Analyzing Male Depression Using Empirical Mode Decomposition

Xavier S

anchez Corrales

1,2 a

, Jordi Sol

e-Casals

2,3 b

, Enrique Arroyo Garc

ıa

4 c

and Diego Palao Vidal

4 d

Researcher, Mental Health Department, Consorci Corporaci

o Sanit

aria Parc Taul

ı, Sabadell, Barcelona, Spain

Data and Signal Processing Research Group, University of Vic–Central University of Catalonia, Vic, Spain

Department of Psychiatry, University of Cambridge, Cambridge, U.K.

Consorci Corporaci

o Sanit

aria Parc Taul

ı, Sabadell, Barcelona, Spain

Keywords:

Depression, IMF (Intrinsic Mode Functions), EMD (Empirical Mode Decomposition), Bootstrapping,

Gaussian Kernel.

Abstract:

This study investigates the differences in male voice between healthy individuals and individuals with depres-

sion, using Empirical Mode Decomposition (EMD) analysis. Voice recordings from 25 men with depression

and 76 without were analyzed. The methodology consisted of extracting 16 Intrinsic Mode Functions (IMFs)

from 20-second voice segments, followed by statistical analyses including bootstrapping of means and stan-

dard deviations with False Discovery Rate (FDR) correction, comparison of probability density functions,

and the application of a Gaussian kernel. The results showed signiﬁcant differences between the means and

standard deviations, The application of the Gaussian kernel revealed more pronounced differences in IMFs 2

to 6, providing more speciﬁc discrimination than traditional statistical methods. The study contributes to the

development of non-invasive and objective diagnostic tools for depression.

1 INTRODUCTION

Depression is a widespread mental health disorder

that affects millions of people worldwide. Tradition-

ally, its identiﬁcation and monitoring has been based

on subjective methods such as clinical interviews and

standardized questionnaires. However, there is a

growing need to develop more objective and quantiﬁ-

able assessment tools.

Speech processing has been widely applied in var-

ious health-related ﬁelds, including the diagnosis of

sleep apnoea (Sol

e-Casals et al., 2014), early detec-

tion of Alzheimer’s disease (L

opez-de Ipi

na et al.,

2015; Lopez-de Ipi

na et al., 2015), Parkinson’s dis-

ease (Mekyska et al., 2018), and stroke recovery

through brain-computer interfaces (Tong et al., 2023).

Additional examples and applications can be found

in the works of Sol

e-Casals et al. (2010) and Es-

posito et al. (2016). Despite these advancements,

research in the ﬁeld of mental health remains rela-

https://orcid.org/0009-0002-4335-6851

https://orcid.org/0000-0002-6534-1979

https://orcid.org/0000-0002-3323-6568

https://orcid.org/0009-0009-8937-3623

tively limited. Within this context, voice analysis has

emerged as a promising tool for the detection and

evaluation of mental disorders, including depression

(Krishnan et al., 2021; Akkaralaertsest and Yingth-

awornsuk, 2019; Alghowinem et al., 2013; Espinola

et al., 2022).

Our study focuses on voice signal analysis to dif-

ferentiate between men with depression and those

who are healthy, using Empirical Mode Decompo-

sition (EMD). Due to space limitations, this study

will focus solely on the male sample. In future stud-

ies, we will address the analysis of the female sam-

ple, considering the comparison between both. This

method, based on the Hilbert-Huang Transform (Liu

et al., 2020), is particularly well-suited for the anal-

ysis of nonlinear and non-stationary signals such as

voice (Chen et al., 2021). EMD decomposes the sig-

nal, the raw signal in our case, into Intrinsic Mode

Functions (IMFs); this method is used to separate the

voice into modes from which differentiable character-

istics between depression and health can be extracted

(Sharma et al., 2017).

The primary objective of this study is to iden-

tify the most representative Intrinsic Mode Functions

(IMFs) in the depressive voice of men by comparing

886

Corrales, X. S., Solé-Casals, J., García, E. A. and Vidal, D. P.

