Glottal Flow Analysis in Parkinsonian Speech

Patrick Corcoran, Arnold Hensman and Barry Kirkpatrick

TU Dublin, Blanchardstown Campus, Dublin, Ireland

Keywords: Glottal Flow, Parkinson’s Disease, Speech.

Abstract: Speech and vocal impairments are one of the earliest symptoms of Parkinson’s disease (PD). Laryngoscope

examinations have identified that patients with the disease show pathological behaviour of the vocal folds.

The behaviour of the vocal folds is investigated by analysing the glottal flow waveform in Parkinsonian

speech in this study. This study aims to determine the appropriate method for estimating the glottal source in

PD speech and to identify glottal parameters that could be indicative of PD. An experiment was conducted to

analyse a selection of glottal parameters (2 time-domain and 3 frequency-domain) measured from the glottal

flow waveform estimated from speech recordings. A database of 52 healthy speakers and 44 speakers with

Parkinson’s disease was considered for this experiment. Two glottal estimation techniques are considered in

the experiment: iterative and adaptive inverse filtering (IAIF) and quasi-closed phase (QCP) inverse filtering.

The results showed that 2 of the 5 glottal parameters (1 time domain and 1 frequency domain) produced values

indicating a difference between healthy and PD speech files in the database. The results also indicate that

glottal estimates from the IAIF method resulted in parameters discriminating between healthy and PD higher

than glottal estimates from the QCP method.

1 INTRODUCTION

Parkinson’s disease (PD) is a chronic

neurodegenerative disorder of the central nervous

system generally observed in elderly people. It is the

second most common neurodegenerative disease,

after Alzheimer’s, affecting an estimated 10 million

people around the world, with these numbers

expected to double in the next 10 years (Dorsey et al.,

2007). Currently there is no cure for PD but early

diagnosis and drug therapies can decrease the

difficulties of the disorder and improve quality of life.

This study aims to investigate the appropriate method

for estimating the glottal flow in PD speech and

identify glottal parameters that could be indicative of

PD by analysing the behaviour of the glottal flow.

The cause of PD is attributed to the progressive

loss of dopamine in the brain which is the chemical

released by nerve cells to interact with other nerve

cells. This interaction between nerve cells is

responsible for controlling the motor and mental

functions of a person, and the reduction in dopamine

levels leads to PD symptoms. Typical motor

symptoms observed in PD are muscular rigidity,

resting tremor and slowness of movement. Along

with these, many patients develop non-motor

symptoms like sustained depression and memory

loss. Individuals with PD experience different

combinations of these symptoms at different severity

levels. The muscles in the face, mouth and throat can

be affected which results in problems with speech and

swallowing. It is estimated that 89% of PD patients

will suffer some form of vocal impairment (Logeman

et al., 1978) and a vocal disorder may be one of the

earliest symptoms of the disease (Harel et al, 2004).

Speech related symptoms that have been reported to

affect PD patients include harsh or breathy voice,

reduced volume and vocal tremor. The most

commonly used scale for the progression of PD is the

Unified Parkinson’s Disease Rating Scale (UPDRS).

Employing the UPDRS is a complex and lengthy

procedure which requires the subjective evaluation by

a clinical expert. Analysing the speech signal in PD

patients may provide a tool to help clinicians evaluate

and diagnose the disease. This could provide a non-

invasive method of indirectly examining the larynx

which may help with further monitoring of the

disease and could be performed remotely.

Previous studies of the larynx in PD patients have

shown incidences of abnormalities of the vocal folds.

These laryngeal dysfunctions have been observed

through laryngoscope examinations where video

116

Corcoran, P., Hensman, A. and Kirkpatrick, B.

Glottal Flow Analysis in Parkinsonian Speech.

DOI: 10.5220/0007259701160123

In Proceedings of the 12th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2019), pages 116-123

ISBN: 978-989-758-353-7

frames of the vocal folds are obtained and analysed

by a clinical expert. Hanson et al (1984) examined 32

PD patients and reported that 94% showed vocal fold

bowing and 81% demonstrated varying degrees of

asymmetry of the vocal folds. Smith et al (1995)

reported that from videostroboscopic examinations of

22 patients, there was a 38% incidence of vocal-fold

bowing and 67% incidence of incomplete glottal

closure. Perez et al (1996) reported irregularities in

the closure and vibration of the vocal folds with 50%

of the patients demonstrating abnormal glottal closure

and 47% demonstrating irregular vibration of the

vocal folds with asymmetric behaviour. Yücetürk et

al. (2002) examined 30 PD patients and reported that

70% had at least one of eight laryngeal dysfunctions

with some of the patients featuring more than one.

