Deep Learning Solution for Pathological Voice Detection using

LSTM-based Autoencoder Hybrid with Multi-Task Learning

Dávid Sztahó, Kiss Gábor and Tulics Miklós Gábriel

Department of Telecommunication and Media Informatics, Budapest University of Technology and Economics,

Magyar tudósok körútja 2., Budapest, Hungary

Keywords: Deep Neural Networks, Long-Short Term Memory, Autoencoder, Multi-Task Learning, Speech, Bio-signal,

LSTM, Voice Pathologies, Parkinson, Depression.

Abstract: In this paper, a deep learning approach is introduced to detect pathological voice disorders from continuous

speech. Speech as bio-signal is getting more and more attention as a discriminant for different diseases. To

exploit information in speech, a long-short term memory (LSTM) autoencoder hybrid with multi-task learning

solution is proposed with spectrogram as input feature. Different speech databases (voice disorders,

depression, Parkinson’s disease) are applied as evaluation datasets. Applicability of the method is

demonstrated by obtaining accuracies 85% for Parkinson’s disease, 86% for dysphonia, and 90% for

depression on test datasets. The advantage of this method is that it is fully data-driven, in the sense that it does

not require special acoustic-phonetic preprocessing separately for the types of disease to be recognized. We

believe that the applied method in this article can be used to other diseases as well and can be used for other

languages also.

1 INTRODUCTION

Speech as bio-signal getting more and more attention

as a discriminant for different diseases. There can be

many alterations in speech production due to

neurological and/or organic disorders caused by

illnesses. In general, any alteration from ‘normal’

speech might be an indication of pathological speech.

Alterations of speech may be caused by various

things, for example psychological conditions such as

depression. Voice disorders are also main causes of

voice alternations. Voice disorder happens once

somebody’s voice quality, pitch, and loudness are

inappropriate for an individual’s age, gender, cultural

background, or geographic location. The American

Speech-Language-Hearing Association divides voice

disorders into two groups: organic voice disorders

and functional voice disorders. Organic disorders can

be structural and neurogenic in nature. Structural

disorders involve physical changes in the voice

mechanism, such as alterations in vocal fold tissues

such as oedema or vocal nodules, polyps,

gastroesophageal reflux disease (GERD), cyst and

vocal cord paralysis. Neurogenic voice disorders on

the other hand are caused by a problem in the nervous

system, that include voice problems caused by

abnormal control, coordination, or strength of voice

box muscles due to an underlying neurological

disease such as stroke, Parkinson’s disease, multiple

sclerosis.

Classifying speech into normal and disordered is

more problematic than it first seems. There are a large

number of works (Dastjerd et al., 2019; Filiou et al.,

2020; Jeancolas et al., 2020; Kiss & Vicsi, 2017a;

Klempíř & Krupička, 2018; Low et al., 2020; Tóth et

al., 2018; Zhang et al., 2019) subjected to

classification of these diseases using various machine

learning techniques.

Deep learning (DL) is one of the most frequently

used machine learning solutions nowadays. There are

many DNN algorithms developed, each is a proper

solution to a given data type processing. In this paper,

we utilize a long-short term memory (LSTM)

autoencoder (AE) hybrid with multi-task learning

(MTL) solution to propose a DL structure for

detecting multiple diseases, using a voice disorder, a

depression and a Parkinson’s disease dataset.

An important disadvantage of these classification

methods is that they may need complex phonetic

preprocessing in order to detect different parts of the

speech and, therefore, they may be language

dependent. The proposed method in this paper is an

Sztahó, D., Gábor, K. and Gábriel, T.

Deep Learning Solution for Pathological Voice Detection using LSTM-based Autoencoder Hybrid with Multi-Task Learning.

DOI: 10.5220/0010193101350141

In Proceedings of the 14th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2021) - Volume 4: BIOSIGNALS, pages 135-141

ISBN: 978-989-758-490-9

135

automatic way without the need of segmenting the

speech by an automatic speech recognizer (ASR).

Feature extraction is a critical step of any

classification method. DL has the ability to learn to

extract the best needed features for the best result.