Analyzing Male Depression Using Empirical Mode Decomposition.

DOI: 10.5220/0013157600003911

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 18th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2025) - Volume 1, pages 886-892

ISBN: 978-989-758-731-3; ISSN: 2184-4305

the probability density in the distributions of IMFs be-

tween the depression and healthy groups. This study

contributes to the identiﬁcation of differentiable char-

acteristics in the male voice through the evaluation of

IMFs. Firstly, the most important statistical features

for extracting these characteristics from voice data are

identiﬁed. Then, the statistical feature that shows sig-

niﬁcant differences based on its distribution density is

compared. Finally, the data are ﬁltered using a kernel

that considers this statistical feature (sigma) as a ﬁlter

for extracting the characteristics.

We must take into account that, the real-time ap-

plicability represents a signiﬁcant advancement in the

clinical and hospital setting, as it allows for the col-

lection and processing of voice signals on-site, di-

rectly during the interaction between the patient and

the healthcare professional. This type of immediate

processing could facilitate faster and more accurate

diagnoses without the need for advanced equipment

or prolonged waiting times for analysis. Instead of

relying on remote servers or lengthy processing times

required by Deep Learning models, an EMD-based

system could analyze the vocal characteristics of the

patient within seconds, contributing to faster clini-

cal decisions, optimizing care, and improving over-

all efﬁciency in high-demand settings such as hospi-

tals or medical consultations. According to the study

of Tasnim and Novikova (2022), the use of Deep

Learning features only resulted in a marginal per-

formance improvement (0.0004%), while consuming

1000 times more memory and 3000 times more com-

putation time compared to Machine Learning mod-

els, like the Gaussian kernel approach we used in this

study. This could have signiﬁcant implications for

rapid and efﬁcient diagnosis in clinical settings.

Our hypothesis is that statistical analysis and the

Gaussian distribution of IMFs will provide robust fea-

tures for identifying differences, and that these fea-

tures will help effectively classify between the voices

of subjects with depression and those who are healthy.

This study aims to contribute to the development

of non-invasive and objective assessment tools for de-

tecting and monitoring depression in clinical settings,

advancing the understanding of voice characteristics

associated with depression in men. The way to con-

tribute knowledge to this ﬁeld is by aiding in the iden-

tiﬁcation of differentiable classiﬁcation characteris-

tics in voice using a model based on a Gaussian ker-

nel.

2 METHODOLOGY

The data used in this study come from the Dis-

tress Analysis Interview Corpus (DAIC) at the Uni-

versity of Southern California (Gratch et al., 2014).

Data download was conducted following the ethical

protocol established by the university, adhering to

anonymization standards to protect the participants

identities.

The data segmentation was performed over the to-

tal length of the voice data for each group individu-

ally, dividing it into 20-second segments, which were

subsequently randomised. The inclusion of more than

one voice in a single segment was not considered.

This methodology ensures that potential uncontrolled

variables in the voice do not interfere with the study.

Our methodological approach involves extracting 16

IMFs from each 20-second voice segment, followed

by statistical analysis that includes bootstrapping cal-

culations with the statistics of mean, median, standard

deviation, kurtosis, and skewness, corrected with the

False Discovery Rate (FDR) method. We also per-

formed a probability density function analysis of the

IMFs comparing the depression and healthy groups,

and conducted comparisons between both groups us-

ing a Gaussian kernel.

To clearly and concisely illustrate the methodol-

ogy employed in this study, a ﬂowchart is presented

in Fig. 1.

From this database, we used the male voice

data ﬁles, totaling 101 wav ﬁles. Of these, 25

are from men with depression and 76 are from

men without depression (healthy). The depression

identiﬁcation was based on results from the PHQ-

8 (Patient Health Questionnaire depression scale)

(Kroenke et al., 2009).