Tsuboi et al. (2015) reported that of 22 PD patients

treated with subthalamic nucleus deep brain

stimulation 77% showed an incidence of incomplete

glottal closure and 50% showed signs of

asymmetrical glottal movement.

Research is ongoing in the studies of diagnosing

and monitoring PD with speech and is showing

positive steps towards establishing an objective

measurement of the disease (Little et al., 2009).

Studies implementing advanced signal processing

algorithms have shown that symptoms of PD can be

predicted on the UPDRS scale remotely using non-

invasive speech recordings (Tsanas et al., 2010).

Speech impairments of PD patients were investigated

by features such as jitter, shimmer and harmonic to

noise ratio (Tsanas et al, 2012). Results obtained from

these parameters showed accuracies of up to 98%.

Sharma (2014) also reported jitter and shimmer

showing different values, when tested on 14 PD and

7 healthy subjects. This study also reported the glottal

pulse of the healthy subjects to be symmetric in nature

when compared to the PD patients. The behaviour of

the glottal flow in PD patients has been studied and

parameters have been identified that discriminate

from healthy speakers with accuracies of over 90%

(Hanratty et al., 2016). Additionally, automatic

detection of PD has been researched by analysing the

non-linear behaviour of the vocal folds which affects

the glottal flow signal with accuracies up to 78%

(Belalcázar-Bolaños et al., 2016). Detection of early

stages of Parkinson’s disease using Mel-frequency

cepstral coefficients was investigated by (Jeancolas et

al., 2017), employing a detection framework similar

to that used in speaker recognition and obtained

results between 60% and 91%. It is difficult to

compare results between studies as they have used

different performance metrics on different test

databases and recording protocols.

The aim of this study is to build on previous

studies and contribute to the research of using speech

files to aid in the diagnosis of PD. This will be

achieved by analysing the behaviour of the glottal

flow waveform in PD speech and identifying glottal

parameters which are distinct to healthy speech. The

glottal flow will be estimated from speech signals by

different methods to identify which is the most

applicable to extracting the glottal flow estimate in

Parkinsonian speech.

The rest of this paper is organised as follows

Section 1.1 describes the background on the glottal

flow, glottal estimation techniques and glottal

parameters. Section 2 describes the experimental

procedure and details the data used in the experiment,

Section 3 presents the results of the experiment, and

Section 4 comprises the conclusions of the study.

1.1 Background

The glottal flow is the airflow that is generated from

the lungs and then passed through the vocal folds,

located in the larynx. The vocal folds vibrate which

causes them to open and close periodically. This

airflow is filtered by the vocal tract cavities to

produce human speech (Quatieri, 2006). The glottal

flow waveform is produced from this airflow and is

depicted in Figure 1 (Drugman et al., 2012).

Figure 1: Glottal flow waveform (upper) and glottal flow

derivative waveform (lower) with open phase and closed

phase displayed.

Each period of the glottal flow waveform can be

separated into three main parts, the open phase, the

return phase and the closed phase. During the open

phase, the air pressure gradually increases until it

comes to an abrupt stop when the glottis closes, which

is called the glottal closure instant (GCI; Drugman

and Dutoit, 2009). In healthy speech, during the

Glottal Flow Analysis in Parkinsonian Speech

117

closed phase, there is no air flow through the vocal

folds and the amplitude of the signal has returned to

zero. The open phase of the glottal flow waveform is

divided into two phases, the opening phase and the

closing phase. The opening phase refers to the

timespan up to the maximum positive amplitude of

the glottal pulse, while the closing phase is the period

after this until the GCI. After the open phase, there is

a period where the waveform returns to the initial

state, this is called the return phase (Drugman et al.,

2012). The glottal flow can be represented as a glottal

flow derivative waveform, as shown in Figure 1, as it

reflects some characteristics that are not represented

in the glottal flow waveform (Plumpe et al., 1999).