There are end-to-end systems that receive raw sound

signals (amplitudes) as input and derive the proper

features themselves. For this, huge available data is

needed. Generally, dealing with speech and diseases,

this is not the case. In our work we use spectrograms

as inputs to the DL method and apply autoencoder

based feature learning (Yu et al., 2019).

Speech production process is time-varying. The

same linguistic content can be said in many durations.

Importantly, this variation may not correlate with the

given task (disease detection) at all. LSTMs, as a

special DL building element, has the property to learn

information across time due to its ability to have

memory. Therefore, it can learn information that

concerns different diseases (Gupta, 2018; Kim et al.,

2018; Mallela et al., 2020; Yang et al., 2016; Zhao et

al., 2019).

Overfitting is always an important concern in

classification trials. There are multiple ways of

overcoming this error, mostly by applying proper

dataset splitting (train-validation-test). Here, beside

the correct dataset splitting, multi-task learning

(Ruder, 2004) is also utilized. MTL is not only used

as regularization, but also for the parallel

classification and autoencoder (feature learning)

implementation.

Recent works dealing with deep learning and

disease classification include various convolutional

network assemblies, recurrent neural networks,

LSTMs and even solutions on mobile devices. Since

the majority of these works use different datasets for

evaluation (even for the same disease), it is hard to

compare their results. Most of them report about 90%

classification accuracy (Gunduz, 2019; Kaur et al.,

2019, 2020; Lam et al., 2019; Mdhaffar et al., 2019;

Mohammed et al., 2020; Rejaibi et al., 2019, p.). The

proposed AE-LSTM hybrid can be considered as

novel architecture among the found studies.

There are several approaches for the binary

classification of a healthy subject from voices

affected by some disorder. The first question is

whether to use sustained vowels or continuous

speech. Researchers achieved high accuracies using

sustained vowels (Orozco-Arroyave, 2015; Zhang,

2008; Ali, 2017; Teixeira, 2017), however, a

significant proportion of researchers use continuous

speech in their research pointing out the benefits of

using continuous speech over sustained vowels

(Vicsi, 2011; Guedes, 2019; Cordeiro, 2015). The

research findings are expected to be more applicable

to practical work since continuous speech is used in

real-world situations. In the work of (Guedes, 2019)

the German Saarbrücken Voice Database with the

phrase “Guten Morgen, wie geht es Ihnen?” to

classify dysphonia and healthy voices. A 66% f1-

score was reached in their experiment with Long-

Short-Term-Memory and Convolutional Network for

classification. In (Tulics, 2019) researchers used

acoustic features and phone-level posterior

probabilities computed by the DNN soft-max layer of

the speech recognition system and used them as an

input for an SVM and a Fully-Connected Deep

Neural Network. Classification accuracies were

ranging from 85% to 88% in their experiments.

The accuracy of distinguishing between depressed

and healthy subjects depends largely on the database

used, such as the size of the database and the severity

of the subjects included in it (Cummins et al., 2015).

The accuracy of the classification also depends on the

methods used, such as the feature extraction, or the

use of gender dependent or independent models (Low

et al., 2020). In (Kiss & Vicsi, 2017b) researchers

used gender dependent models and used selected

acoustic features as an input for an SVM, achieving

86% accuracy with a database which can be

considered similar as ours.

The paper is structured as follows: in Section 2,

the used speech datasets are described. Section 3

discusses the methods applied. Section 4, 5 and 6

contains the results, discussions and conclusions.

2 DATABASE

The proposed DL structure was tested on three

datasets of three disease types. Each dataset contained

the recording of reading a short folk tale ‘The North

Wind and the Sun’ in Hungarian. For each dataset,

healthy speakers are included as a control population

with the same sample number as the given disease

dataset and age and gender distribution matching the

patients’ statistics. Although the classification task

described is binary, for the sake of completeness, the

severity of the disease is also noted here for each

dataset. The audio samples were recorded with

external USB sound card and clip-on microphone in

PCM format using 44 kHz sampling rate and 16-bit

quantization. Informed consents were signed by each

patient before recordings.

