Voice quality analysis to detect neurological diseases

Abstract

Neurological degenerative diseases are becoming a growing concern in modern society. The successful treatment of these diseases depend greatly in early detection. Speech has been routinely used by specialists as a valuable correlate in the assessment of pathological disease. Specifically voicing can serve as a very introspective correlate for this practice. The present paper uses a methodology previously employed in organic pathology voice quality assessment to explore to what extent specific low-level correlates of neurological diseases may be established. The methodology uses voiced recordings of sustained vowels to estimate vocal fold visco-elastic parameters from inverse filtering. These parameters show to be clearly influenced by unstable neuronal spiking resulting in tremor which affects many phonation cycles. The possible modeling of tremor could be used as an index to neuro-motor problems in phonation and help in differential diagnose of the pathology at an early stage. The paper presents examples on parameter estimations from study cases of spasmodic dysphonia and Parkinson Disease. Further development of research lines on this estimation methodology is also addressed.

Full text

Draft of the publication presented at MABEVA 2013, Florence, Italy, August 25-27
Copy for personal non-profit use only. Copyright holder: Universitá di Firenze
Abstract: Neurological degenerative diseases are
becoming a growing concern in modern society. The
successful treatment of these diseases depend greatly
in early detection. Speech has been routinely used by
specialists as a valuable correlate in the assessment of
pathological disease. Specifically voicing can serve as
a very introspective correlate for this practice. The
present paper uses a methodology previously
employed in organic pathology voice quality
assessment to explore to what extent specific low-level
correlates of neurological diseases may be established.
The methodology uses voiced recordings of sustained
vowels to estimate vocal fold visco-elastic parameters
from inverse filtering. These parameters show to be
clearly influenced by unstable neuronal spiking
resulting in tremor which affects many phonation
cycles. The possible modeling of tremor could be used
as an index to neuro-motor problems in phonation
and help in differential diagnose of the pathology at
an early stage. The paper presents examples on
parameter estimations from study cases of spasmodic
dysphonia and Parkinson Disease. Further
development of research lines on this estimation
methodology is also addressed.
Keywords: Inverse Filtering, Vocal Fold Biomechanics,
Parkinson Disease, Voice Quality Assessment, Tremor
I. INTRODUCTION
Classically Voice Quality Analysis has been focused to
detect and establish the organic pathology in voice
resulting from pathological alterations of larynx
physiology. The study of other sources of dysphonic
voice finding their ultimate reasons in the alterations of
the neurological paths controlling phonation have been
tagged as "functional" or "non-organic". Voice resulting
from altered phonation due to neurological reasons may
be a most valuable report of the etiology and progress of
neural diseases affecting the production of voice, such as
pathologies resulting in voice tremor [1]. These would
include some kinds of spasmodic dysphonia, stammering
and Parkinson. The possibility of early detection in the
first stages of Parkinson's Disease (PD) may grant a
better preventive treatment reducing the progress of the
illness [2]. Monitoring treatment by objective methods is
also an important goal, especially in modifying or
defining new protocols. The deepest foundations of the
methodology proposed in this paper are to be found in
tracking the malfunctioning of neurological and
neuromuscular paths involved in voice production (see
Fig. 1).
Fig. 1. Simplified view of main neural pathways involved in
the production of phonation: 1. Links from linguistic
neuromotor cortex to Basal Ganglion relay stages. 2. Branch of
the X nerve acting on the naso-pharingeal switch. 3. Idem acting
on the retro-lingual switch connected to the epiglottal switch. 4.
Branch of the laryngeal nerve acting on the transversal and
oblique arytenoid and cricothyroid muscles responsible for the
VOICE QUALITY ANALYSIS TO DETECT NEUROLOGICAL DISEASES
1P. Gómez, 1V. Rodellar, 1V. Nieto, 1L. M. Mazaira, 1C. Muñoz, 2M. Fernández, 2E. Toribio
Grupo de Informática Aplicada al Procesado de Señal e Imagen, Universidad Politécnica de Madrid
1Campus de Montegancedo, s/n, 28660 Boadilla del Monte, Madrid, Spain
Tel.: +34913367384, fax: +34913366601, e-mail: pedro@pino.datsi.fi.upm.es
2ENT and Neurology Services, Hospital del Henares, Avda. Marie Curie s/n, 28822 Coslada, Madrid, Spain