1.1.1 Estimating the Glottal Source

A speech signal can be represented as being made up

of two main components, the glottal flow (source) and

the vocal tract (filter) (Fant, 1971). Glottal inverse

filtering (GIF) is a technique used to estimate the

glottal flow waveform from a speech signal. The idea

of GIF is to estimate a model for the vocal tract filter

from a recorded speech signal and then filter the

recorded signal through the inverse of this model to

cancel the effects, resulting in an estimate of the

glottal flow signal (Alku, 2011). Modern GIF

methods can be categorised as (1) closed-phase

methods (2) iterative methods and (3) spectral

decomposition methods. Closed-phase methods use

the closed-phase of the glottal flow signal as there is

said to be less interaction from the vocal tract and it

provides a more accurate model of the vocal tract,

resulting in more accurate glottal estimates (Wong et

al., 1979). Iterative methods utilise the whole pitch

period to remove the influence of the glottal

waveform and estimate the vocal tract. This vocal

tract estimate is then used by inverse filtering to

provide an estimate of the glottal flow (Alku, 1992).

Spectral decomposition methods involve estimating

the glottal flow by separating the speech by maximum

and minimum phase components (Alku, 2011). All

methods except iterative methods require accurate

identification of glottal closure instants (GCI) and

glottal opening instants (GOI; Drugman and Dutoit,

2009). Most studies when evaluating GIF methods

will use synthetic speech because the glottal flow

signal cannot be measured directly from the human

larynx (Airaksinen, 2014). A recent study (Chien et

al., 2017) has shown that closed-phase and iterative

methods perform well and show stability on different

voice qualities of sustained synthetic vowels, while

spectral decomposition methods provided a less

stable performance on the tested database. Breathy

voice quality is one of the reported speech disorders

of Parkinson’s disease and this paper reported that

Figure 2: Glottal flow pulse (top) and glottal flow derivative

pulse (bottom) with measurements for quasi-open quotient

(QOQ) and normalised amplitude quotient (NAQ).

closed-phase and iterative methods display

robustness on this voice quality across a number of

synthetic vowels.

1.1.2 Glottal Flow Parameters

Many parametrisations of the glottal flow exist but

not all are suitable for Parkinsonian speech.

Parkinsonian speech is known to show harsh and

breathy characteristics among other pathologies. This

indicates that the glottal parameters must be robust to

noise for effective measurement. This section gives

an overview of the glottal parameters selected to be

considered in this study.

The time-domain parameters, Quasi-open

quotient (QOQ) and normalised amplitude quotient

(NAQ), were selected as they are known to be robust

measurements of the glottal waveform in adverse

conditions. Previous studies have reported that they

show potential for separating PD and healthy speech

(Hanratty et al., 2016).

Quasi-open Quotient (QOQ): This parameter

measures the duration of the open phase from when

the amplitude of the glottal pulse crosses the 50%

marker line at the point, t

until it falls below it again

at t

as shown in Figure 2. The marker t

is defined

as the point at which the amplitude of the glottal pulse

reaches 50% of its maximum value and t

marks the

point at which the glottal pulse goes below 50% of the

maximum value. The timing distance between these

two points is referred to as the quasi-open phase. This

duration is subsequently normalised with respect to

the pitch period, T

. It was designed as a more robust

version of the open quotient (OQ) (Airas, 2008; Kane,

BIOSIGNALS 2019 - 12th International Conference on Bio-inspired Systems and Signal Processing

118

2012) and mitigates the issue of noise corrupting the

measurement of the instant of glottal opening. The

formula for QOQ is displayed in Equation 1 (Hacki,

1989).

QOQ  



t





(1)

Normalised Amplitude Quotient (NAQ): This

parameter is measured from the peak amplitude of the

glottal flow, E

and maximum negative amplitude of

the glottal flow derivative, E

, as shown in Figure 2.

This is then normalised with respect to the pitch

period, T

, as shown in Equation 2 (Alku et al., 2002).

This is selected as a glottal parameter as it

representative of the glottal pulse and glottal

derivative pulse. It is robust to variations in recording

conditions as it is normalised with respect to

amplitude.