BIOSIGNALS 2021 - 14th International Conference on Bio-inspired Systems and Signal Processing

136

2.1 Hungarian Parkinson’s Speech

Dataset (HPSD)

Speech samples were collected from patients

diagnosed with Parkinson’s disease (PD) by two

health institutes in Budapest: Virányos Clinic and

Semmelweis University. The severity of PD was

labelled according to the Hoehn & Yahr scale (H-Y)

(Hoehn & Yahr, 1967). The H-Y scale ranges from 1

to 5, where 1 indicates minimal PD while 5 is the

worst PD condition. We did not filter the patients

according to the taken medications. All patients were

in ON state.

83 speech sample were collected from patients

with PD: 43 male speakers (mean H-Y score:

2.74(±1.05); mean age: 64(±9.5)) and 40 female

speakers (mean H-Y score: 2.74(±1.10); mean age:

65.4(±9.4)).

2.2 Voice Disorder Speech Dataset

(VDSD)

Voice samples from patients were collected during

patient consultations in a consulting room at the

Department of Head and Neck Surgery of the

National Institute of Oncology. The collected speech

database contains voices from people suffering from

diseases like tumors at various places of the vocal

tract, gastroesophageal reflux disease, chronic

inflammation of larynx, bulbar paresis, amyotrophic

lateral sclerosis, leukoplakia, spasmodic dysphonia,

etc. The recorded voice samples in this experiments

were classified by a leading phoniatric according to

the RBH scale. The RBH scale gives the severity of

dysphonia, where R stands for roughness, B for

breathiness and H for overall hoarseness. The degree

of the category H cannot be less than the highest rate

of the other two categories. For example, if B = 3 and

R = 2, H is 3, and cannot be 2 or 1. A healthy voice’s

code is R0B0H0; the maximum H and respectively

RBH value is 3, so a voice’s code with severe

dysphonia is R3B3H3. Here the H score is given.

The database contains a total of 261 recordings

from patients with dysphonia: 159 females (mean H

score: 1.72(±0.77); mean age: 57.3(±14.8)) and 102

males (mean H score: 2 (±0.83); mean age:

53.7(±15.1)).

2.3 Hungarian Depressed Speech

Dataset (HDSD)

Speech samples were collected from patients

diagnosed with depression by the Psychiatric and

Psychotherapeutic Clinic of Semmelweis University,

Budapest. Patients with antipsychotic medication

which can affect the acoustical features of speech

were left out. The degree of severity of depression

was recorded using the Beck Depression Inventory II

(BDI) scale (Beck et al., 1996)). The BDI-II scale

ranges from 0 to 63, where 0 indicates a healthy state,

while 63 is the worst depression condition. The BDI-

II scale uses the following rating: 0-13 healthy, 14-19

mild depression, 20-28 moderate depression, 29-63

severe depression.

A total of 107 speech sample were recorded from

depressed patients: 64 female subjects (mean BDI

score: 28.0(±9.0); mean age: 37.5(±13.5) and 43

males subjects (mean BDI score: 26.1(±8.0); mean

age: 40.8(±13.6)).

3 METHODS

3.1 Implemented Deep Learning

Architecture and Input

The DL architecture, training and evaluation was

implemented in Tensorflow 2.1.0. The implemented

DL architecture consists of an LSTM and an

autoencoder part. Multi-task learning is applied in

order to train the network for the task-specific labels

and the autencoder-based feature extraction. Figure 1

shows the structure of the implemented network.

Spectrogram is fed to the network as input. For each

audio sample, the spectrogram was extracted by 10

ms timestep and 256 FFT size (resulting in 16 ms

window size), commonly used in speech analysis.

Because Tensorflow did not manage the varying

duration of the samples, this was solved by using a

Masking layer at the input. Technically, each

spectrogram was padded with 0.0 elements to reach

the duration of the longest audio sample. By using the

Masking layer, the 0.0 elements were skipped during

training and prediction processes.