The data were downloaded from the aforemen-

tioned source (DAIC) in an Excel ﬁle for diagnostic

differentiation. After separating the depression and

healthy data, silence longer than 0.5 seconds was re-

moved, and all recordings were consolidated into a

single ﬁle. The voice recordings from the depression

group amounted to 20 minutes and 35 seconds, while

the healthy group data totaled 1 hour and 00 seconds.

Silence removal was performed using the open-source

program Audacity version 3.5.1 (Audacity, 2023).

The original voice signal was recorded at a sam-

pling frequency of 16 kHz. However, upon analyzing

the spectrograms, it was observed that the relevant

content of the signal did not exceed 10 kHz in any

case. Given that processing data at 16 kHz incurs a

high computational cost, the sampling frequency was

reduced to 10 kHz. This reduction was performed

using the Librosa library (version 0.10.1) in Python

Analyzing Male Depression Using Empirical Mode Decomposition

887

Figure 1: Flowchart of the methodological procedure used

in this study. In statistical analysis, STD refers to standard

deviation, Skew to skewness, and Kur to kurtosis.

(version 3.11.9), which is the programming environ-

ment used for all analysis in this study. This opti-

mization allowed for more efﬁcient processing with-

out loss of information in the voice signal. During the

frequency reduction, the procedure with the Librosa

library also performed a min-max normalization to

the range [-1, 1].

Next, both audio ﬁles (depression and healthy)

were segmented into 20-second parts. The audio

from the depression group was divided into 62 20-

second segments, and after discarding the last one

due to its smaller size after the cut, 61 segments re-

mained. For the healthy group, 179 segments were

initially obtained, and after applying the same pro-

cedure, 178 segments were retained. Both groups of

segments, depression and healthy, were randomized

individually. From these two voice groups, 16 Intrin-

sic Mode Functions (IMFs) were extracted from each

20-second signal.

For this, Empirical Mode Decomposition (EMD)

was used from the library of the same name in the

open-source programming language Python (version

3.11.9). The number of IMFs was determined by

stopping the extraction of modes when the informa-

tion in the last IMF was practically ﬂat.

Next, we performed a comparison of means, stan-

dard deviations, kurtosis, and skewness of the 16

IMFs using the bootstrapping method for depression

and healthy groups. For this, the statistics for each

vector corresponding to a 20-second voice segment

(depression n = 61, healthy n = 178) were calculated

for each IMF (n = 16), resulting in a matrix of (61,

16) for the depression group and a matrix of (178, 16)

for the healthy group.

On these matrices, we performed bootstrapping

calculations comparing the aforementioned statistics

of the ﬁrst IMF from the depression group with the

ﬁrst IMF from the healthy group, and so on for each

IMF. In this process, the number of iterations in the

bootstrapping method (n iterations) was set to 10,000.

Subsequently, the False Discovery Rate (FDR)

method was applied to the 16 p-values resulting from

each statistic to control for the possibility of false pos-

itives among all signiﬁcant results. In biological sig-

nal data such as voice, which can be inherently vari-

able and not always follow a normal distribution, the

bootstrapping method with FDR correction provides

a more robust and conservative approach, as reﬂected

in Fig. 2.

Next, to compare the probability density func-

tion estimates between depression and healthy groups

across different IMFs, we plotted the mean and stan-

dard deviation statistics for each IMF, comparing de-

pression (dark bar) with healthy (light bar). The re-

sults can be seen in Fig. 3.

Subsequently, a Gaussian kernel was applied. In

equation (1), we can see the equation applied directly

to the data derived from the Empirical Mode Decom-

position, speciﬁcally the Intrinsic Mode Functions

(IMFs). Where x and x’ in the formula, represent spe-

ciﬁc points in these IMFs that are being compared (for

example: x represents the points of IMF1 from the de-

pression group and x’ the points of the same IMF in

the healthy group). The comparison is made at each

point of the IMFs, and then the average of all the point

similarities is taken for each particular IMF.