NAQ  





(2)

Frequency-domain parameters, H1-H2, harmonic

richness factor (HRF) and parabolic spectral

parameter (PSP), were selected for this study.

Measurements in the glottal source spectrum give an

alternative approach to the time-domain parameters

and they are known to distinguish between voice

qualities (Alku, 2011). Two of these glottal

parameters, H1-H2 and HRF, have been tested in

combination with time-domain parameters on

Parkinsonian and healthy speech (Belalcázar-Bolaños

et al., 2016). Discrimination between healthy speech

and Parkinsonian speech was made using a Support

Vector Machine (SVM) and produced accuracies of

up to 78%. These parameters were selected to

determine their individual performance in detecting

PD.

H1-H2: This is a measure of the change in amplitudes

of the first two harmonics, H1 and H2, of the

differentiated glottal source spectrum (Fant, 1995).

This measurement has been used as a glottal

parameter as it is reported that changes in the open-

quotient of the glottal cycle produce a corresponding

change in H1-H2 (Doval et al., 2006). This has been

used to detect different phonation types by analysing

the measurement.

Harmonic Richness Factor (HRF): This spectral

parameter is a measurement computed by the sum of

the amplitudes of the harmonics above the

fundamental harmonic. This is then normalised with

respect to the first harmonic, H

, and is shown in

Equation 3 (Childers and Lee, 1991). This parameter

represents the spectral tilt of the glottal flow and has

been used to identify different phonation types.

HRF 

∑







(3)

Parabolic Spectral Parameter (PSP): This is a

measure to model frequency domain characteristics

within the glottal signal. It is computed by fitting a

parabola to the lower frequencies in the glottal source

spectrum (Alku et al., 1997). This parameter was

introduced as a robust measurement of the spectral

decay in the glottal signal to detect phonation type.

2 EXPERIMENTAL PROCEDURE

The objective of this experiment was to identify the

appropriate method for estimating the glottal source

in Parkinsonian speech. This was completed by

analysing the glottal signal from PD and healthy

speech recordings using different estimation

techniques and identifying parameters that behave

different. These parameters would then be tested to

quantify if a separation exists between PD and heathy

speech.

The performance of the parameters was quantified

using receiver operating characteristic (ROC) curves

and the area under the ROC curve (AUC) (Fawcett,

2006). The ROC curve and AUC value quantify the

performance of the glottal parameters in their task to

separate between PD and healthy speech. The AUC

value can range from 0 and 1 and it can be interpreted

as the probability of making the correct decision on

classifying a particular file correctly. An AUC value

of 0.5 indicates no separation.

2.1 Data

The data used in the experiment was taken from three

components to create one database, containing

healthy and Parkinsonian speech files.

2.1.1 Parkinsonian Speech

The Parkinson’s disease speech recordings consisted

of a combination of two databases from different

sources.

The first database was recorded in a quiet

environment with a Zoom H2n recorder at St. Mary’s

Hospital in Dublin, Ireland as reported in (Hanratty et

al., 2016). The signals were sampled at 44.1 kHz per

channel with a 16 bit resolution. The database

contains 22 Parkinson’s disease patients who were

asked to make a sustained sound of the vowel ‘a’ for

as long as possible, therefore the signals are of

varying durations.

Glottal Flow Analysis in Parkinsonian Speech

119

The second Parkinson’s disease speech database

was recorded by a Trust MC-1500 microphone placed

10 cm from the speaker’s lips as reported in (Sakar et

al., 2013). The signals were sampled at 44.1 kHz per

channel with a 16 bit resolution. The database

contains 28 Parkinson’s disease patients with an age

range from 39 – 79 and who are suffering with the

disease for 0 – 13 years. The patients recorded

sustained vowels ‘a’ and ‘o’ three times with varying

durations. For this study, the ‘a’ sounds were taken

from this database to correspond with the recordings

taken from the previous database.

