A Machine Learning-Based Approach for Non-Invasive Stress

Analysis Using Speech Features for Enhanced Detection and

Classification

Devika M, Shruthi R, Sahana K and Madhubala V

Department of Computer Science and Engineering, SRM Institute of Science and Technology, Ramapuram, Chennai,

Tamil Nadu, India

Keywords: K-NN (K-Neighbour Nearest), Stress, HRV, Machine Learning, Speech Analysis, Feature Extraction.

Abstract: Stress has its impact on our physical and mental health in many forms, and thus it becomes even more

important to detect it at an early stage so that we can manage it better. Standard approaches to monitoring

the stress response typically involve tracking a person’s physiological state or behavioural differences with

the help of invasive sensors or elaborate data collection systems. As a more appealing non-intrusive option,

this research paper proposes a machine learning model, to observe our speech patterns to monitor our stress

levels on a real-time basis. By analysing features such as pitch, tone, frequency and the rate at which

someone speaks, the system can gain a better understanding of a person’s emotional state. These pieces of

information are collected by techniques like Mel Frequency Cepstral Coefficient (MFCC), Fast Fourier

Transform (FFT), and Spectral Analysis. These extracted features are then used to train multiple machine

learning classifiers such as Decision Trees and K-Nearest (K-NN) Neighbours. This enables you to classify

stress levels into three specifics categories (Neutral, Mild Stress, High Stress). This analysis allows us to

make sure the model is reliable and its performance is understood using standard metrics including but not

limited to accuracy, precision, recall and F1-score. The idea is to make stress detection more achievable

and operable every day.

1 INTRODUCTION

Stress represents a major challenge to an individual's

adaptive capacity, triggering both cerebral and

physiological changes that can pave the way for

severe health conditions. These include

hypertension, coronary artery disease, cardiac arrest,

stroke, and mental health issues such as depression

and anxiety. Stress can be categorized into two

distinct types: short term acute stress and long-term

chronic stress. Ultimately, the body relies on the

parasympathetic nervous system (PNS) to restore

balance and return to homeostasis.

To evaluate the stress response, we can assess

perceptual, behavioural, and physical indicators.

Self-report questionnaires serve as a crucial tool in

measuring an individual's subjective perception of

stress levels. Many physiological markers

responsive to stress have been thoroughly studied.

Stress analysis via machine learning, where

frequency domain features derived from audio

signals are employed in classifying emotional states

reflecting varying levels of stress, is the thrust of

this project. The dataset consists of FFT (Fast

Fourier Transform) coefficients, which represent the

frequency content of speech signals. By examining

these characteristics, we train and test two machine

learning models- Decision Tree Classifier and K-

Nearest Neighbours (KNN) Classifier. The main

aim is to establish how effectively the models can

classify emotions (Negative, Neutral, and Positive)

and map them to possible stress levels. Model

performance is tested using accuracy scores and

confusion matrices to determine patterns of

classification. This research has applications in

mental health analyzing, workplace stress

management, and healthcare, where automated

stress detection can provide valuable insights for

intervention and support.

The structure of this paper is divided as

following, section II discusses about the various use

cases of the Stress analysis, section III covers the

related works of the Stress analysis, section IV

M., D., R., S., K., S. and V., M.

A Machine Learning-Based Approach for Non-Invasive Stress Analysis Using Speech Features for Enhanced Detection and Classiﬁcation.

DOI: 10.5220/0013896800004919

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

In Proceedings of the 1st International Conference on Research and Development in Information, Communication, and Computing Technologies (ICRDICCT‘25 2025) - Volume 3, pages

285-291

ISBN: 978-989-758-777-1

285

illustrates the existing system, section V discusses

the proposed system of the project, section VI

discusses the result and section VII discusses the

Experimental result of the project. VII discusses the

conclusion of the project.

1.1 Characteristics of Stress Analysis

Below are the characteristics of Stress analysis

1.1.1 Feature Extraction from Speech

Signals

Stress analysis often relies on extracting features

from speech, such as FFT (Fast Fourier Transform)

coefficients, Mel Frequency Cepstral Coefficients

(MFCCs), and pitch variations, which help in

identifying emotional states.