The DL architecture consists of two parts. An

autoencoder part learns feature representation

(dimensionality reduction) for the audio sample

spectra. This is intended to encode information in the

spectra. Part of the bottleneck layer (fc

autoencoder

) has

shared neurons that are also trained to the task-

specific target labels. This tries to ensure that part of

the encoded spectra contains information that is

specific to the given disease. This also serves as a

regulation technique to avoid overfitting. The idea

behind this multi-task learning is that this forces the

encoded spectra to contain information partly about

the disease characteristics.

Deep Learning Solution for Pathological Voice Detection using LSTM-based Autoencoder Hybrid with Multi-Task Learning

137

The other part of the DL architecture performs

disease-specific classification. This part contains

two fully connected layers (one is a shared layer in

the bottleneck layer) with relu activation functions,

dropout layers (dropout parameter set to 0.5) and

softmax layer at the end. Before the softmax layer an

Figure 1: Structure of the implemented network.

Table 1: Number of units in DL layers.

layer name

units

lstm 100

autoencoder1

autoencoder2

200

time distributed layer 128

average pooling layer is needed in order to make one

decision for one audio sample (by averaging the

outputs of layer fc

A shared LSTM layer is also applied in order to

learn time varying information.

Layer sizes are shown in Table 1. The appropriate

numbers were selected according to preliminary

experiments.

3.2 Dataset Splitting and Network

Training

Evaluation was performed by splitting each database

into training, validation and test sets by the following

method.

Due to the limited number of audio samples

available, all samples were used for testing by a 10-

fold cross-validation process. 10 test sets were

created by 10-fold cross-validation (stratified). In

each cross-validation iteration, the remaining 90% of

the samples were split into training and validation

sets, 70% and 20% respectively, by stratified random

sampling. Training was done on the training set and

an early stopping was applied on the validation set.

Minimum cross-entropy of the disease classification

was used as a cost function for early stopping.

Maximum 1000 training epochs were done with 50

patience steps for early stopping. During training,

‘Adam’ optimizer was used.

4 RESULTS

Accuracy, sensitivity and specificity was used as

evaluation metrics. Tables 2, 3 and 4 show the results

in each set during tests (training, validation and

testing). The confusion matrices for the test sets are

also shown in Table 5. The values in the cells are the

number of samples (subjects).

The results show that the applied DL performs

well on all three datasets. The lowest accuracy on the

test sets is 0.86, which means that 86% of the samples

are correctly classified into healthy or disease

categories. The highest score is 0.90 in the case of

depression.

The DL method doesn’t seem to overfit. Balanced

results metrics are achieved in the training, validation

and test steps.

Also, sensitivity and specificity scores are well

balanced in the case of VDSD and HDSD datasets.

Parkinson samples are a little bit unbalanced

according to these metrics. Higher specificity is

reached, which means that healthy samples are

classified more accurate than the Parkinson samples.

BIOSIGNALS 2021 - 14th International Conference on Bio-inspired Systems and Signal Processing

138

Table 2: Results on the HPDS samples.

accuracy sensitivity specificity

training 0.86 0.77 0.95

validation 0.88 0.80 0.96

test 0.85 0.75 0.95

Table 3: Results on the VDSD samples.

accuracy sensitivity specificity

training 0.90 0.89 0.91

validation 0.87 0.86 0.89

test 0.86 0.85 0.87

Table 4: Results on the HDSD samples.

accuracy sensitivity specificity

training 0.93 0.92 0.93

validation 0.90 0.90 0.89

test 0.90 0.91 0.89

Table 5: Confusion matrices obtained on the test datasets.

PDSD

predicted positive predicted negative

true positive 62 21

true negative 4 79

PDSD

predicted positive predicted negative

true positive 226 35

true negative 28 164

HDSD

predicted positive predicted negative

true positive 95 12

true negative 9 93

5 DISCUSSION

The achieved results on the applied three datasets

show that the proposed DL method is able to extract

information from the samples in order to make

distinction between negative (healthy) and positive

cases. These actual accuracy, sensitivity and

specificity scores are, naturally, dependent on the

datasets. However, they can be considered large

enough, that some statements could be concluded

based on them. With the extension of the recordings

and by applying more datasets in different languages

may also prove the proposed DL method to be more

robustly usable.