K(x, x

′

) = exp



−γ∥x − x

′

∥



, where γ =

2σ

(1)

BIOSIGNALS 2025 - 18th International Conference on Bio-inspired Systems and Signal Processing

888

Figure 2: Heatmap of p-values (p < 0.05 in blue) from

applying bootstrapping to statistics with FDR correction.

This kernel works by applying a convolutional ﬁl-

ter to each point in the signal, thereby smoothing the

signal without affecting its signiﬁcant patterns.

The purpose of this type of kernel is to ﬁnd dis-

tribution similarities between the elements (points)

of the IMFs. In our case, we compared the IMFs

from the depression group with those from the healthy

group, as obtained from the Empirical Mode Decom-

position. The results are shown in Fig. 3.

3 RESULTS

The results of the statistical comparisons with boot-

strapping corrected using the False Discovery Rate

method (Fig. 2). indicate signiﬁcant differences in the

means of the ﬁrst 8 IMFs and the last IMF between

depression and healthy groups, except for IMF 2 and

IMF 5.

For standard deviations, differences are observed

in the ﬁrst two IMFs, IMF 7, and from IMF 11 on-

wards. The relationship between the comparisons of

means and standard deviations appears to be inverse;

up to the midpoint of the IMFs, the means show sig-

niﬁcant differences (p value < 0.05), except for IMFs

2, 5, and 16. In contrast, signiﬁcant differences in

standard deviations are seen in the higher IMFs, ex-

cept for IMF 1 and IMF 2. Notably, both IMF 1 and

IMF 16 show signiﬁcant differences between depres-

sion and healthy groups in both statistics. In compar-

isons of kurtosis and skewness using the bootstrap-

ping method, we did not ﬁnd signiﬁcant differences

in p-values for any IMF.

The probability density functions of the means of

the IMFs (Fig. 3) show differences between the prob-

abilities of the IMFs in depression and healthy groups,

with an inverse relationship between means and stan-

Figure 3: Comparisons of the mean and std probability den-

sity function between depression and health of the IMFs.

dard deviations. Despite being on a logarithmic scale,

the present graph includes some bars with values so

close to zero that they are not visible.

For the means, density signiﬁcantly decreases af-

ter IMF 1 and then decreases more gradually, with

more pronounced differences in density probabilities

between depression and healthy groups in the ﬁrst

IMFs, which stabilize from IMF 9 onwards. Con-

versely, for standard deviations, density progressively

increases in the IMFs, with generally less signiﬁcant

inequalities between depression and healthy groups,

except in the last IMF.

The results of applying the Gaussian kernel, ac-

cording to equation (1) and shown in Fig. 4, high-

light signiﬁcant differences between depression and

healthy groups in the ﬁrst 7 IMFs. These differences

are most pronounced from IMF 2 to IMF 6, inclu-

sive. From IMF 7 onwards, the similarity between

the IMFs is very high.

4 DISCUSSION

As has been veriﬁed in other studies such as Krish-

nan et al. (2021) or Liu et al. (2020), in the compari-

son of IMFs, the ﬁrst (high-frequency) IMFs are more

important for differentiating characteristics. Accord-

ing to our data, the means are a more reliable statistic

than the standard deviations for the ﬁrst IMFs. Con-

Analyzing Male Depression Using Empirical Mode Decomposition

889

Figure 4: Comparison of similarities in the distributions of

IMFs between depression and health applying the Gaussian

kernel, with the x-axis presented on a logarithmic scale.

versely, the standard deviations are less reliable in this

case. This is because as the IMF increases, the em-

pirical decomposition data approaches zero, and our

standard deviation data, as shown in Fig. 3, deviates

from zero. This indicates a greater dispersion of the

data, which does not effectively discriminate between

depression and healthy groups.

The differences between groups in the means for

IMF 1 may be due to low frequencies in the voices of

individuals with depression overlapping with noise,

which is more common in this IMF during the initial

phase of Empirical Mode Decomposition (EMD).