1.1.2 Emotion-Based Stress Classification

Stress levels are linked to emotions, commonly

classified into Negative, Neutral, and Positive states.

Machine learning models analyse these emotions to

infer stress intensity.

1.1.3 Supervised Learning Models

Algorithms like Decision Tree Classifier and K-

Nearest Neighbours (KNN) are trained on labeled to

identify stress pattern from input features

1.1.4 Data Pre-processing and Label

Mapping

Raw data undergoes pre-processing methods such as

normalization, feature selection, and label encoding

to improve model performance and interpretability.

1.2 Advantages of Stress Analysis

1. Early stress detection

2. Automated and Objective Analysis

3. Personalized Stress Management

4. Non-Intrusive and User-Friendly

This is just a fraction of the benefits of machine

learning-based stress analysis that promise a new era

of support for mental health, workplace well-being

and healthcare. Its greatest benefit is early detection

of stress, enabling interventions sooner that can

have a significant impact in preventing later mental

and physical health problems. In contrast to

conventional techniques which tend to depend on

self-reports and can be subjective and biased,

machine learning provides a more objective and

automated solution which results in higher accuracy.

Because these models can learn from individual

stress profiles, they can provide tailored advice. The

cost-effective and large-scale stress measurement

through smart devices minimizes the need of

frequent visits to doctor. In comparison, combining

speech, heart rate, EEG signals, and facial

expressions from these advanced systems provides a

more comprehensive and reliable stress detection

profiling compared to single channel systems.

2 USE CASES OF STRESS

ANAYSIS USING MACHINE

LEARNING

This stress analysis model uses machine learning to

analyse speech-based features and classify emotions

into different stress levels. Below is an in-depth

explanation of how this model can be applied in

various industries and real-world applications.

1. Mental Health Monitoring & Early

Intervention

2. Workplace Stress Management

3. Call Centre & Customer Service Analysis

4. Education & Student Stress Tracking

5. High-Risk Professions (Military, Aviation,

Healthcare, etc.)

2.1 Mental Health Monitoring & Early

Intervention

The model processes speech data, extracts

frequency-based features, and classifies the

emotional state into negative (high stress), neutral

(moderate stress), or positive (low stress)..It can be

integrated into mental health applications to monitor

users' emotions over time. Helps in early detection

of stress, anxiety, and depression. Provides data-

driven insights to mental health professionals for

better patient care.

2.2 Workplace Stress Management

The model can be used to monitor employees' stress

levels through voice-based interactions during

meetings, calls, or check-ins. HR teams can analyze

collective stress trends and make data-driven

ICRDICCT‘25 2025 - INTERNATIONAL CONFERENCE ON RESEARCH AND DEVELOPMENT IN INFORMATION,

COMMUNICATION, AND COMPUTING TECHNOLOGIES

286

workplace improvements. Improves employee well-

being and productivity. Helps prevent burnout and

workplace dissatisfaction.

2.3 Call Centre & Customer Service

Analysis

The model can analyse customer service call

recordings to detect stress levels in both employees

and customers. If high stress is detected in

employees, breaks or training programs can be

recommended. If high stress is detected in

customers, the system can escalate the call to a

human representative with special skills. Enhances

customer satisfaction by detecting and responding to

stress signals.

Reduces employee burnout by managing high-stress

interactions.

2.4 Education & Student Stress

Tracking

Teachers and education platforms can use stress

analysis to track students’ emotional well-being

during virtual or in-person classes.

Stress levels can be analysed to adjust teaching

methods and exam difficulty. Helps prevent

academic burnout and anxiety. Improves teaching

strategies based on real-time student feedback.

2.5 High-Risk Professions (Military,

Aviation, Healthcare, etc.)

The model can monitor stress in high-pressure

professions like pilots, surgeons, soldiers, and

emergency responders. It can detect stress levels

from voice-based communication and alert

supervisors when stress levels are dangerously high.