Draft of the publication presented at MABEVA 2013, Florence, Italy, August 25-27
Copy for personal non-profit use only. Copyright holder: Universitá di Firenze
vocal fold adduction and abduction. 5. Branch of the vagus
nerve (phrenic) actuating on the diaphragmatic muscles. 6.
Feedback loop in Basal Ganglia damping muscular tone.
These comprise links from the neuromotor linguistic
cortex [3] to the subthalamic region [4] and through the
laryngeal nerve and their associated pathways [5][6] to
the muscles activating the thyro-arytenoid structure,
responsible in the last term of vocal fold stretching,
adduction and abduction (Superior Laryngeal Nerve,
Internal and External Laryngeal Branches of the Inferior
Laryngeal Nerve, Transverse and Oblique Arytenoid
Muscles -TOAM-, and Cricothyroid Muscles -CM). Any
alteration in the functionality of these pathways and in the
associated muscles will result in temporary distortions of
the parameters of tension and dynamic mass contribution
of the vocal folds, both on the body and the cover
biomechanics. Correlates of these alterations will be
found in the pitch, and in long term jitter and shimmer, as
the periodicity of these alterations may be of hundreds of
milliseconds [7]. The aim of this paper is to give some
phenomenological account in detecting and grading the
neurological disease using biomechanical correlates
obtained from the inverse filtering of voice. The
technology has been tested in monitoring pre-post
treatment of organic pathology, and due to its ubiquitous
character can be applied as well to the neurological
disease.
II. METHODOS
A database of voice recordings from neurological
disease-affected patients is being recorded in Hospital del
Henares of Madrid. This geographical area South East of
Madrid is specially sensitive to PD. Being a heavy
industrial area it is believed that some environmental
factors may be responsible of the largest incidence of PD
among the aging population compared to other regions of
Madrid. For the preliminary and explorative character of
the present study some specific cases are selected, these
being strong spasmodic and PD voice samples,
pathological voice of organic origin and voice from
normophonic patients to serve as a contrast (all of them
females). These voices are inverse filtered and some
biometrical and biomechanical parameters are estimated,
as the glottal closure sharpness, the mucosal/average
ratio, the first two cepstral coefficients of the glottal
source power spectral density, and the tension of the
vocal fold body. It may be shown that these indices show
a strong correlation with the spasmodic episodes both in
their timely evolution and statistical dispersion. The
methodology is based on the following steps:
1. Three emissions of the vowel /a/ are recorded at
44,100 Hz under normal phonation conditions.
2. For specific statistical comparison they are low-pass
filtered and re-sampled to 22,050 Hz. High-pass
filtering at 25Hz is also applied to eliminate low
frequency flickering effects. Frames of 0.4 s long are
used in the analysis.
3. Inverse Filtering is applied, and the glottal source is
reconstructed [8].
Estimations of the glottal closure sharpness,
noise/glottal ratio, dynamic mass and tension of the vocal
fold body are derived following [8].
III. RESULTS
An episode of spasmodic dysphonia (SD) has been
selected from the database to show the possibilities of the
methodology, corresponding to a female voice (32 year
old) manifesting about 2-3 spasms per second. The record
is a segment of 0.4 s long from a sustained phonation of
vowel /a/ (see Fig. 2 and Fig. 3).
Fig. 2. Episode of spasmodic dysphonia. Templates from top to
bottom: Voice signal. Inverse filtering residual. Glottal source.
Glottal flow.
Fig. 3. Left templates from top to bottom: Phonation cycle-
synchronous estimates of the dynamic mass component of the
vocal folds, friction losses and body stiffness. Right templates
from top to bottom: Statistical distributions of the left templates
given as boxplots.