Based on the results, the highest evaluation scores

were achieved on the depression database test set (and

also on training and validation sets). Based on

personal experience of clinical experts, intonation is

also highly affected by depression. By applying an

LSTM layer, this intonational disorder (which can be

captured through temporal analysis) can be more

accurately modelled.

The suggested method not only captured

intonational features of speech, but voice

characteristics related to dysphonia, such as

hoarseness and breathiness.

In case of Parkinson’s disease lower sensitivity

scores are achieved than specificity scores. Although

a high sensitivity-specificity balance is more

desirable, the present case doesn’t mean that the

method can’t be used as pre-screening. Less positive

samples will be detected, but the overall accuracy can

be considered sufficient for the task. In fact, every

method is usable with over random performance.

The results achieved are comparable to the

previous results in the field. Since actual results are

dataset dependent, direct comparison of accuracy and

other metrics is problematic. (Kiss & Vicsi, 2017b)

reported 86% accuracy for a former version of the

depression dataset. Here, we achieved 90%. In case

of PD, (Sztahó et al., 2019) reported around 88%

accuracy using cross-validation setup, without

separate independent test set. The 85% achieved here

is comparable, especially if we add that here we

applied an appropriate test set. In the case of VDSD

dataset, a previous result is reported in (Tulics, 2019).

In that, a 95% accuracy was described with possible

overfitting effect, and between 85-88% without

overfitting. Here, we achieved 86% without probable

overfitting. All these researches used segmentation

information to obtain the highest performance. In our

case here, this computationally intensive step is not

needed.

Among many other usages, actual practical

applicability can be pre-screening in general

practitioner offices or home-care environments. A

cheap, easy to use devices (software) can be

implemented to detect various diseases that affect

speech.

Deep Learning Solution for Pathological Voice Detection using LSTM-based Autoencoder Hybrid with Multi-Task Learning

139

6 CONCLUSION

The goal of this work is to introduce a novel method

for pathological speech recognition using continuous

speech without the need of a voicing detection or

speech segmentation application. The proposed DL

architecture consists of two parts: an autoencoder part

learns feature representation and a disease-specific

classification.

We demonstrated the applicability of the method

by classifying three different diseases: Parkinson’s

disease, dysphonia related voice disorders and

depression. The method achieved 0.85 for

Parkinson’s disease, 0.86 for dysphonia, and 0.90 for

depression on the test datasets. These classification

accuracies correspond to the classification accuracies

mentioned in the literature. The advantage of this

method is that it is fully data-driven, in the sense that

it does not require special acoustic-phonetic

preprocessing separately for the types of disease to be

recognized. The speech recordings can be directly

given to the deep neural network (using

spectrographic extraction only).

We believe that the applied method in this article

can be used to other diseases as well and can be used

for other languages also.

ACKNOWLEDGEMENTS

Project no. K128568 has been implemented with the

support provided from the National Research,

Development and Innovation Fund of Hungary,

financed under the K_18 funding scheme. The

research was partly funded by the CELSA

(CELSA/18/027) project titled: “Models of

Pathological Speech for Diagnosis and Speech

Recognition”.

REFERENCES

Ali, Z., Talha, M., & Alsulaiman, M., 2017. A practical

approach: Design and implementation of a healthcare

software for screening of dysphonic patients. IEEE

Access, 5, 5844-5857.

Beck, A. T., Steer, R. A., Ball, R. & Ranieri, W. F., 1996.

Comparison of beck depression inventories -IA and -II

in psychiatric outpatients. Journal of Personality

Assessment 67, 588–597.

Cordeiro, H., Meneses, C., & Fonseca, J., 2015. Continuous

speech classification systems for voice pathologies

identification. In Doctoral Conference on Computing,

Electrical and Industrial Systems, pp. 217-224.

