CHAPTER 1

1.1 INTRODUCTION

Mental well-being is as important as physical health in today’s fast-paced and

often stressful lifestyle. However, traditional mental health support systems
can be expensive, inaccessible, or avoided due to social stigma and fears,
leaving many individuals—especially youth, students, and working
professionals—without timely help for managing stress, anxiety, and
emotional imbalances. Advances in technology and artificial intelligence (AI)
now offer innovative ways to bridge this gap by providing interactive,
personalized, and accessible mental health support. The AI Mental Health
Buddy is designed to be a virtual companion that uses Natural Language
Processing (NLP) to understand a user’s emotions from voice or text input,
detect mood states, and respond with personalized healing suggestions such as
motivational quotes, calming music, breathing exercises, and mood tracking.
This web-based application features a clean, user- friendly interface that
supports both voice and text interaction, making it easier for users to
communicate according to their preference. The system also prioritizes user
privacy by storing mood data locally without requiring login or signup,
ensuring confidentiality and fostering trust. By leveraging advanced sentiment
analysis and machine learning techniques, this AI companion empowers users
to gain better insight into their emotional well-being and offers real-time
assistance tailored to their individual needs, making mental health support
more accessible and effective for those facing daily emotional challenges.

The AI Mental Health Buddy incorporates sophisticated emotion detection

capabilities that go beyond basic sentiment analysis to identify a wide range of
emotional states, including complex feelings like burnout, hopelessness, and
anxiety, as well as critical states such as suicidal ideation, enabling timely and
appropriate support. The platform further enhances user engagement and
efficacy by integrating mood tracking analytics, allowing users to visualize
their emotional trends over time and better understand their mental health
patterns. Through this holistic and empathetic approach, the AI Mental Health
Buddy aims to reduce the barriers to mental health care, and always-available
support system that complements traditional therapy and promotes emotional
resilience and well-being in everyday life.
 CHAPTER 2
SYSTEM ANALYSIS

2.1 LITERATURE SURVEY

2.1.1 PURPOSE

The aim of this research paper is to tackle the emerging mental health
epidemic that is hampering students, working professionals, and young adults
by creating a novel AI-driven solution that surmounts the profound
impediments that exist in the conventional mental health care systems. As
mental illnesses like stress, anxiety, depression, and emotional instability are
spreading fast in the fast-moving digital age, most people still refuse to go for
help owing to social stigma, exorbitant prices, lack of accessibility, and long
queues for expert services.

This research seeks to illustrate how advanced technologies such as

tokenization, dependency parsing, semantic analysis, and machine learning
algorithms like RoBERTa and GPT-4 Mini can be successfully combined to
design an all-encompassing emotional intelligence system that can identify not
only simple emotions but also sophisticated mental health-oriented states such
as overwhelm, burnout, hopelessness, and crisis situations needing expert
intervention. The ultimate goal is to determine if this AI-powered method has
the capability to effectively offer real-time, personalized emotional care
through voice and text interactions, mood monitoring features, and
customized healing recommendations in a bid to democratize mental health
care and turn emotional well-being support as accessible and ubiquitous as
physical health tracking Existing problem .

2.1.2 EXISTING SYSTEM

In digitally connected world, millions of people—especially students, young

adults, and those balancing high-stress careers—battle stress, anxiety,
depression, and emotional exhaustion, but have trouble or are unable to access
timely mental health care because of financial prohibitive costs, extensive
waiting lists, geographical barriers, and the lingering stigma of seeking help.
Although a growing number of AI-driven chatbots and wellness apps promise
round-the-clock assistance, most rely on simplistic sentiment analysis that fails
to grasp the nuance of complex emotional states, offer only generic responses,
mishandle crisis situations, compromise user privacy through data sharing,
and lack genuine empathy or personalized engagement. Thus, users tend to
drop out in the long run, depriving this at-risk population of the supportive,
adaptive, and secure interventions they must have in order to cope with their
mental health issues.

2.1.2 PROPOSED SOLUTION

The proposed solution integrates deep learning techniques, specifically a

CNN model, combined with MediaPipe for real-time pose estimation. By
leveraging artificial intelligence, this system provides accurate posture
classification and feedback. The use of angle detection enables precise analysis
of key body joints, allowing for an effective correction mechanism that works
seamlessly in real-world environments. Unlike traditional methods, this
solution adapts to individual body types and environmental variations,
ensuring a personalized experience. By implementing real-time tracking, users
receive instant posture alerts, reducing the chances of developing chronic
musculoskeletal disorders. Furthermore, the integration of cloud-based
analytics allows for long-term tracking and assessment of user progress. The
proposed system is designed to be cost-effective and easily deployable across
multiple platforms, making it accessible to a wide range of users. Additionally,
the system incorporates an intuitive user interface that enhances accessibility
for individuals with varying levels of technical expertise. By utilizing AI-
driven insights, users can track their posture trends over time, receiving
personalized recommendations for improvement. The integration of a feedback
loop allows the model to continuously learn and adapt to user-specific
movement patterns, thereby refining its accuracy with prolonged usage.
Furthermore, this solution can be extended to workplaces, fitness centers, and
rehabilitation clinics, providing a holistic approach to posture correction and
ergonomic well-being. With the ability to function on both mobile and desktop
platforms, the system ensures seamless usability, making it a versatile tool for
maintaining long-term spinal health and overall body alignment.
 2.2 SYSTEM REQUIREMENTS