Reduces human errors and accidents in critical

operations. Improves mental resilience and decision-

making in high-risk environments.

3 RELATED WORK

Stress response is defined as an evoked response

when our body perceives any stimuli that exceeds an

organism’s adaptive capacity and disrupt

homeostasis.

This paper (J. Lee et al., 2024) analysed the

impact of the personal Physical stress factor Stress

response and recovery and one operated one

Comparative analysis of stress reactivity and

Recovering between HRV and EDA parameters

Group. Stress situation may begin Human body is

identified and measured Represented for stress with

a defence mechanism (A. Ferrarotti et al., 2024) by

physical variation.

This paper (A. Ferrarotti et al., 2024) mainly

relies on invasive systems and infrequently focuses

on AR/VR applications. this work investigates

whether the head movements of a user wearing an

AR HMD vary due to the presence of a stress factor

while performing static tasks. In order to induce

stress, the SCWT has been used.

This paper ( S. Santhiya et al., 2024) indicates

how audio-based Stress detection system can be

used to improve systems General welfare and

preventive health measures. To make accurate

production and Trusted results, this study proposes a

novel method of detecting stress from deep learning

Algorithm with sophisticated signal processing

Technology.

This paper (A. Singh et al., 2024) proposes a

comprehensive approach which uses the

ccombination of computational power, Software,

technology and sympathy care to determine

similarities in frequency pattern stress level. In our

sharp world, stress is affected Countless people

affect their mental Goodness. 48% gene z adults

report Accepted, sad, depressed, (N. Oryngozha et

al., 2024) experience Fofo (fear of disappearance),

and reduced Self -esteem or insecurity.

This research N. Oryngozha et al., 2024)

proposes ML and NLP to recognize and evaluate

stress related posts and comments within Reddit’s

academic communities which can be used to analyze

comments from text. Stress monitoring is a crucial

component (Z. Lei el al., 2024) of any strategy used

to intervene in the case of stress.

This study (H. A. Khan et al., 2023) proposes for

an outline using a wearable sensor, safely an

artificial intelligence driven in associated with

Cloud-based server to initial detection High blood

pressure and an intervention facility System.

This study (Mittalakod et al., 2023) improves

sentiment and emotion analysis to measure the stress

levels of people by analysing their social posts and

comments using complex machine learning

techniques and the deep learning model BERT.

This paper (J. G. Jayawickrama and R. A. H. M.

Rupasingha, 2022) explores how we can detect

human stress by analysing sleep patterns, using a

technique called ensemble learning. In the first part

of the research, five different machine learning

algorithms were tested for classification, including

Decision Tree, Logistic Regression, and Naive

A Machine Learning-Based Approach for Non-Invasive Stress Analysis Using Speech Features for Enhanced Detection and Classiﬁcation

287

Bayes. Then, in the second part, they applied an

ensemble learning algorithm that combined the

results of these five methods using an average

probability approach. The findings were promising,

with ensemble learning achieving an 94.25%

accuracy in classifying the data. This suggests that

using a combination of algorithms can significantly

enhance our ability to understand and identify stress

based on sleep patterns.

This paper (F. J. Ming et al., 2023) proposes to

develop a Facial Emotion Recognition System

designed to help identify mental stress among

university students. The goal is to provide support

both to the students themselves and to the

counselling departments within institutions. The

system will analyse facial expressions to recognize

various emotions, including happiness, sadness,

anger, and fear.

4 EXISTING SYSTEM

Figure 1: Existing Stress Analysis model.

The following study presents a non-invasive means

of detecting stress within Augmented Reality (AR)

applications based on head movement during the use

of Head-Mounted Displays (HMDs). Using this

system, engaged with various AR tasks, stress levels

were identified using machine learning with the

main algorithm being Support vector machines

(SVMs). Headset data is logged and converted to be

analysed to identify similarities across participants.

Using a weighted decision model, an SVMs training

process is initiated in an effort to enhance their

classification accuracy in predicting stress levels.