Draft of the publication presented at MABEVA 2013, Florence, Italy, August 25-27
Copy for personal non-profit use only. Copyright holder: Universitá di Firenze
The voice segment studied is a part of a recording of a
sustained /a/ 0.4 s long where an episode of spasm is
clearly recognizable by the amplitude decay. The
reconstruction of the glottal source (template c) does not
show such a strong decay in amplitude. Pitch ranges from
208-203 Hz in the sections out of the spasm to a
minimum of 185 Hz during the spasm, following an
almost regular fluctuation (tremor of about 2.5 Hz). It
may be seen that the estimates of the dynamic mass of the
vocal fold body, and especially the fold tension are highly
correlated with the spasm, reporting changes of about
25% and 50% of variation respectively. Similar
fluctuations are found in other distortion parameters, such
as the sharpness of the closure spike in the glottal source,
the noise/glottal energy ratio and some cepstral
parameters of the glottal source spectral density.
Fig. 4. Phonation 0.4 s long from a patient affected from
Parkinson Disease. Templates from top to bottom: Voice signal.
Inverse filtering residual. Glottal source. Glottal flow.
Fig. 5. Left templates from top to bottom: Phonation cycle-
synchronous estimates of the dynamic mass of the vocal folds,
friction losses and body stiffness. Right templates from top to
bottom: Statistical distributions of the left templates given as
boxplots.
A second example from a patient (72 year old) affected
by Parkinson Disease (PD) corresponding also to female
voice has been analyzed following the same
methodology. The record is a segment of 0.4 s long from
a sustained phonation of vowel /a/. The results of the
analysis are reported in Fig. 4 and Fig. 5. In this case the
changes in amplitude are not as relevant as in the
spasmodic case. The reconstruction of the glottal source
(c) does not show important changes in amplitude as
well. Pitch ranges from 240-256 Hz following an
irregular fluctuation (tremor) of about 5 Hz. The
estimates of the vocal fold body dynamic mass and
stiffness report changes of around 20% . To put the
analysis into context at this point it would be worth to
compare some overall results for these two cases against
results from a normal female speaker and a pathological
female speaker. The normal speaker (NF) is a 34 year old
female, non-smoker not having reported any problem
with voice, volunteering for the study. Normal condition
was assessed by endoscopy and EGG. The case with
organic pathology corresponded to a female 22 year old
having been diagnosed from a left vocal fold cyst (LVFC)
affecting the contralateral fold (contact lesion). The case
was graded 2 (severe) in GRBAS scale. Endoscopy and
EGG availed the diagnose. The acoustic processing of the
four cases included the extraction of pitch, relative jitter
and shimmer and the noise/glottal energy ratio (NGE).
The stiffness of the vocal fold body was estimated as
well. The results are given in Table 1 at the end of the
paper.
IV. DISCUSSION
From the results in Table 1 the first consequence is that
pathological data (except for PD) are clearly
differentiated from normal data in the value of the
dispersion (standard deviation) and in the stiffness of the
vocal fold body. Mean values of the classical distortion
estimates as jitter, shimmer or NGE do not show
important differences among the pathological cases
except in PD. This case shows distortion parameters
which could be considered normal. The problem is that
tremor in PD is observed as FM-modulations which do
not leave clues in the jitter, whereas the SD case may be
traced in shimmer. Going to the causes, it seems that the
effects of FM-affected spiking producing tremor in SD
may be observed on the specific muscles affecting vocal
fold abduction and adduction (TOAM-CM) as well as in
the muscles responsible for pressure build-up and
sustenance in lungs during phonation (diaphragm). The
influence of FM-affected spiking in the modulation of the
vocal tract (naso-velar switch, glossomuscular and oro-
labial complexes) could also introduce changes in the
production of voice, interfering with vocal-fold induced
tremor. These differences may affect the results observed,
as in the two cases studied. In the spasmodic case the