In the last IMF, the dispersion of data in the de-

pression group might increase the likelihood that the

trend toward low frequencies in the voices of individ-

uals with depression results in a higher probability of

belonging to this data range.

The results regarding the comparison of proba-

bility density indicate a more stable discrimination

of mean distributions across the ﬁrst IMFs compared

to standard deviations in the depression and healthy

groups. The use of means in identifying differences

appears to be more robust. In comparisons of stan-

dard deviation distributions, the distribution does not

seem to be a discriminative criterion. The probability

density functions of the means of the IMFs in differ-

entiating between depression and healthy characteris-

tics reﬂect a situation similar to that found with the

bootstrapping method corrected.

By examining the results in Fig.2 and Fig.3 and

comparing them with those from the Gaussian kernel

application in Fig. 4, we see that the Gaussian kernel,

by simulating a ’locally weighted mean’, is able to

extract differences more speciﬁcally than the mean or

standard deviation with the bootstrapping method or

density distributions.

In the kernel, the standard deviation controls the

extent of this inﬂuence: a small sigma considers only

the closest neighbors, preserving more details but

with less smoothing, while a large sigma covers a

broader region, resulting in more intense smoothing

that removes more noise but also more ﬁne details.

In this way, by adjusting sigma, a balance can be

achieved between noise reduction and the preserva-

tion of important features in the signal.

The analysis using various methods, particularly

the Gaussian kernel approach, reveals signiﬁcant dif-

ferences between depression and healthy states in the

ﬁrst 7 IMFs of voice data in men. This method proves

more effective than simple means or standard devia-

tions in discriminating between these groups, espe-

cially for IMFs 2 through 6. These differences in

IMFs could reﬂect subtle changes in voice charac-

teristics associated with depression in men. For in-

stance, the ﬁrst IMFs (1-3) might represent changes

in high-frequency components of the voice, possibly

indicating alterations in speech clarity or crispness in

depressed men. The intermediate IMFs (4-6) could

capture modiﬁcations in the mid-frequency range, po-

tentially related to changes in pitch and voice modula-

tion, which are often described as more monotonous

in depression. Lastly, IMF 7 might reﬂect varia-

tions in low-frequency components, possibly associ-

ated with changes in speech rhythm or cadence, which

tend to be slower in depressive states.

However, it is crucial to note that whilst these

mathematical differences are evident, the study lacks

a detailed explanation of how these differences trans-

late into clinically actionable insights for depression

diagnosis. While identifying these statistical dispar-

ities in voice characteristics is a promising start, the

clinical utility of such ﬁndings remains unclear with-

out additional investigation. Future studies could aim

to correlate speciﬁc IMF patterns with established

diagnostic criteria for depression in men, making

the ﬁndings more relevant to real-world applications.

Moreover, there is a need to explore how these IMF

changes relate to speciﬁc symptoms of male depres-

sion, such as irritability, aggression, or social with-

drawal. Such insights could enhance the understand-

ing of how voice analysis might aid in detecting or

monitoring depression and pave the way for develop-

ing non-invasive diagnostic tools.

BIOSIGNALS 2025 - 18th International Conference on Bio-inspired Systems and Signal Processing

890

5 CONCLUSION

Due to the minimal differences in the data ranges

and the rapid variations between them, distinguish-

ing their characteristics using basic statistical calcu-

lations, such as mean and standard deviation with the

bootstrapping method, is challenging. Despite apply-

ing the FDR correction, this method remains unsta-

ble. In contrast, the distribution density of the data

is somewhat more robust; however, due to the nar-

row range of the data, it has difﬁculties in differen-

tiating between depressed and healthy data based on

the IMF, especially when using the standard deviation

as a statistic.

The Gaussian kernel, while not complex or com-

putationally expensive, is better at distinguishing

characteristics by accounting for variance and per-

forming local weighting through the ﬁlter. It again

highlights the ﬁrst IMFs as relevant for differentiat-

ing features between depression and health.