2.2.1 NON-FUNCTIONAL REQUIREMENTS

2.2.2 FUNCTIONAL REQUIREMENTS

The system shall support both voice and text user input and automatically
transcribe voice to text with a minimum of 90% accuracy. The system shall
examine every user input in real time to identify emotional states—
happiness, sadness, worry, burnout, or crisis—utilizing a fine-tuned NLP
model. The system will provide and send customized healing
recommendations (e.g., inspirational quotes, soothing music links, guided
breathing techniques, CBT-informed prompts) according to the recognized
emotion in two seconds of processing time. The system will enable users to
 see and export a timestamped mood-tracking history, with detected emotions
and suggested interventions. The system will automatically initiate a crisis
alert workflow—offering emergency resources and prompting user
outreach— upon detection of life-threatening emotions (e.g., suicidal
ideation). The system will allow for smooth switching between voice and
text modes in one conversation, maintaining context across input types. The
system will store all the interaction data (transcripts, emotions detected, user
feedback) locally on the user's device in encrypted form, not needing user
accounts or cloud storage. The system will enable users to personalize
notification settings (e.g., frequency of wellness check-ins, times for exercise
reminders) and adjust its interaction style with time to accommodate each
user's changing mood patterns.

2.3 SOFTWARE AND HARDWARE REQUIREMENTS

2.3.1 SOFTWARE REQUIREMENTS

OPERATING SYSTEM: Windows 10 or later,

macOS 10.14+ or any recent Linux distribution
(Ubuntu 18.04+)

RUNTIME ENVIRONMENT: [Link] 14+ for front-end server, Python

3.8+ for back-end AI services

WEB FRAMEWORK: [Link] or Angular for UI,

[Link] or Flask for API routing

NLP LIBRARIES: NLTK, spaCy, HuggingFace

Transformers (BERT, RoBERTa)

VOICE PROCESSING: Python Whisper or Google

Cloud Speech-to- Text SDK

MACHINE LEARNING: PyTorch or TensorFlow for

model training and inference

DATA STORAGE: Local IndexedDB or SQLite for

mood logs, AES- 256 encryption libraries
 SECURITY: TLS 1.3 support, OpenSSL for
encryption, OAuth 2.0 libraries for optional
integrations

TESTING & CI: Jest/Mocha for JavaScript testing,

PyTest for Python, GitHub Actions or Jenkins for
pipelines

MONITORING : Prometheus + Grafana or

equivalent for real-time health metrics

2.3.2 HARDWARE REQUIREMENTS

CLIENT DEVICES: Any microphone and speaker-capable

Device at least dual-core CPU, 2 GB RAM, recent web browser

SERVER (HOSTED DEPLOYMENTS): Multi-core CPUs

(minimum 4 cores), 8 GB+ RAM, SSD storage, and optionally
GPU support (NVIDIA T4 or similar) for ML model inference
acceleration

NETWORK: Minimum 10 Mbps upload/download broadband

for API call and voice data transport

BACKUP AND STORAGE: Backup to External HDD or

encrypted backup to secure cloud storage

SECURITY: Firewall and network security appliances that

support TLS and intrusion detection to secure data in transit
 CHAPTER 3

METHODOLOGIES

The process entails comprehension of user inputs in form of voice or text,

speech-to-text conversion using models such as Whisper or Google Speech
API, text processing using NLP methods such as tokenization and sentiment
analysis, emotion detection using sophisticated models such as BERT or
RoBERTa, and subsequently producing empathetic responses based on the
emotional state of the user. The system keeps learning from user interactions
to enhance its support, retains data locally for privacy reasons, and makes the
platform safe by recognizing crisis situations and giving immediate assistance.
This step-by-step process assists in developing a responsive, precise, and
empathetic mental health chatbot.

3.1. Natural Language Processing (NLP) Framework

3.1.1. Architecture
AI Mental Health Chatbot is meant to give compassionate, real-time
emotional assistance via voice and text conversations. It employs sophisticated
natural language processing to identify a large variety of human emotions from
user inputs. The system responds by creating customized healing
recommendations such as inspirational quotes, soothing music, and breathing
techniques based on the user's emotional state. It espouses user privacy through
local data storage without logins, facilitating hassle-free, accessible mental
health support.

The architecture includes a number of principal elements:

• The User Interface (UI) enables users to communicate through voice or

text.

• Voice inputs are translated into text through speech-to-text modules

such as Whisper or Google API.

• Text is preprocessed with NLP libraries such as spaCy and NLTK.

• Emotion detection models such as BERT and RoBERTa categorize
emotional states from preprocessed text.
• The response generator generates personalized responses on the
basis of identified emotions.

• Mood tracking stores user emotional patterns locally for persistent

assistance.

• Privacy and data protection are provided through local encrypted

storage and zero cloud data holding.

• This multi-layered design provides seamless, accurate emotion

recognition and assistive interaction with robust privacy in mind.