Draft of the publication presented at MABEVA 2013, Florence, Italy, August 25-27
Copy for personal non-profit use only. Copyright holder: Universitá di Firenze
important changes in amplitude observed could be
associated with some influence of the spasm on the
diaphragm and other muscles inducing subglottal
pressure, besides affecting strongly to the vocal fold
stiffness as a result of the TOAM-CM action. The result
during the spasm is a dystonic relaxation of the vocal
folds (abduction) accompanied by a decay in subglottal
pressure. The case of the PD patient may have to see only
with the action of the TOAM-CM, resulting in a
relatively cyclic dystonic behavior of the vocal fold but
not in important changes of the subglottal pressure. It
seems that parameters tracking amplitude changes as
shimmer or APPQ measured directly on the glottal
source, as well as the indirect estimates of vocal fold
tension may serve as important marks to produce
differential diagnose in tremor-affected dysphonias, and
this line should be further studied. Other possible
correlates could be the sharpness of the closure instant
and the lowest cepstral coefficientes of the glottal source
spectral profile. This means that the study of tremor as a
result of neurodegenerative diseases may require complex
time-frequency analytical techniques. Chaotic modeling
of tremor in stiffness and other correlates, and Wavelet
Transform may be good candidates out these studies.
V. CONCLUSIONS
The first conclusion from this phenomenological
description is that tremor appears as a mark in certain
biomechanical estimates of vocal fold dynamics as body
stiffness. Therefore the monitoring and modeling of
tremor could be based on the study of these correlates.
Indications that differential diagnose could also be based
in combined amplitude-stiffness indices are plausible
enough for the issue to deserve further study. The
analysis of the mentioned correlates estimated directly
from the glottal source obtained after vocal tract
inversion instead of whole voice may be a beneficial
methodology to unveil and quantize the extent or degree
of the spasmodic or tremor illness. As the
characterization of tremor in voice shows quasi-cyclic
information, techniques to model this characteristic as
chaotic attractors, wavelets, or ARMA coefficients may
be of much higher resolution than the analysis of full
voice. The monitoring of neurological diseases is of most
importance in a world where the aging of general
population will demand important resources for health
care. The early detection and monitoring of these
problems may help in devising more efficient treatment
protocols. Routine voice tests may help in this task. The
validation of this methodology for PD is in due course in
cooperation with the ENT and Neurology Services at
Hospital del Henares.
REFERENCES
[1] Pantazis, Y., Koutsogiannaki, M., Stylianou, Y., "A
novel method for the extraction of tremor", Proc. of
MAVEBA07, Florence University Press 2009, pp.
107-110.
[2] Das, R. “A comparison of multiple classification
methods for diagnosis of Parkinson disease”, Expert
Systems with Applications, Vol. 37, 2010, 1568-
1572.
[3] Rauschecker, J. P., Scott, S. K., "Maps and streams
in the auditory cortex: nonhuman primates illuminate
human speech processing", Nature Neuroscience,
Vol 12, 2009, pp. 718-724.
[4] Törnqvist, A. L., Schalén, L., Rehncrona, S., "Effects
of different electrical parameter settings on the
intelligibility of speech in patients with Parkinson's
Disease treated with subthalamic deep brain
stimulation", Mov. Disord., Vol. 20, 2004, pp. 416-
423.
[5] Eckley, C. A., Sataloff, R. T., Hawkshaw, M.,
Spiegel, J. R., Mandel, S., "Voice range in superior
laryngeal nerve paresis and paralysis", J. Voice, Vol.
12, 1998, pp. 340-348.
[6] Luschei, E. S., Ramig, L. O., Baker, K. L., Smith, M.
E., "Discharge characteristics of laryngeal single
motor units during phonation in young and older
adults and in persons with Parkinson disease", J.
Neurophysiol., Vol. 81, 1999, pp. 2131-2139.
[7] Tsanas, A., Little, M. A., McSharry, P. E., Ramig, L.
O., "Accurate telemonitoring of Parkinson's disease
progression by non-invasive speech tests", IEEE
Trans. Biomed. Eng., Vol. 57, 2009, pp. 884-893.
[8] Gómez, P., Fernández, R., Rodellar, V., Nieto, V.,
Álvarez, A., Mazaira, L. M., Martínez, R, Godino, J.
I., “Glottal Source Biometrical Signature for Voice
Pathology Detection”, Speech Communication, Vol.
51, 2009, pp. 759-781.
Table 1. Parameter values for the four cases studied (standard deviations between parentheses)
Subject/Parameter
Pitch Hz
Jitter %
Shim %
NGE %
Stiffness (g.s-2)
#346 (34y NF)
199 (1.15)
0.7 (0.6)
1.9 (1.3)
8.4 (0.5)
19542 (185)
#341 (22y LVFC)
215 (7.04)
4.2 (3.6)
3.8 (2.4)
6.6 (1.9)
24857 (4487)
#308 (45y SD)
199 (6.02)
1.5 (1.5)
6.5 (3.3)
8.9 (1.7)
22168 (2656)
#337523 (72y PD)
248 (3.87)
0.7 (0.5)
1.4 (1.0)
6.5 (1.1)
25138 (988)