2.1.2 Healthy Speech

The healthy speech database was obtained from

(Childers, 1999). This database was recorded in a

professional single-wall sound room with an Electro-

Voice RE-10 cardioid microphone. The microphone

was placed 15 cm from the speaker’s lips and the

signals were sampled at 10 kHz per channel with a

16-bit resolution. The database contains 52 subjects

(25 male and 27 female) with a normal larynx and an

age range from 20 – 80 years old. All subjects

recorded 28 tasks which included 12 sustained

different vowel sounds with a duration of

approximated 2 seconds. This database also included

full words and spoken sentences from the speakers.

For this study, the sustained vowel ‘aa’ was taken

from this database to be consistent with the

Parkinsonian database.

2.1.3 Glottal Estimation from Database

For this experiment all speech recordings with a

sustained vowel ‘a’ were investigated. As all speech

recordings were of various durations, a window of

500ms of continuous voiced speech was extracted

from the centre section of each recording. This also

ensured there was no transient effects included for the

glottal analysis. 5 of the 22 speech files from the

source (Hanratty et al., 2016) were excluded as they

did not meet the protocol for requirements of 500 ms

of continuous speech. The recorded sustained vowel

‘o’ from the source (Sakar et al., 2013) were excluded

as the ‘a’ recordings were only considered for this

experiment. The overall database included 52 healthy

speech files and 44 Parkinsonian speech files of a

sustained ‘a’ sound from each speaker for a duration

of 500ms.

For this experiment, the GIF methods chosen were

closed phase methods and iterative methods as they

have shown to be robust in extracting the glottal

source in varying phonations. Spectral decomposition

methods were not selected as they consider the closed

phase of the glottal signal to be zero (Alku, 2011) and

this would not be appropriate for Parkinsonian speech

knowing the vocal fold disorders attributed to the

disease. The closed phase technique chosen was

quasi-closed phase (QCP) inverse filtering

(Airaksinen et al., 2014) and the iterative method

chosen was iterative and adaptive inverse filtering

(IAIF) (Alku, 1992). The QCP method needs

identification of GCIs and GOIs and these were

computed by the SEDREAMS algorithm (Drugman

and Dutoit., 2009). The glottal signal was estimated

from each speech recording by both methods.

Algorithms for these methods were implemented

from the sources (Degottex et al., 2014; Alku et al.,

2017).

For each glottal estimate from all speakers, the

selected five glottal parameters were measured on

every pitch period across the 500ms window. The

median value of the parameter was computed to

represent the value for the parameter on each file. The

median value was selected to remove any outliers and

to represent a true single value of one parameter from

the glottal signal. This single value for each of the five

glottal parameters was used to create ROC curves to

show the performance of distinguishing between

healthy and Parkinsonian speech. A value for the

AUC and the SE was computed to illustrate the

performance of the parameters. Glottal parameters

were extracted twice, using the IAIF and QCP

methods to indicate if one is producing a better

performance.

3 RESULTS

The performance of the separation of healthy and PD

speech for each parameter for the IAIF method is

shown in Figure 3. This is presented as individual

ROC curves for each parameter in different colours,

where the dashed line represents the value 0.5. It can

be seen from the curves that three of the tested glottal

parameters produce a good performance; QOQ, PSP

and H1-H2. The parameter QOQ shows the highest

performance for separating healthy and PD speech for

the IAIF method.

The performance of the separation of healthy and

PD speech for each parameter for the QCP method is

shown in Figure 4. This is presented as individual

ROC curves for each parameter in different colours,

where the dashed line represents the value 0.5. It can

be seen from the graph that the performance of all the

glottal parameters is slightly above the 0.5 line with

no single parameter showing an excellent

performance of separating PD and healthy speech.

BIOSIGNALS 2019 - 12th International Conference on Bio-inspired Systems and Signal Processing

120

Figure 3: ROC curves for glottal parameters tested by IAIF

method.

Figure 4: ROC curves for glottal parameters tested by QCP

method.

The ROC curves for the glottal parameters are

shown in Figure 3 and Figure 4. For further analysis

of the performance, the AUC was computed for each

parameter for both estimation techniques. The

obtained values from this experiment are presented in

Table 1. The results are presented in terms of AUC

and SE, with each individual glottal parameter result

from the two glottal estimation techniques, IAIF and

QCP.

Table 1: Results obtained for glottal parameters tested.