We believe that future research should focus on

applying a more speciﬁc kernel, considering the range

and rapid ﬂuctuations in voice data. Additionally, in-

corporating a larger database with samples from both

genders would be beneﬁcial to analyze if there are

gender differences and whether these overlap with the

discrepancies between depression and health. In any

case, identifying depression through voice is a very

promising ﬁeld where, as we have seen, we are likely

to establish a diagnostic differentiation method poten-

tially useful in digital screening of depression.

Although it is not the aim of this study, we be-

lieve that it could be interesting to compare Gaussian

analysis of IMFs with machine learning methods or

to incorporate the results into such models to test pre-

diction methods. Furthermore, the integration of this

type of diagnosis into the clinical setting would be

crucial, allowing for the real-time assessment of pa-

tients with depression in hospital environments. Ethi-

cal applicability must be considered, as it involves the

collection and analysis of patients voices, which re-

quires their consent for the collection and handling of

their voice data.

ACKNOWLEDGEMENTS

X.S.C. carried out this work as part of the PhD

programme in Experimental Sciences and Technol-

ogy at the University of Vic - Central University

of Catalonia. We would like to thank the Univer-

sity of Southern California for providing voice data

and questionnaire information, without which this re-

search would not have been possible. We ﬁnally

thank the support of the Spanish Ministry of Sci-

ence and Innovation/ISCIII/FEDER (PI21/01148));

the Secretaria d’Universitats i Recerca del Departa-

ment d’Economia i Coneixement of the Generalitat de

Catalunya (2021 SGR 01431); the CERCA program

of the I3PT; the Instituto de Salud Carlos III; and the

CIBER of Mental Health (CIBERSAM).

CONFLICT OF INTEREST

D.P. has received grants and also served as a consul-

tant or advisor for Rovi, Angelini, Janssen and Lund-

beck, with no ﬁnancial or other relationship relevant

to the subject of this article. The other authors declare

no conﬂicts of interest.

REFERENCES

Akkaralaertsest, T. and Yingthawornsuk, T. (2019). Classi-

ﬁcation of depressed speech samples with spectral en-

ergy ratios as depression indicator. In 2019 IEEE Sym-

posium Series on Computational Intelligence (SSCI).

Alghowinem, S., Goecke, R., Wagner, M., Epps, J.,

Gedeon, T., Breakspear, M., and Parker, G. (2013).

A comparative study of different classiﬁers for de-

tecting depression from spontaneous speech. In 2013

IEEE International Conference on Acoustics, Speech

and Signal Processing (ICASSP), pages 8022–8026.

IEEE.

Audacity (2023). http://sourceforge.net/projects/audacity/.

(accessed 1 October 2023).

Chen, L., Wang, C., Chen, J., Xiang, Z., and Hu, X. (2021).

Voice disorder identiﬁcation by using hilbert-huang

transform (hht) and k nearest neighbor (knn). Jour-

nal of Voice, 35(6):932.e1–932.e11.

Espinola, C. W., Gomes, J. C., Pereira, J. M. S., and dos

Santos, W. P. (2022). Detection of major depressive

disorder, bipolar disorder, schizophrenia and general-

ized anxiety disorder using vocal acoustic analysis and

machine learning: An exploratory study. Research on

Biomedical Engineering, 38(3):813–829.

Esposito, A., Faundez-Zanuy, M., Esposito, A. M., Cor-

dasco, G., Drugman, T., Sol

e-Casals, J., and Mora-

bito, F. C. (2016). Recent Advances in Nonlin-

ear Speech Processing: Directions and Challenges.

Springer.

Gratch, J., Artstein, R., Lucas, G., Stratou, G., Scherer,

S., Nazarian, A., Wood, R., Boberg, J., DeVault, D.,

Marsella, S., Traum, D., Rizzo, S., and Morency,

L.-P. (2014). The distress analysis interview corpus

of human and computer interviews. In LREC 2014

- Ninth International Conference on Language Re-

sources and Evaluation, pages 3123–3128. European

Language Resources Association (ELRA).