However, it is worth mentioning that there are some

limitations to this system. The figure 1 shows the

Existing Stress Analysis model. For one, the study

had a relatively small number of people, meaning its

results may not be generalizable to everyone. There

isnt a lot of diversity among participants, either, in

terms of age and gender and backgrounds — which

could affect how the model performs across various

groups of people.

5 PROPOSED SYSTEM

Figure 2: Proposed System for stress Analysis.

The existing system relies on a fairly limited sample

size, which really holds it back from being as

effective as it could be. In contrast, our new system

ICRDICCT‘25 2025 - INTERNATIONAL CONFERENCE ON RESEARCH AND DEVELOPMENT IN INFORMATION,

COMMUNICATION, AND COMPUTING TECHNOLOGIES

288

takes a big step forward by including a larger and

more diverse group of participants, significantly

boosting our data set. One of the key differences is

that while the existing approach only analyses head

movement data after tasks are completedessentially

looking at it afterward we’ve developed an

innovative solution that detects stress in real time.

We leverage streaming machine learning models to

continuously analyse head movement data, allowing

us to gain immediate insights and react promptly.

The figure 2 shows the Proposed System for stress

Analysis.In our evaluation process, we closely

compared Support Vector Machine (SVM) and K-

Nearest Neighbours (K-NN) to see which method

fits our goals best. We ultimately decided to go with

K-NN because of its strength in classification.

This algorithm is great at categorizing data based

on class membership, which means it can effectively

decide whether a data point belongs to group A or

group B. K-NN works by looking at the nearest

neighbours, with 'k' representing the number of

closest points considered—usually a small, positive

integer. We also looked into Decision Trees (DT),

which create classification models in a way that's

easy to understand by forming a tree structure. This

technique breaks down the dataset into smaller

pieces while building a corresponding decision tree.

The result is a model that’s clear and interpretable,

showing decision nodes and leaf nodes that represent

the underlying class data. Our approach not only

pushes the limits of what can be achieved in real-

time stress detection but also establishes a new

benchmark in the field.

Working of Stress Analysis

The Stress analysis project is structured into multiple

functional steps based on its working

1. Data Acquisition Module

2. Data Pre-processing Module

3. Feature Visualization Module

4. Model Training Module

5. Prediction Module

5.1 Data Acquisition Module

The Data Acquisition Module plays a crucial role in

gathering the data needed for stress analysis. It uses

a CSV file that includes speech recordings along

with their associated stress labels. This dataset

features numerical characteristics extracted from the

speech recordings, alongside labels that indicate

various levels of stress. Essentially, this module

makes sure that we have all the necessary

information at hand for the next steps in processing

and analysis.

5.2 Data Pre-Processing Module

pre-processing is an inevitable procedure in any

machine learning process Similar to the way in that

we are working out the Functions here now is the

data cleaning, feature selection or label encoding.

First, we handle any non-existent values to ensure

uniformity of the data. And then we will replace the

categorical one like NEGATIVE, NEUTRAL and

POSITIVE into numbers (0,1,2) that will help us to

becoming compatible with different machine

learning algorithms. Lastly, we ready the data for

exploration and modelling.

5.3 Feature Visualization Module

Understanding the characteristics of the features is

critical before you start training the model. In this

module we utilize Matplotlib to plot sample

distributions of the features so that we can visualize

the variation of the frequency domain features with

stress levels. Visualizing this data allows us to

identify patterns and trends that may not be so

obvious, which greatly improves our feature

selection and increases the performance of the

model.

5.4 Model Training Module

Module 4 Training ML models to classify the level

of stress These features are fed to algorithms

(Decision Tree Classifier, K-Nearest Neighbours

(KNN)) capable of recognising patterns in the

training data. These classifiers generate predictive

models based on the correlation between the features

of head movement over the three different levels of

stress labels seen in Figs. It’s a cool example of how

to use data and technology to sharpen knowledge

about human emotions.