Figure 1. Convolutional Neural Network

 3.1.2. Activation and Regularization

Activation in the MindMate AI framework is facilitated through nonlinear

activation functions integrated into its deep learning models to enable encoding
of intricate emotional patterns. In the transformer-based emotion detection
layers (e.g., RoBERTa and GPT-4 Mini), Gaussian
Error Linear Units (GELU) are used after every self-attention and feed-forward
sublayer, allowing for smooth, probabilistic activations that enhance gradient
flow and model expressiveness. Within the response generation modules—
especially in fully connected layers projecting emotion embeddings to the
response categories—Rectified Linear Units (ReLU) are employed to impose
sparsity and mitigate vanishing-gradient effects, promoting rapid convergence
during training. To avoid overfitting and have the chatbot generalize well
across varied user inputs, the architecture utilizes a variety of regularization
techniques.

Dropout at rate 0.1 is used after every attention block and between dense
layers to randomly shut off neurons at training time, promoting the model to
learn redundant and robust representations of emotional signals. Layer
normalization is used throughout transformer blocks to regularize training by
normalizing intermediate activations, and also as a weak form of
regularization. Additionally, L2 weight decay is applied to all trainable
parameters with a small coefficient (e.g., 0.01) to penalize large weights and
further improve generalization. Together, these activation and regularization
techniques ensure that the MindMate AI’s emotion detection and response
generation modules remain both flexible enough to model nuanced human
emotions and robust against overfitting to specific training data.

• Data Preprocessing and Dataset

Normalization in MindMate AI plays a crucial role in preparing user input for

accurate emotion detection and response generation. Once speech is
transcribed into text, the normalization process standardizes the raw text by
converting all characters to lowercase, removing punctuation and extraneous
symbols, and expanding common contractions (e.g., “I’m” to “I am”). Stop
words like "the," "and," or "but" are removed to concentrate on emotionally
relevant words, and lemmatization lowers words to their root forms—e.g.,
"running" to "run" —so that various word forms are treated equally. This
cleansing and normalizing step also fixes standard misspellings and normalizes
emotive abbreviations (e.g., "lol" to "laugh out loud") so that downstream NLP
models can properly understand user emotion. By providing a reliable, noise-
free text input to the tokenization and parsing modules, normalization aids the
system in perceiving subtle emotional states and producing empathetic,
context-sensitive suggestions.

3..1 Dataset Collection

MindMate AI dataset is an abundant, ethically composed corpus of

anonymized user conversations intended for training and testing the system's
emotional intelligence features. It consists of tens of thousands of instances
collected from various demographic segments such as students dealing with
academic pressure, working professionals dealing with burnout, and young
adults dealing with social pressures—each recorded in actual everyday
conversational sessions. Each interaction registers the initial raw user input,
either voice transcription or text prompt, and then the tokenized, normalized
input that is fed into the NLP pipeline. The platform labels each interaction
with one or more emotional states—ranging from simple states like happiness,
sadness, anger, and fear to more advanced mental-health–oriented conditions
like overwhelm, hopelessness, burnout, and suicidal ideation—and also
includes confidence scores measuring model confidence. To facilitate
personalized response generation, the dataset also contains the particular
intervention selected (e.g., motivational quotes, guided breathing exercises,
relaxing audio suggestions, or CBT-based prompts) and asks for optional user
ratings in terms of perceived helpfulness of each suggestion. Metadata fields
record the date and time, input modality (voice vs. text), session ID, and
anonymized user profile features (e.g., age group, overall stress level) to
facilitate longitudinal trend analysis. Historical interaction sequences for each
user are stored to support mood-tracking analytics and adaptive
personalization, enabling the AI to learn from aggregate emotional patterns
over time. All entries are stored locally in encrypted SQLite or IndexedDB
formats to respect privacy, with consented, aggregate exports used only for
model retraining, performance benchmarking, and ongoing refinement of
response strategies. This comprehensive dataset underpins MindMate AI’s
ability to detect nuanced emotional cues, deliver empathetic support, and
evolve through real-world usage.

The dataset includes more than 50,000 anonymized interaction logs gathered
from a diverse population of users—students, working individuals, and young
adults—who are interacting with the system both vocally and textually. About
60% of records come from short voice clips (10–30 seconds long each)
transcribed using Whisper or Google Speech-to-Text, while the other 40% are
typed text inputs averaging between 15–50 words. Each entry contains the raw
input as originally submitted, a normalized and tokenized copy of the text, and
up to three labeled emotional states taken from a taxonomy of 40 categories
(from simple emotions such as joy, sadness, and anger through subtle mental-
health states like burnout, overwhelm, hopelessness, and suicidal ideation).
With each label comes a confidence measure of model certainty. For every
interaction, the dataset records the system's selected intervention—e.g.,
motivational quotes (20% of interventions), guided breathing exercises (25%),
soothing music links (15%), CBT-based prompts (30%), and crisis resources
(10%)—and optional binary feedback on helpfulness. Metadata fields include
exact timestamps, session IDs to group sequence interactions (average session
length: 5 entries), input modality, and coarse user- profile attributes (age
bracket, self-reported stress level). All data are securely stored in encrypted
SQLite tables on user devices, with aggregated summaries exported for model
retraining and performance analysis. Model Training And Fine Tuning.
MindMate AI is an innovative AI Mental Health Buddy developed to address
the growing prevalence of stress, anxiety, and depression, especially among
students and working professionals who often avoid traditional support due to
social stigma or inaccessibility. Functioning as a virtual companion, the system
provides real-time emotional support through an interactive, web-based
application. Its core functionality relies on AI-driven mood detection using a
sophisticated Natural Language Processing (NLP) pipeline. The system
captures user input through voice or text interaction, converts speech to text
(using technologies like Google Speech API or Whisper AI), and processes the
text using techniques like Tokenization, Lemmatization, and Stopword
removal (via spaCy/NLTK). The emotion is then classified using advanced
models like DistilRoBERTa or fine-tuned deep learning models, which are
trained to recognize an expanded set of emotional states far beyond basic
sentiment, including nuanced categories such as Mental Health-Focused States
(e.g., Burnout, Hopeless, Suicidal thoughts) and Recovery & Growth Emotions
(e.g., Resilience, Healing). Based on the detected mood, the system generates
personalized healing suggestions, which include motivational quotes, calming
music, relaxation tips (like CBT-based responses), and breathing exercises.
Furthermore, the system incorporates a Mood Tracker Module with visual
charts to help users track and understand their emotional patterns over time. A
primary advantage is the focus on user privacy, as the system is designed to be
accessible anytime, work without mandatory login, and store mood data
locally. The ultimate goal is to bridge the accessibility gap and empower users
to proactively manage and improve their emotional well-being.

