics of patients with Parkinson’s
disease: III. Articulation, diadochokinesis, and over-all speech
adequacy,” Journal of Speech and Hearing Disorders, vol. 30,
no. 3, pp. 217–224, 1965.
[5] J. A. Logemann, H. B. Fisher, B. Boshes, and E. R. Blonsky,
“Frequency and cooccurrence of vocal tract dysfunctions in the
speech of a large sample of Parkinson patients,” Journal of
Speech and Hearing Disorders, vol. 43, no. 1, pp. 47–57, 1978.
[6] J. E. Sussman and K. Tjaden, “Perceptual measures of speech
from individuals with Parkinson’s disease and multiple sclerosis:
Intelligibility and beyond,” Journal of Speech, Language, and
Hearing Research, vol. 55, no. 4, pp. 1208–1219, 2012.
[7] J. A. Logemann and H. B. Fisher, “Vocal tract control in
Parkinson’s disease,” Journal of Speech and Hearing Disorders,
vol. 46, no. 4, pp. 348–352, 1981.
[8] S. Skodda, and U. Schlegel, “Speech rate and rhythm in
Parkinson’s disease,” Movement Disorders, vol. 23, no. 7, pp.
985–992, 2008.
[9] S. Skodda, “Aspects of speech rate and regularity in Parkinson’s
disease,” Journal of the neurological sciences, vol. 310, no. 1–2,
pp. 231–236, 2011.
[10] S. G. Hoberman, “Speech techniques in aphasia and
Parkinsonism,” Journal - Michigan State Medical Society, vol.
57, no. 12, pp. 1720–1723, 1958.
[11] I. J. Rusz, R. Cmejla, T. Tykalova, H. Ruzickova, J. Klempir, V.
Majerova, J. Picmausova, J. Roth, and E. Ruzicka, “Imprecise
vowel articulation as a potential early marker of Parkinson’s
disease: Effect of speaking task,” The Journal of the Acoustical
Society of America, vol. 134, no. 3, pp. 2171–2181, 2013.
[12] C. G. Goetz, B. C. Tilley, S. R. Shaftman, G. T. Stebbins, S.
Fahn, P. Martinez‐Martin, ... and N. LaPelle, “Movement
Disorder Society–sponsored revision of the Unified Parkinson’s
Disease Rating Scale (MDS‐UPDRS): Scale presentation and
clinimetric testing results,” Movement Disorders, vol. 23, no. 15,
pp. 2129–2170, 2008.
[13] T. F. Quatieri and R. B. Dunn, “Speech enhancement based on
auditory spectral change,” in Proceedings of the IEEE
International Conference on Acoustics, Speech, and Signal
Processing, 2002, pp. I-257.
[14] T. F. Quatieri and R. A. Baxter, “Noise reduction based on
spectral change,” in Proceedings of the IEEE Workshop on
Applications of Signal Processing to Audio and Acoustics
(WASPAA), 1997.
[15] D. A. Reynolds, T. F. Quatieri, and R. B. Dunn, “Speaker
verification using adapted Gaussian mixture models,” Digital
signal processing, vol. 10, no. 1, pp. 19–41, 2000.
[16] F. Eyben, F. Weninger, F. Gross, and B. Schuller, “Recent
developments in openSMILE, the Munich Open-Source
Multimedia Feature Extractor,” in Proceedings of ACM
Multimedia (MM), Barcelona, Spain, 2013, pp. 835–838.
[17] T. Thiede, W. C. Treurniet, R. Bitto, C. Schmidmer, T. Sporer, J.
G. Beerends, and C. Colomes, “PEAQ-The ITU standard for
objective measurement of perceived audio quality,” Journal of
the Audio Engineering Society, vol. 48, no. 1/2, pp. 3–29, 2000.
[18] E. Godoy and Y. Stylianou, “Unsupervised acoustic analyses of
normal and Lombard speech, with spectral envelope
transformation to improve intelligibility,” in 13th Annual
Conference of the International Speech Communication
Association, September 9–13, Portland, Oregon, Proceedings,
2012, pp. 1472–1475.
[19] W. Shen, C. White, T. J. Hazen, “A comparison of query-by-
example methods for spoken term detection,” in Proceedings of
the IEEE International Conference on Acoustics Speech and
Signal Processing, 2010.
[20] P. Boersma and D. Weenink, “Praat, a system for doing
phonetics by computer,” 2001.
[21] D. D. Mehta, D. Rudoy, and P. J. Wolfe, “Kalman-based
autoregressive moving average modeling and inference for
formant and antiformant tracking,” The Journal of the Acoustical
Society of America, vol. 132, no. 3, pp. 1732–1746, 2012.
[22] J. R. Williamson, T. F. Quatieri, B. S. Helfer, G. Ciccarelli, and
D. D. Mehta, “Vocal and facial biomarkers of depression based
on motor incoordination and timing,” in Proceedings of the 4th
ACM International Workshop on Audio/Visual Emotion
Challenge (AVEC), 2014, pp. 65–72.
[23] A. Trevino, T. F. Quatieri, and N. Malyska, “Phonologically-
based biomarkers for major depressive disorder,” EURASIP
Journal on Advances in Signal Processing, vol. 42, pp. 1–18,
2011.
[24] J. R. Williamson, T. F. Quatieri, B. S. Helfer, R. Horwitz, B. Yu,
and D. D. Mehta, “Vocal biomarkers of depression based on
motor incoordination,” in Proceedings of the 3rd ACM
International Workshop on Audio/Visual Emotion Challenge,
2013, pp. 41–48.
[25] B. Yu, T. F. Quatieri, J. W. Williamson, and J. Mundt,
“Prediction of cognitive performance in an animal fluency task
based on rate and articulatory markers,” in 15th Annual
Conference of the International Speech Communication
Association, September 9–13, Portland, Oregon, Proceedings,
2014.
[26] B. S. Helfer, T. F. Quatieri, J. R. Williamson, L. Keyes, B.
Evans, W. N. Greene, J. Palmer, and K. Heaton, “Articulatory
dynamics and coordination in classifying cognitive change with
preclinical mTBI,” in 15th Annual Conference of the
International Speech Communication Association, September 9–
13, Portland, Oregon, Proceedings, 2014.
[27] J. R. Williamson, D. Bliss, D. W. Browne, and J. T. Narayanan,
“Seizure prediction using EEG spatiotemporal correlation
structure,” Epilepsy and Behavior, vol. 25, no. 2, pp. 230–238,
2012.
[28] F. H. Guenther, S. S. Ghosh, and J. A. Tourville, “Neural
modeling and imaging of the cortical interactions underlying
syllable production,” Brain and Language, vol. 96, no. 3, pp.
280–301, 2006.
[29] B. Schuller, S. Steidl, A. Batliner, S. Hantke, F. Hönig, J. R.
Orozco-Arroyave, E. Nöth, Y. Zhang, F. Weninger, “The
INTERSPEECH 2015 Computational Paralinguistics Challenge:
Nativeness, Parkinson’s & Eating Condition,” in
INTERSPEECH 2015 – 16th Annual Conference of the
International Speech Communication Association, September 6–
10, Dresden, Germany, Proceedings, 2015.
522