5.5 Predicting Module

When the training phase is complete, the system

uses the models it learned from the training data to

assess stress on new data that was not seen in the

training phase. These trained classifiers determine

whether a certain sample is NEGATIVE,

NEUTRAL or POSITIVE stress level when the

testing set is provided. These predictions are then

A Machine Learning-Based Approach for Non-Invasive Stress Analysis Using Speech Features for Enhanced Detection and Classiﬁcation

289

checked against the actual stress levels to measure

classifier accuracy.

Figure 3: Comparison between Existing system VS

Proposed system.

Figure 3 shows a comparative graph that clearly

depicts the performance improvement between the

existing system and the system that is proposed in

this work. The model showed the best computational

efficiency, training speed, interpretability, data

intelligence, memory sage, and real-time

performance. More specifically, the proposed system

achieves high scores for computational efficiency,

interpretability, memory, and real-time use, whereas

the existing system performs low in these aspects.

Further, compared to other systems that require large

amounts of data but achieve similar accuracy rates,

the proposed system is more optimized and requires

less data to train on. Once again, the training speed

is better, allowing the model to learn faster. In this

way, these developments contribute to the proposed

system to be used in real time stress analysis

applications where speed and efficiency of data

acquisition and decision making are critical.

6 CONCLUSIONS

By analysing how we move our heads, we can have

a contemporary and more useful method of assessing

level of stress using machine learning. The system

can classify stress into 3 types: NEGATIVE,

NEUTRAL, POSITIVE using smart algorithms like

Decision Tree Classifier and K-Nearest Neighbours

(KNN). It is also structured into multiple

components data collection, data cleaning,

visualization, model training, and model evaluation

so it’s also accurate and reliable. Its most notable

benefit, compared to current solutions, is that there

are no physiological sensors required, making it

more practical for usage in real-life scenarios,

particularly in Augmented Reality (AR)

environments. Tools such as confusion matrices and

performance metrics help boost the model’s strength

over time, so that its predictive power can continue

improving. As for the future, there are exciting

potential upgrades, including real-time stress

monitoring, more physiological indicators, and even

advanced deep learning techniques to boost

accuracy.

In short, this machine learning system is a

considerable step towards stress analysis in real-

time. It has potential use in the fields of health care,

workplace efficiency, and communication between

humans and computers. You know, with a few more

nudges and improvements it can be a powerful tool

to help better understand and cope with stress in our

lives.

REFERENCES

J. Lee et al., "Analysis of Acute Stress Reactivity and

Recovery in Autonomic Nervous System Considering

Individual Characteristics of Stress Using HRV and

EDA," in IEEE Access, vol. 12, pp. 115400-115410,

2024, doi: 10.1109/ACCESS.2024.3437671

10, pp. 6970- A. Ferrarotti, S. Baldoni, M. Carli and F.

Battisti, "Stress Assessment for Augmented Reality

Applications Based on Head Movement Features," in

IEEE Transactions on Visualization and Computer

Graphics, vol. 30, no.6983, Oct. 2024.

S. Santhiya, P. Jayadharshini, N. Abinaya, K. Sruthi, N.

Kavin Vishnu and V. Suganth, "Advancing Stress

Detection through Deep Learning in Human-Machine

Interactions with Speech Signals Analysis," 2024 15th

International Conference on Computing Communicati

on and Networking Technologies (ICCCNT), Kamand,

India, 2024, pp. 1-5, doi:

10.1109/ICCCNT61001.2024.10726085.

A. Singh, A. Agarwal, A. Kumar, U. Rastogi and S.

Kumar, "A Comprehensive Approach to Stress

Detection and Management Using Machine Learning,

Deep Learning, and Chatbot Integration," 2024 15th

International Conference on Computing Communicat-

ion and Networking Technologies (ICCCNT),

Kamand, India, 2024, pp. 1-5, doi:

10.1109/ICCCNT61001.2024.10725990.N.