3..2 Technologies and Tools used

Frontend Technologies

HTML (HyperText Markup Language) serves as the foundation for the web
interface, providing the structural markup for all pages. The frontend utilizes
clean and semantic HTML5 elements to create an accessible user interface
that supports both text and voice interactions. CSS (Cascading Style Sheets)
handles all visual styling and responsive design aspects. The implementation
focuses on creating a user-friendly, calming interface appropriate for mental
health applications, with careful attention to color psychology and
accessibility standards. JAVASCRIPT powers the interactive elements and
dynamic functionality. The frontend JavaScript handles user interactions,
manages the voice-to-text conversion using the Web Speech API, and
orchestrates communication between the user interface and backend services.

Natural Language Processing (NLP) and Machine Learning

NLTK (Natural Language Toolkit)

• It performs Tokenization, which is breaking text into individual words and
sentences.
• It handles Stop Word Removal, filtering out common words that don't
help in emotional analysis.
• In Stemming and Lemmatization, which reduces words to their base
forms.
• NLTK also does Part-of-Speech (POS) Tagging, identifying the
grammatical roles of words.
 • spaCy provides advanced NLP capabilities and offers superior
performance.
• Fast Processing due to its Cython-optimized implementation,
making it suitable for production use.
• Dependency Parsing, which helps in understanding the
grammatical relationships between words.
• Named Entity Recognition, identifying important entities within the
user's input.
• Its Lemmatization is considered more accurate for word
normalization compared to simple stemming.

Machine Learning Frameworks for Emotion Detection

Transformers (Hugging Face) library enables access to state-of-the-art pre-

trained models:

• Scikit-learn: For traditional machine learning approaches and model

evaluationBERT (Bidirectional Encoder Representations from
Transformers): For understanding context in user messages
• RoBERTa: Enhanced version of BERT with improved
accuracy for emotion classification
• DistilBERT: Lightweight version of BERT for faster inference
while maintaining accuracy
• TensorFlow serve as the underlying deep learning frameworks:
• PyTorch: Preferred for research and experimentation with dynamic
computation graphs
• TensorFlow: Suitable for production deployment with robust serving
infrastructure

Voice-to-Text Conversion

• The system supports multiple speech-to-text options:

• Google Speech-to-Text API: Cloud-based solution with high
accuracy and real-time processing.
• OpenAI Whisper: Open-source, offline-capable solution with
strong multilingual support.
 Database Technologies:
Local Storage Options
• For development and privacy-focused deployment, SQLite
provides an efficient embedded database solution
• Serverless: No separate database server required
• Lightweight: Single-file database perfect for local mood data storage
• Zero Configuration: Easy to set up and maintain

Monitoring and Analytics

In Application Monitoring, The system incorporates several observability

features, including Error Tracking for comprehensive error logging and
notification systems, Performance Monitoring to gather real-time application
performance metrics, Health Checks to ensure automated system health
monitoring, and Usage Analytics to understand user interaction patterns while
strictly maintaining privacy.

ML Model Monitoring

The system incorporates advanced model management features, including

Model Performance tracking to monitor the emotion detection accuracy over
time, A/B Testing to compare different model versions and various approaches
for classification, Bias Detection for monitoring potential biases that might
arise in emotion classification, and Continuous Learning mechanisms that
allow the model to improve constantly based on ongoing user feedback.
Sources
 CHAPTER 4

MODULES
4.1 Data Collection and Preprocessing Module

The Data Preprocessing and Preprocessing module is a crucial first step in the
system's Natural Language Processing (NLP) pipeline. Its sole purpose is to
clean the raw user input, whether it be transcribed voice or text, to allow for
proper analysis by the emotion classification model. This is done through the
use of tools such as NLTK and spaCy. Preprocessing steps of key importance
are Tokenization —splitting the text up into smaller components such as words
or sentences —and Lowercasing for uniformity. It also does Stop Word
Removal so that frequent words that do not aid emotional analysis can be
eliminated, and Lemmatization (and even Stemming) to bring words down to
their roots. The module can further use the transformers library (HuggingFace)
for expert BERT tokenization and preprocessing. Following cleaning and
normalization of the text, the Text Vectorization step transforms it into
numerical form employing methods such as TF-IDF so the data is now ready
to feed into the machine learning model