Glottal

Parameter

IAIF QCP

AUC SE AUC SE

QOQ 0.857 0.040 0.631 0.057

NAQ 0.467 0.059 0.547 0.059

H1H2 0.613 0.058 0.593 0.058

HRF 0.533 0.059 0.520 0.059

PSP 0.708 0.053 0.619 0.058

According to the results obtained from the IAIF

estimation algorithm the glottal parameter, QOQ,

computed an AUC value of over 0.857 which

indicates this parameter was different in healthy and

PD speech. PSP was found to have an AUC value

exceeding 0.71 which again indicates a good

separation between healthy and PD. NAQ was found

to have the lowest performance for the IAIF method

scoring an AUC value of 0.47 indicating this

parameter could not distinguish between healthy and

PD speech from this database.

The results for the QCP method show that the

parameters QOQ and PSP perform the highest at

separating healthy and PD speech for this technique

with both obtaining AUC values exceeding 0.61.

HRF has the lowest performance with an AUC value

of 0.52 indicating this does not perform well at

distinguishing between healthy and PD speech

signals.

The performance of the two estimation

techniques, IAIF and QCP, was analysed by

comparing the AUC and SE values of the parameters

obtained by both. The estimation technique that had

higher AUC values was considered to perform better

at separating healthy and Parkinsonian speech. IAIF

obtained higher values for the AUC in all parameters

except one, NAQ. For the parameter QOQ, IAIF

scored a higher AUC value, obtaining 0.857

(SE=0.040) with QCP obtaining 0.631 (SE=0.057).

The frequency domain parameter, PSP, also scored a

higher value using the IAIF method. This indicates

that overall, with these parameters, the IAIF method

performs better at discriminating between healthy and

Parkinsonian speech from the speech files in this

database.

Laryngoscope studies on PD patients reported that

vocal fold disorders are evident in Parkinsonian

speech. It would be expected that vocal fold disorders

could lead to pathological features in the glottal

signal. The results found in this study suggest that the

glottal flow exhibits different characteristics in

Parkinsonian speech when compared with healthy

Glottal Flow Analysis in Parkinsonian Speech

121

speech. Hanratty et al. (2016) reported that the

parameter QOQ scored a performance of over 90% at

separating healthy and PD speech files when the

glottal source was estimated by the IAIF method. In

this study, the database was increased to include more

PD speech files and QOQ still produced a high AUC

value of 0.857 when distinguishing between healthy

and PD.

4 CONCLUSIONS

Speech impairments are a common occurrence in PD

patients and this could be related to the vocal fold

abnormalities found in the patients. The results in this

study indicate that different behaviour is evident in

the glottal flow signal, with two glottal parameters

showing separation between PD and healthy speech

recordings from the test database.

The results indicate that the timing based

parameter, QOQ, and the frequency domain

parameter, PSP, show significant results when tasked

with distinguishing between healthy and PD speech.

According to the results from this experiment the

estimation technique IAIF outperformed the QCP

method with the selected glottal parameters. IAIF

obtained higher AUC values for all parameters except

one, indicating it is the appropriate method for

estimating the glottal source from Parkinsonian

speech

This experiment selected five glottal parameters

and two glottal estimation techniques but note that

many more possibilities exist that were not

considered in this study. Estimating the glottal flow

from speech signals can be a challenging task and is

particularly difficult for pathological speech, such as

that found in PD. Sources of variations exist in the

test dataset, which include different recording

protocols, severity of the disease in the PD group and

age of participants not matched to the control group

of healthy speech. Based on these critiques, there

must be caution on drawing broader conclusions.

Future work with an improved database with more

participants will be considered to fully understand

how parameters behave in the glottal flow of

Parkinsonian speech.

ACKNOWLEDGEMENTS

The authors would like to acknowledge

Technological University Dublin for funding this

research project.

The authors would also like to acknowledge

participants and staff that contributed to this study

from St. Mary’s Hospital, Dublin.

REFERENCES

Airaksinen, M. et al. (2014) Quasi Closed Phase Glottal

Inverse Filtering Analysis with Weighted Linear

Prediction. IEEE/ACM Transactions on Audio, Speech

and Language Processing (TASLP). 22 (3), pp. 596-

607.