Krishnan, P., Joseph Raj, A., and Rajangam, V. (2021).

Emotion classiﬁcation from speech signal based on

Analyzing Male Depression Using Empirical Mode Decomposition

891

empirical mode decomposition and non-linear fea-

tures: Speech emotion recognition. Complex and In-

telligent Systems, 7(4):1919–1934.

Kroenke, K., Strine, T. W., Spitzer, R. L., Williams, J.

B. W., Berry, J. T., and Mokdad, A. H. (2009).

The phq-8 as a measure of current depression in the

general population. Journal of Affective Disorders,

114(1–3):163–173.

Liu, Z., Xu, Y., Ding, Z., and Chen, Q. (2020). Time-

frequency analysis based on hilbert-huang transform

for depression recognition in speech. In 2020

IEEE International Conference on Bioinformatics and

Biomedicine (BIBM), pages 1072–1076.

Lopez-de Ipi

na, K., Alonso-Hern

andez, J., Sol

e-Casals,

J., Travieso-Gonz

alez, C. M., Ezeiza, A., Faundez-

Zanuy, M., Calvo, P. M., and Beitia, B. (2015). Fea-

ture selection for automatic analysis of emotional re-

sponse based on nonlinear speech modeling suitable

for diagnosis of alzheimer’s disease. Neurocomput-

ing, 150:392–401.

opez-de Ipi

na, K., Sol

e-Casals, J., Eguiraun, H., Alonso,

J. B., Travieso, C. M., Ezeiza, A., Barroso, N., Ecay-

Torres, M., Martinez-Lage, P., and Beitia, B. (2015).

Feature selection for spontaneous speech analysis to

aid in alzheimer’s disease diagnosis: A fractal di-

mension approach. Computer Speech & Language,

30(1):43–60.

Mekyska, J., Galaz, Z., Kiska, T., Zvoncak, V., Mucha, J.,

Smekal, Z., Eliasova, I., Kostalova, M., Mrackova,

M., Fiedorova, D., Faundez-Zanuy, M., Sol

e-Casals,

J., Gomez-Vilda, P., and Rektorova, I. (2018). Quan-

titative analysis of relationship between hypokinetic

dysarthria and the freezing of gait in parkinson’s dis-

ease. Cognitive Computation, 10(6):1006–1018.

Sharma, R., Prasanna, S. R. M., Bhukya, R. K., and Das,

R. K. (2017). Analysis of the intrinsic mode func-

tions for speaker information. Speech Communica-

tion, 91:1–16.

Sol

e-Casals, J., Munteanu, C., Mart

ın, O. C., Barb

e, F.,

Queipo, C., Amilibia, J., and Dur

an-Cantolla, J.

(2014). Detection of severe obstructive sleep ap-

nea through voice analysis. Applied Soft Computing,

23:346–354.

Sol

e-Casals, J., Zaiats, V., and Monte-Moreno, E. (2010).

Non-linear and non-conventional speech process-

ing: Alternative techniques. Cognitive Computation,

2:133–134.

Tasnim, M. and Novikova, J. (2022). Cost-effective mod-

els for detecting depression from speech. In Wani,

M. A., Kantardzic, M., Palade, V., Neagu, D., Yang,

L., and Chan, K. Y., editors, 2022 21ST IEEE INTER-

NATIONAL CONFERENCE ON MACHINE LEARN-

ING AND APPLICATIONS (ICMLA), pages 1687–

1694. IEEE Computer Soc.

Tong, J., Xing, Z., Wei, X., Yue, C., Dong, E., Du, S.,

Sun, Z., Sol

e-Casals, J., and Caiafa, C. F. (2023). To-

wards improving motor imagery brain–computer in-

terface using multimodal speech imagery. Journal of

Medical and Biological Engineering, 43(3):216–226.

BIOSIGNALS 2025 - 18th International Conference on Bio-inspired Systems and Signal Processing

892