4.2 NLP Model Training Module

The NLP Model Training Module is responsible for developing and refining
the core intelligence of the system. This process involves using a specialized
dataset, specifically the HuggingFace Emotion dataset, which contains
emotion labels necessary for training. The input text, after undergoing
preprocessing (Tokenization, Lemmatization), is converted into a numeric
format using Text Vectorization techniques like TF-IDF. This vectorized data
is then fed to the Emotion Classification model, which can be Traditional
Machine Learning algorithms such as SVM, Random Forest, or Naive Bayes,
or modern deep learning models like BERT, LSTM, a fine-tuned version of
DistilRoBERTa. The goal is for the model to accurately classify emotions in
the user's text and fetch the top-scoring emotion to trigger an appropriate
empathetic response or suggestion from the chatbot.
 4.3 Chat Response And Prediction Module

Chat Response and Prediction Module serves as the system's final output
engine, utilizing the emotion detected from the user's input to generate a
meaningful and helpful reply. After the Emotion Classification stage
identifies the top-scoring emotion (e.g., "lonely") , this module triggers the
appropriate empathetic response, emoji, or healing suggestion. The system
uses Rule-based or Retrieval-based methods, often employing cosine
similarity, to select the best response from a pre-defined list. This
personalized content includes motivational quotes, calming music,
breathing exercises, and relaxation tips, such as CBT-based calming
responses. Essentially, this module takes the emotional prediction and
translates it into real-time, personalized support and recommended healing
content.

4.4 Sentiment And Emotion Analysis Module

The Sentiment and Emotion Analysis Module is the core intelligence of the
system, responsible for detecting the user's emotional state from their input.
This process begins by converting any spoken input into text using either the
Google Speech-to-Text API or Whisper AI. The resulting text then undergoes
Text Preprocessing via tools like spaCy or NLTK (performing Tokenization,
Lemmatization, etc.). The preprocessed text is then passed to the Emotion
Classification Model, which can be a fine-tuned advanced model such as
DistilRoBERTa , or more traditional models like BERT, LSTM, or Traditional
ML (SVM, Random Forest). The model's job is to analyze the text, classify the
emotion, and fetch the top- scoring emotion (e.g., "lonely"). This detected
emotion is finally used to trigger the personalized and empathetic response or
suggestion from the chatbot

4.5 User Interface Module

The User Interface (UI) Module provides a clean and interactive platform,
serving as the user's primary point of interaction with the AI Mental Health
Buddy. Designed for maximum accessibility, the UI supports flexible input via
both Voice and Text, leveraging the Web Speech API to capture and transcribe
spoken communication in real-time. Beyond conversational support, the UI
integrates a Mood Tracker, which visually displays charts and logs to help
users reflect on their emotional journey and patterns over time. Furthermore,
the system prioritizes user privacy by functioning without mandatory login or
signup and stores mood data locally, ensuring a completely secure and
personalized experience.

4.6 Model Integration and Deployment

The Model Integration and Deployment Module is responsible for

operationalizing the trained Emotion Classification Model and ensuring its
robust performance in a live environment. The finalized NLP model,
potentially a fine-tuned version of DistilRoBERTa, is integrated into the back-
end application. Deployment utilizes Platforms-as-a-Service (Heroku,
Railway, or Render) for simplified, scalable hosting, or the comprehensive
cloud infrastructure of AWS (Amazon Web Services), leveraging services
like EC2 or AWS Lambda. Crucially, this module incorporates advanced
monitoring, including Model Performance tracking for emotion detection
accuracy, Testing to evaluate different model versions, Bias Detection, and
Continuous Learning mechanisms for iterative improvement based on real-
time user feedback. The front-end can be hosted separately on platforms like
Vercel or Netlify, which provide features like serverless architecture and
automatic scaling
 CHAPTER 5

SYSTEM DESIGN

5.1 DATA FLOW DIAGRAM

5.2 MODEL ARCHITECTURE

5.3 API Workflow

 CHAPTER 6

6. MATHEMATICAL CONCEPT USED

The AI Mental Health Buddy is based on a number of basic mathematical

and statistical principles that form the basis of its Natural Language
Processing (NLP) and machine learning functions. The principles are mainly
applied to allow the model to process and measure emotions based on text
data.

Text Vectorization: The model employs methods such as TF-IDF (Term

Frequency-Inverse Document Frequency) and Word2Vec (or comparable
word embedding processes) to translate human language into numerical
vectors.

TF-IDF is a quantitative measure that indicates how much a word

contributes to a document relative to a group of documents. Word2Vec
employs a neural network to map words into dense vectors in a high-
dimensional vector space where the distance between vectors
mathematically encodes for semantic [Link] Similarity is a
measure to determine how close two non-zero vectors are by measuring the
cosine of the angle between them. Here, it is employed to match the
vectorized user input (emotion) against the vectorized pre-defined responses
to select the closest match.

Machine Learning Models: The statistical and probabilistic foundations

form the basis of the core Emotion Classification Model. ML Models such
as Support Vector Machines (SVM), Random Forest, and Naive Bayes (NB)
are constructed using statistical concepts and probability to label the
vectorized text into one of the 35+ emotions. Deep Learning Models such as
BERT, LSTM, and DistilRoBERTa employ intricate linear algebra,
calculus, and matrix operations in their neural network architecture to learn
very non-linear patterns in the text data. Probability & Scoring: Probability
scores are utilized by the system to retrieve the highest-scoring emotion from
the model output (e.g., {sad: 0.82, angry: 0.21, lonely: 0.88} → "lonely").
 CHAPTER 7

7. SYSTEM TESTING

A comprehensive system testing plan is vital to ensure the AI Mental

Health Buddy is accurate, reliable, and provides an effective user
experience. Testing is structured across several specialized phases, including
Unit, Integration, Performance, and Usability Testing, to validate both
functional accuracy and non-functional requirements.