N. Oryngozha, P. Shamoi and A. Igali, "Detection and

Analysis of Stress-Related Posts in Reddit’s Acamedic

Communities," in IEEE Access, vol. 12, pp. 14932-

14948, 2024, doi: 10.1109/ACCESS.2024.3357662

Z. Lei, Y. Zhou, G. Zhou, H. Su, Z. Yan and Q. Xiong,

"Welding Residual Stress Analysis and Its Influence

Study on Warm Rotor of an HTS Motor," in IEEE

Transactions on Applied Superconductivity, vol. 34,

no. 8, pp. 1-5, Nov. 2024, Art no. 5206105, doi:

10.1109/TASC.2024.3446284.

H. A. Khan, T. N. Nguyen, G. Shafiq, J. Mirza and M. A.

Javed, "A Secure Wearable Framework for Stress

Detection in Patients Affected by Communicable

Diseases," in IEEE Sensors Journal, vol. 23, no. 2, pp.

ICRDICCT‘25 2025 - INTERNATIONAL CONFERENCE ON RESEARCH AND DEVELOPMENT IN INFORMATION,

COMMUNICATION, AND COMPUTING TECHNOLOGIES

290

981-988, 15 Jan.15 2023, doi:

10.1109/JSEN.2022.3204586

A. G, K. B, P. S. Mittalakod, S. K and B. Ambore,

"Detection of Stress Over Social Interactions Using

Machine Learning and Natural Language Processing,"

2023 6th International Conference on Contemporary

Computing and Informatics (IC3I), Gautam Buddha

Nagar, India, 2023, pp. 1771-1775, doi:

10.1109/IC3I59117.2023.10397916.

J. G. Jayawickrama and R. A. H. M. Rupasingha,

"Ensemble Learning Approach to Human Stress

Detection Based on Behaviours During the Sleep,"

2022 4

International Conference on Advancements in

Computing (ICAC), Colombo, Sri Lanka, 2022, pp.

132-137, doi: 10.1109/ICAC57685.2022.10025175.

F. J. Ming, S. Shabana Anhum, S. Islam and K. H. Keoy,

"Facial Emotion Recognition System for Mental Stress

Detection among University Students," 2023 3rd

International Conference on Electrical, Computer,

Communications and Mechatronics Engineering

(ICECCME), Tenerife, Canary Islands, Spain, 2023,

pp. 1-6, doi:

10.1109/ICECCME57830.2023.10252617.

S. K. Sriramulugari, V. A. K. Gorantla, A. H. Mewada, K.

Gupta and T. Kiruthiga, "The opinion-based analysis

for stressed adults using sentimental mining model,"

2023 IEEE Technology & Engineering Management

Conference - Asia Pacific (TEMSCON-ASPAC),

Bengaluru, India, 2023, pp. 1-6, doi:

10.1109/TEMSCON-ASPAC59527.2023.10531336.

A. Patel, D. Nariani and A. Rai, "Mental Stress Detection

using EEG and Recurrent Deep Learning," 2023 IEEE

Applied Sensing Conference (APSCON), Bengaluru,

India, 2023, pp. 13,

doi:10.1109/APSCON56343.2023.10100977.

K. Singh, S. K. Chawla, G. Singh and P. Soni, "Stress

Detection using Machine Learning Techniques: A

review," 2023 5th International Conference on

Advances in Computing, Communication Control

and Networking (ICAC3N), Greater Noida, India,

2023, pp. 255-260, doi:

10.1109/ICAC3N60023.2023.10541696.

K. Tao, Y. Huang, Y. Shen and L. Sun, "Automated

Stress Recognition Using Supervised Learning

Classifiers by Interactive Virtual Reality Scenes," in

IEEE Transactions on Neural Systems and

Rehabilitation Engineering, vol. 30, pp. 2060-2066,

2022, doi: 10.1109/TNSRE.2022.3192571

A. Akella et al., "Classifying Multi-Level Stress

Responses from Brain Cortical EEG in Nurses and

Non-Health Professionals Using Machine Learning

Auto Encoder," in IEEE Journal of Translational

Engineering in Health and Medicine, vol. 9, pp. 1-9,

2021

A Machine Learning-Based Approach for Non-Invasive Stress Analysis Using Speech Features for Enhanced Detection and Classiﬁcation

291