Cummins, N., Scherer, S., Krajewski, J., Schnieder, S.,

Epps, J., & Quatieri, T. F., 2015. A review of depression

and suicide risk assessment using speech analysis.

Speech Communication, 71, 10-49.

Dastjerd, N. K., Sert, O. C., Ozyer, T., & Alhajj, R., 2019.

Fuzzy Classification Methods Based Diagnosis of

Parkinson’s disease from Speech Test Cases. Current

Aging Science, 12(2), 100–120.

Filiou, R.-P., Bier, N., Slegers, A., Houzé, B., Belchior, P.,

& Brambati, S. M., 2020. Connected speech assessment

in the early detection of Alzheimer’s disease and mild

cognitive impairment: A scoping review. Aphasiology,

34(6), 723–755.

Guedes, V., Teixeira, F., Oliveira, A., Fernandes, J., Silva,

L., Junior, A., & Teixeira, J. P., 2019. Transfer

Learning with AudioSet to Voice Pathologies

Identification in Continuous Speech. Procedia

Computer Science, 164, 662-669.

Gunduz, H., 2019. Deep learning-based Parkinson’s disease

classification using vocal feature sets. IEEE Access, 7,

115540–115551.

Gupta, V., 2018. Voice disorder detection using long short

term memory (lstm) model. ArXiv Preprint

ArXiv:1812.01779.

Hoehn, M. & Yahr, M. D., 1967. Parkinsonism onset,

progression, and mortality. Neurology 17, pp. 427–427

Jeancolas, L., Petrovska-Delacrétaz, D., Mangone, G.,

Benkelfat, B.-E., Corvol, J.-C., Vidailhet, M., Lehéricy,

S., & Benali, H., 2020. X-vectors: New Quantitative

Biomarkers for Early Parkinson’s Disease Detection

from Speech. ArXiv:2007.03599 [Cs, Eess, q-Bio].

http://arxiv.org/abs/2007.03599

Kaur, S., Aggarwal, H., & Rani, R., 2019. Diagnosis of

Parkinson’s Disease Using Principle Component

Analysis and Deep Learning. Journal of Medical

Imaging and Health Informatics, 9(3), 602–609.

Kaur, S., Aggarwal, H., & Rani, R., 2020. Hyper-parameter

optimization of deep learning model for prediction of

Parkinson’s disease. Machine Vision and Applications,

31(5), 32.

Kim, M. J., Cao, B., An, K., & Wang, J., 2018. Dysarthric

Speech Recognition Using Convolutional LSTM

Neural Network. INTERSPEECH, 2948–2952.

Kiss, G., & Vicsi, K., 2017a. Comparison of read and

spontaneous speech in case of automatic detection of

depression. 2017 8th IEEE International Conference on

Cognitive Infocommunications (CogInfoCom), 213–

218.

Kiss, G., & Vicsi, K., 2017b. Mono-and multi-lingual

depression prediction based on speech processing.

International Journal of Speech Technology, 20(4),

919-935.

Klempíř, O., & Krupička, R., 2018. Machine learning using

speech utterances for Parkinson disease detection.

Lékař a Technika - Clinician and Technology, 48(2),

66–71.

Lam, G., Dongyan, H., & Lin, W., 2019. Context-aware

deep learning for multi-modal depression detection.

ICASSP 2019-2019 IEEE International Conference on

BIOSIGNALS 2021 - 14th International Conference on Bio-inspired Systems and Signal Processing

140

Acoustics, Speech and Signal Processing (ICASSP),

3946–3950.

Low, D. M., Bentley, K. H., & Ghosh, S. S., 2020.

Automated assessment of psychiatric disorders using

speech: A systematic review. Laryngoscope

Investigative Otolaryngology, 5(1), 96–116.

Mallela, J., Illa, A., Suhas, B. N., Udupa, S., Belur, Y.,

Atchayaram, N., Yadav, R., Reddy, P., Gope, D., &

Ghosh, P. K., 2020. Voice based classification of

patients with Amyotrophic Lateral Sclerosis,

Parkinson’s Disease and Healthy Controls with CNN-

LSTM using transfer learning. ICASSP 2020-2020

IEEE International Conference on Acoustics, Speech

and Signal Processing (ICASSP), 6784–6788.