Airas, M. (2008) TKK Aparat: An Environment for Voice

Inverse Filtering and Parameterization. Logopedics

Phoniatrics Vocology. 33 (1), pp. 49-64.

Alku, P. (2011) Glottal Inverse Filtering Analysis of

Human Voice production—a Review of Estimation and

Parameterization Methods of the Glottal Excitation and

their Applications. Sadhana. 36 (5), pp. 623-650.

Alku, P. (1992) Glottal Wave Analysis with Pitch

Synchronous Iterative Adaptive Inverse

Filtering. Speech Communication. 11 (2-3), pp. 109-

118.

Alku, P., Bäckström, T. & Vilkman, E. (2002) Normalized

Amplitude Quotient for Parametrization of the Glottal

Flow. The Journal of the Acoustical Society of

America. 112 (2), pp. 701-710.

Alku, P., Pohjalainen, H. & Airaksinen, M. (2017) Aalto

Aparat–A Freely Available Tool for Glottal Inverse

Filtering and Voice Source Parameterization

In Subsidia: Tools and Resources for Speech Sciences.

Alku, P., Strik, H. & Vilkman, E. (1997) Parabolic Spectral

Parameter-A New Method for Quantification of the

Glottal Flow.

Belalcázar-Bolaños, E.A., Orozco-Arroyave, J.R., Vargas-

Bonilla, J.F., Haderlein, T. & Nöth, E. (2016) Glottal

Flow Patterns Analyses for Parkinson’s Disease

Detection: Acoustic and Nonlinear

Approaches. International Conference on Text, Speech,

and Dialogue. : Springer, pp. 400.

Chien, Y. et al. (2017) Evaluation of Glottal Inverse

Filtering Algorithms using a Physiologically Based

Articulatory Speech Synthesizer. IEEE/ACM

Transactions on Audio, Speech, and Language

Processing. 25 (8), pp. 1718-1730.

Childers, D.G. (1999) Speech Processing. : John Wiley &

Sons, Inc.

Childers, D.G. & Lee, C. (1991) Vocal Quality Factors:

Analysis, Synthesis, and Perception. The Journal of the

Acoustical Society of America. 90 (5), pp. 2394-2410.

Degottex, G., Kane, J., Drugman, T., Raitio, T. & Scherer,

S. (2014) COVAREP—A Collaborative Voice

Analysis Repository for Speech

Technologies. Acoustics, Speech and Signal

Processing (ICASSP), 2014 IEEE International

Conference on. : IEEE, pp. 960.

Dorsey, E.R. et al. (2007) Projected Number of People with

Parkinson Disease in the most Populous Nations, 2005

through 2030. Neurology. 68 (5), pp. 384-386.

BIOSIGNALS 2019 - 12th International Conference on Bio-inspired Systems and Signal Processing

122

Doval, B., d'Alessandro, C. & Henrich, N. (2006) The

Spectrum of Glottal Flow Models. Acta Acustica

United with Acustica. 92 (6), pp. 1026-1046.

Drugman, T., Bozkurt, B. & Dutoit, T. (2012) A

Comparative Study of Glottal Source Estimation

Techniques. Computer Speech & Language. 26 (1), pp.

20-34.

Drugman, T. & Dutoit, T. (2009) Glottal Closure and

Opening Instant Detection from Speech Signals. Tenth

Annual Conference of the International Speech

Communication Association.

Fant, G. (1995) The LF-Model Revisited. Transformations

and Frequency Domain Analysis. Speech

Trans.Lab.Q.Rep., Royal Inst.of Tech.Stockholm. 2 (3),

pp. 40.

Fant, G. (1971) Acoustic Theory of Speech Production:

With Calculations Based on X-Ray Studies of Russian

Articulations. : Walter de Gruyter.

Fawcett, T. (2006) An Introduction to ROC

Analysis. Pattern Recognition Letters. 27 (8), pp. 861-

874.

Hacki, T. (1989) Classification of Glottal Dysfunctions on

the Basis of Electroglottography. Folia Phoniatrica. 41

(1), pp. 43-48.

Hanratty, J., Deegan, C., Walsh, M. & Kirkpatrick, B.