Unit Testing

Unit testing focuses on validating the smallest individual components, or

"units," of the application in isolation to ensure each piece of code is working
correctly. In this system, key units include the functions for Text
Preprocessing and Emotion Classification. For preprocessing, this involves
testing that the Tokenization function correctly splits text into tokens, that
Stop Word Removal accurately filters out common words, and that
Lemmatization reduces words to their proper base forms. For the Emotion
Classification Model, unit tests would verify the mathematical core of the
model: given a known vectorized input, the model must consistently output
the expected probability scores, with the correct emotion (e.g., "lonely")
being the top-scoring result. Finally, the functions responsible for selecting
the correct Healing Suggestion (e.g., a CBT tip) based on an input emotion
must be verified to ensure the mapping logic is flawless.

Integration Testing

Integration testing verifies that different modules of the system work

together smoothly as a whole. This is crucial for the AI pipeline, which relies
on a sequential flow of data. The primary test scenario is the full Sentiment
Analysis Workflow: a user's voice input is passed to the Voice Recognition
Module (using Web Speech API or Whisper AI), which successfully outputs
transcribed text. This text is then passed to the Text Preprocessing functions
(using NLTK/spaCy), and the cleaned text is passed to the Emotion
Classification Model. The final, successful integration is proven when the
fetched emotion triggers the correct, personalized response from the
Response Generator Module. Integration testing also covers the interaction
between the Front-end UI and the Local Storage to confirm that mood data
is saved and correctly retrieved to populate the Mood History & Analytics
charts.

Performance Testing

Performance testing evaluates the system's responsiveness, speed, and

stability under various load conditions to ensure a good user experience. The
key metric is End-to-End Latency: the time taken from a user providing
voice or text input until the system displays
the final therapeutic response. The system must process the voice
conversion, NLP pipeline, and response generation in a near-real-time
manner to maintain a natural conversational flow. Load testing will simulate
multiple concurrent users interacting with the chatbot to ensure the Cloud
Deployment Platforms (like Heroku or AWS) can handle the load and scale
effectively without service degradation.

Usability Testing

Usability testing focuses on the user experience, ensuring the interface is

intuitive and the system provides effective, accessible support. Testing
confirms that the User Interface (UI) is clean, interactive, and supports both
Voice and Text Interaction equally well for better accessibility. Users will
test features like the ease of providing voice input and the clarity of the
Visual Mood Logs. Since the system emphasizes User Privacy, testing
verifies that it functions effectively without requiring login or signup and
that the local storage of mood data is seamless and non-intrusive. A final
check ensures the therapeutic content is easy to find and understand,
fulfilling the goal of providing personalized healing content.

Real Time Accuracy Evaluation

The system's accuracy in real time is persistently tested by means of a two-

pronged process: Model Performance Tracking and Response Relevance
checking. Model performance mainly monitors Accuracy Over Time and
examines Confidence Scores of the Emotion Classification Model in order to
promptly recognize degradation or weakness in recognizing particular
emotions. This is accompanied by Bias Detection to maintain classification
equity. The system also employs techniques such as A/B Testing to contrast
various model variants or answers in a live setting, leveraging user
engagement as an indirect proxy for therapeutic efficacy. Lastly, the process
is backed by Continuous Learning mechanisms, which assist in iteratively
learning and improving the model and its response logic based on the repeated
feedback loops that occur while users interact.
 CHAPTER 8
RESULTS AND CONCLUSION

The final deliverable is an AI Mental Health Buddy, a web-based application

designed to be a virtual companion offering emotional support. It is primarily
aimed at students and working professionals facing mental health challenges
like stress, anxiety, and depression, who often hesitate to seek traditional help
due to social stigma or access issues. The application allows interaction via
voice or text and uses Natural Language Processing (NLP) to detect the user's
emotional state in real-time. Based on the detected mood, the system provides
personalized healing suggestions, including motivational quotes, calming
music, breathing exercises, and relaxation tips. Key advantages include its
focus on user privacy (no login/signup, local data storage) and its utility as a
Mood Tracker that visually charts a user's emotional patterns over time. The
project, developed by a team from HICET, aims to bridge the gap in mental
health support by offering an accessible, friendly, and private platform.

Conclusion:
The AI Mental Health Buddy project successfully developed a web-based
application designed to act as a virtual companion for emotional support,
primarily targeting students and working professionals facing rising levels of
stress and anxiety. Addressing the drawbacks of traditional support systems—
such as social stigma, lack of access, and cost —the proposed system utilizes
AI-driven mood detection via both voice and text input.