Mdhaffar, A., Cherif, F., Kessentini, Y., Maalej, M.,

Thabet, J. B., Maalej, M., Jmaiel, M., & Freisleben, B.,

2019. DL4DED: Deep Learning for Depressive

Episode Detection on Mobile Devices. International

Conference on Smart Homes and Health Telematics,

109–121.

Mohammed, M. A., Abdulkareem, K. H., Mostafa, S. A.,

Khanapi Abd Ghani, M., Maashi, M. S., Garcia-

Zapirain, B., Oleagordia, I., Alhakami, H., & AL-Dhief,

F. T., 2020. Voice Pathology Detection and

Classification Using Convolutional Neural Network

Model. Applied Sciences, 10(11), 3723.

Orozco-Arroyave, J. R., Hönig, F., Arias-Londoño, J. D.,

Vargas-Bonilla, J. F., Skodda, S., Rusz, J., & Nöth, E.,

2015. Voiced/unvoiced transitions in speech as a

potential bio-marker to detect Parkinson's disease. In

Sixteenth Annual Conference of the International

Speech Communication Association.

Rejaibi, E., Komaty, A., Meriaudeau, F., Agrebi, S., &

Othmani, A., 2019. MFCC-based Recurrent Neural

Network for Automatic Clinical Depression

Recognition and Assessment from Speech. ArXiv

Preprint ArXiv:1909.07208.

Ruder, S., 2017. An overview of multi-task learning in deep

neural networks. ArXiv Preprint ArXiv:1706.05098.

Sztahó, D., Valálik, I., & Vicsi, K., 2019. Parkinson’s

Disease Severity Estimation on Hungarian Speech

Using Various Speech Tasks. In International

Conference on Speech Technology and Human-

Computer Dialogue (SpeD). pp. 1-6.

Tóth, L., Hoffmann, I., Gosztolya, G., Vincze, V.,

Szatlóczki, G., Bánréti, Z. & Kálmán, J., 2018. A

speech recognition-based solution for the automatic

detection of mild cognitive impairment from

spontaneous speech. Current Alzheimer

Research, 15(2), 130-138.

Tulics, M. G., Szaszák, G., Mészáros, K., & Vicsi, K., 2019.

Artificial Neural Network and SVM based Voice

Disorder Classification. In 10th IEEE International

Conference on Cognitive Infocommunications

(CogInfoCom). pp. 307-312.

Vicsi, K., Imre, V., & Mészáros, K., 2011. Voice disorder

detection on the basis of continuous speech. In 5th

European Conference of the International Federation

for Medical and Biological Engineering, pp. 86-89.

Yang, T.-H., Wu, C.-H., Huang, K.-Y., & Su, M.-H., 2016.

Detection of mood disorder using speech emotion

profiles and LSTM. 10th International Symposium on

Chinese Spoken Language Processing (ISCSLP), pp. 1–

Yu, J., Zheng, X., & Wang, S., 2019. A deep autoencoder

feature learning method for process pattern recognition.

Journal of Process Control, 79, pp. 1–15.

Zhang, H., Song, C., Wang, A., Xu, C., Li, D., & Xu, W.,

2019. PDVocal: Towards Privacy-preserving

Parkinson’s Disease Detection using Non-speech Body

Sounds. The 25th Annual International Conference on

Mobile Computing and Networking, pp. 1–16.

Zhang, Y., & Jiang, J. J., 2008. Acoustic analyses of

sustained and running voices from patients with

laryngeal pathologies. Journal of Voice, 22(1), pp. 1-9.

Zhao, J., Mao, X., & Chen, L., 2019. Speech emotion

recognition using deep 1D & 2D CNN LSTM networks.

Biomedical Signal Processing and Control, 47, pp.

312–323.

Deep Learning Solution for Pathological Voice Detection using LSTM-based Autoencoder Hybrid with Multi-Task Learning

141