(2016) Analysis of Glottal Source Parameters in

Parkinsonian Speech. Engineering in Medicine and

Biology Society (EMBC), 2016 IEEE 38th Annual

International Conference of the. : IEEE, pp. 3666.

Hanson, D.G., Gerratt, B.R. & Ward, P.H. (1984)

Cinegraphic Observations of Laryngeal Function in

Parkinson's Disease. The Laryngoscope. 94 (3), pp.

348-353.

Harel, B., Cannizzaro, M. & Snyder, P.J. (2004) Variability

in Fundamental Frequency during Speech in Prodromal

and Incipient Parkinson's Disease: A Longitudinal Case

Study. Brain and Cognition. 56 (1), pp. 24-29.

Jeancolas, L., Benali, H., Benkelfat, B., Mangone, G.,

Corvol, J., Vidailhet, M., Lehericy, S. & Petrovska-

Delacrétaz, D. (2017) Automatic Detection of Early

Stages of Parkinson's Disease through Acoustic Voice

Analysis with Mel-Frequency Cepstral

Coefficients. Advanced Technologies for Signal and

Image Processing (ATSIP), 2017 International

Conference on. : IEEE, pp. 1.

Kane, J. (2012) Tools for Analysing the Voice. PhD Diss.,

Trinity College, Dublin, Ireland.

Little, M.A. et al. (2009) Suitability of Dysphonia

Measurements for Telemonitoring of Parkinson's

Disease. IEEE Transactions on Biomedical

Engineering. 56 (4), pp. 1015-1022.

Logemann, J.A. et al. (1978) Frequency and Cooccurrence

of Vocal Tract Dysfunctions in the Speech of a Large

Sample of Parkinson Patients. Journal of Speech and

Hearing Disorders. 43 (1), pp. 47-57.

Perez, K.S. et al. (1996) The Parkinson Larynx: Tremor and

Videostroboscopic Findings. Journal of Voice. 10 (4),

pp. 354-361.

Plumpe, M.D., Quatieri, T.F. & Reynolds, D.A. (1999)

Modeling of the Glottal Flow Derivative Waveform

with Application to Speaker Identification. IEEE

Transactions on Speech and Audio Processing. 7 (5),

pp. 569-586.

Quatieri, T.F. (2006) Discrete-Time Speech Signal

Processing: Principles and Practice. : Pearson

Education India.

Sakar, B.E. et al. (2013) Collection and Analysis of a

Parkinson Speech Dataset with Multiple Types of

Sound Recordings. IEEE Journal of Biomedical and

Health Informatics. 17 (4), pp. 828-834.

Sharma, R. (2014) Early Detection of Parkinson's Disease

through Voice. Advances in Engineering and

Technology (ICAET), 2014 International Conference

on. : IEEE, pp. 1.

Smith, M.E. et al. (1995) Intensive Voice Treatment in

Parkinson Disease: Laryngostroboscopic

Findings. Journal of Voice. 9 (4), pp. 453-459.

Tsanas, A. et al. (2010) Accurate Telemonitoring of

Parkinson's Disease Progression by Noninvasive

Speech Tests. IEEE Transactions on Biomedical

Engineering. 57 (4), pp. 884-893.

Tsanas, A. et al. (2012) Novel Speech Signal Processing

Algorithms for High-Accuracy Classification of

Parkinson's Disease. IEEE Transactions on Biomedical

Engineering. 59 (5), pp. 1264-1271.

Tsuboi, T. et al. (2015) Characteristic Laryngoscopic

Findings in Parkinson’s Disease Patients After

Subthalamic Nucleus Deep Brain Stimulation and its

Correlation with Voice Disorder. Journal of Neural

Transmission. 122 (12), pp. 1663-1672.

Wong, D., Markel, J. & Gray, A. (1979) Least Squares

Glottal Inverse Filtering from the Acoustic Speech

Waveform. IEEE Transactions on Acoustics, Speech,

and Signal Processing. 27 (4), pp. 350-355.

Yücetürk, A. et al. (2002) Voice Analysis and

Videolaryngostroboscopy in Patients with Parkinson's

Disease. European Archives of Oto-Rhino-

Laryngology. 259 (6), pp. 290-293.

Glottal Flow Analysis in Parkinsonian Speech

123