The core strength of the system lies in its Sentiment and Emotion Analysis
Module, which employs advanced Natural Language Processing (NLP)
techniques, including the Google Speech API or Whisper AI for voice
conversion, and deep learning models like DistilRoBERTa or ERNIE 4.0 for
Emotion Classification. This allows the system to not only recognize 35
existing emotions but also an expanded range of Mental Health-Focused States
(like Burnout, Hopeless, Panic), Recovery & Growth Emotions, and
Complex/Mixed Emotional States. Based on the detected emotion, the
Response Generator provides real-time personalized healing suggestions,
including motivational quotes, calming music, and CBT- based calming
responses. The AI Mental Health Buddy is a user-friendly, accessible, and
privacy-focused solution, working without a login and storing mood data
locally. By offering personalized emotional insights and visualization through
its Mood Tracking & Analytics module, the system fulfills its goal of
empowering users to understand and proactively improve their emotional well-
being.
10 FUTURE SCOPE

o Deeper Emotion Classification: The system can be enhanced to

accurately recognize and respond to the Complex / Mixed Emotional
States (like "Bittersweet" or "Shame about feeling sad") and the detailed
Recovery & Growth Emotions (such as "Resilience" or "Forgiveness").
o Advanced Parsing Techniques: Implementing more sophisticated
parsing techniques like Constituency Parsing and Semantic Parsing can
allow the chatbot to better understand complex sentence structures and
map user queries to formal therapeutic actions (like API calls for
resources).
o Multilingual Support: While existing systems like Leora and a
ChatGPT-4.0-based chatbot show multilingual capabilities, the AI
Buddy can be expanded to support regional languages and dialects,
particularly important for a diverse user base like students and
professionals.
o Context and Memory: Developing deeper context awareness to
remember past conversations and emotional trends would allow for more
relevant and "human-like" conversation flow, addressing a current
drawback of existing systems.
o Proactive Crisis Handling: While the system currently flags inputs like
"Suicidal thoughts," the scope can be expanded to integrate a robust,
automated crisis alert handling procedure that connects users
immediately to verified, local emergency resources or crisis hotlines.
o Integrated CBT Modules: Expanding the Response Generator module
from simple CBT- based calming responses to full, interactive Cognitive
Behavioral Therapy (CBT) modules and guided exercises, similar to
features seen in apps like Woebot and Kaya.
o Interactive Exercises: Integrating mood-specific exercises and activities,
like guided meditation or mindfulness practice (similar to MindBot's
emotion-specific exercises), directly into the chat interface.
o Advanced Mood Tracking & Analytics: Moving beyond simple visual
trend charts, the Mood History & Analytics module can be enhanced to
provide deeper personalized insights and actionable suggestions based
 on long-term emotional patterns.
o Remote Monitoring/Sharing (Optional): Developing a secure, opt-in
feature that allows users to share their emotional data trends with a
trusted therapist or family member (with explicit user consent) for better
care coordination.
o Enhanced Data Privacy and Security: Continuously auditing the Local
Storage mechanisms to ensure that user mood data remains absolutely
private and secure, reinforcing the system's core advantage of being User
Privacy Focused and working without mandatory login/signup.

REFERENCES

o Kang, Hong. (2025). Development and Evaluation of a Mental Health

Chatbot Using ChatGPT-4.0: Mixed Methods User Experience Study
with Korean Users. JMIR Medical Informatics.
o This paper discusses the application of GPT-4 in mental health chatbots,
covering multilingual capabilities, user experience, and safety features.
o Sehgal, et al. (2025). Exploring Sociocultural Challenges and
Opportunities in Designing Mental Health Chatbots for Adolescents in
India. arXiv preprint.
o Focuses on cultural adaptation, multilingual support, and tailored
interaction for youth mental health.
o Woebot Health USA. (2023). Updates on Woebot Emotion AI V2,
including anxiety and stress management through CBT-based responses
and conversational AI.
o Highlights technology and behavioral therapy integration in chatbot
platforms.
o Luka Inc. USA. (2023). Replika Emotion AI V2: Advances in
conversational avatars for companionship and emotional support.
o Describes friendship simulation, AR modes, and emotion detection
techniques.
o Chopra et al. (2024). Voice-to-text accuracy comparisons of Whisper AI
and Google Speech-to-Text API. Journal of Speech Technology.
o An evaluation of offline vs cloud-based voice recognition accuracy
critical for privacy- focused mental health apps.
o Devlin, J., et al. (2019, ongoing updates). BERT: Pre-training of Deep
Bidirectional Transformers for Language Understanding. (Adapted
 usage through Hugging Face Transformers).
o foundational model for transformer-based emotion classification.
o Liu, Y., et al. (2019). RoBERTa: A Robustly Optimized BERT Pretraining
Approach.
▪ Extended transformer model enhancing NLP emotion
detection accuracy.
o Radford, A., et al. (2022). OpenAI Whisper: Multilingual speech-to-text
model.
o Open-sourcetechnology for voice-to-text used in
privacy-sensitive AI
applications.
o Singh, A., et al. (2025). Natural Language Processing applications in
AI mental health systems. AI in Healthcare Journal.
o Overview of tokenization, dependency parsing, semantic parsing
relevant to chatbot understanding.
o Sharma, P., et al. (2025). Emotion detection using DistilRoBERTa and
ERNIE models for conversational AI. NLP and AI Journal.
o Discusses optimizations for lightweight emotion classification
suited for real- time chatbot application.
International Journal of Scientific Research and Engineering Development-– Volume 8 Issue 5, Sep-Oct 2025
Available at [Link]
RESEARCH ARTICLE OPEN ACCESS

Mindmate Artificial Intelligence: Healing Through Conversion

Mr. [Link] M.E(Ph.D), Ranjith Kumar R, Tamilarasu M, Vimal S, Vineeth S

(Computer Science and Engineering, Hindusthan College of Engineering and Technology, Coimbatore,
[Link]@[Link])
************************
Abstract:
The prevalence of mental health issues like stress, anxiety, and depression is rising, especially among
professionals and students, but many people don't seek help because of stigma or lack of access. This project
introduces AI Mental Health Buddy, a web-based virtual companion that uses both voice and text interfaces
to provide personalized emotional support. Using cutting-edge natural language processing for real-time
mood detection, the system provides customized interventions, such as guided breathing exercises, calming
music, motivational quotes, and mood tracking, to help users develop emotional resilience. With an emphasis
on privacy and usability, AI Mental Health Buddy does not require sign-in and stores data locally to maintain
confidentiality. Through interactive, AI-driven engagement, the platform enables users to monitor and
improve their mental health at any time and from anyplace.

************************
suit the user's emotional state, including
I. INTRODUCTION motivational quotes, cool music, guided breathing
Mental welfare has become equally exercises, and comprehensive mood tracking
important as physical health in today's fast-transport features. In the form of existing mental health
lifestyle, yet using mental health aid is challenging applications, which may lack genuine sympathy, or
for many individuals. The increasing prevalence of limited privatization, AI mental health friendship,
stress, anxiety and depression, especially among when it can be compulsory registration and useful.
students and working professionals, has created an The system is designed to be a user friendly,
immediate requirement of accessible emotional accessible and particularly beneficial.
support systems. Traditional mental health services
often present obstacles including high cost, limited II. EASY TO USE
access, and social stigma that prevents individuals Mental health is an important part of our life
from demanding their help. The emergence of AI and just like physical health. Many people today suffer
natural language processing technologies provides from stress, anxiety and depression, especially
innovative opportunity to bridge this difference in students and working professionals. However, due to
mental health care distribution. Technology- social stigma, high cost, or lack of access to doctors,
operated solutions can provide immediate, many are hesitant to seek appropriate help. To solve
anonymous and personal assistance to users who this problem, technology and artificial intelligence
experience emotional crisis, providing mental health (AI) can be used to create a simple and auxiliary
aid anytime and anywhere. This web-based system. Our project, called AI Mental Health Buddy,
application takes advantage of the abilities to detect acts like a virtual partner who can listen to people
advanced AI-operated mood using natural language through text or voice, understand their feelings, and
processing to understand the user's feelings from provide comfort. The system uses natural language
both voice and text input. The system provides real - processing (NLP) to detect feelings such as
time, individual treatment suggestions suggested to happiness, sadness, loneliness, or stress. Once the

ISSN : 2581-7175 ©IJSRED: All Rights are Reserved Page 946

International Journal of Scientific Research and Engineering Development-– Volume 8 Issue 5, Sep-Oct 2025
Available at [Link]
mood is detected, it suggests useful materials such as methods user experience evaluation with built-in
motivational quotes, cool music, or breathing practice guardrails and safety features to handle more
to make the user feel better. It can also track the demanding emotional states. Sehgal et al. (2025)
emotional changes of the user over time using charts considered socio-cultural considerations in the
and graphs, which can help them understand their design of chatbots for Indian adolescents,
mental health patterns. Using without the need for emphasizing the need for regional language support,
chatbot login is private and always accessible, which culturally sensitive response content, and culturally
ensures complete privacy for users. Like existing based coping strategies.
systems, which lack sympathy and privatization, are
designed to make the proposed system a user friendly There are numerous open-source research projects
and attractive. It combines multiple modules, exemplifying the different features and architectures:
including voice identity, emotion detection, response 1. MindBot (IIIT-Hyderabad, July 2025): Engages
production and mood history. Together, these users via voice chat and features exercises or
modules make chatbot interactive and accessories for activities focused on particular emotions with
users facing emotional challenges. Although it is not detection for feelings, such as FOMO, empowerment,
a replacement for professional therapy, AI mental relief, and self-doubt.
health friend serves as a safe and simple tool for daily
stress management, awareness and self-care. 2. Leora (IIT Delhi & NIMHANS, January 2025):
Functions as a multilingual interaction bot as well as
III LITERATURE REVIEW an "emotions radar" that can detect feelings of
Improvements in artificial intelligence and anxiety, overwhelm, burnout, hopelessness, and
natural language processing (NLP) have enabled the peace.
creation of chatbots that provide mental health
support. Several recent studies and prototypes 3. Kaya Beta (AIIMS Delhi, December 2024):
highlight the promise and drawbacks of virtual Leverages the recording of mood journaling,
companions driven by AI providing focused, timely, approach to dialects, as well as breathing module
and personalized emotional support. Early while demonstrating detection for feelings of anger,
investigations of AI mental health chatbots included confusion, insecurity, calm, or trust, etc.
rule-based systems and frameworks based on
cognitive-behavioral therapy (CBT). Woebot 4. Each of these systems has adopted NLP toolkits
Health’s Woebot (2023) offers a sensible solution such as the Natural Language Toolkit (NLTK) for
with CBT conversation scripts and journaling tokenization, stemming and part-of-speech tagging;
prompts to address anxiety and depression through spaCy for rapid lemmatization and dependency
24/7 text chat. Woebot is successful in engaging parsing, as well as using transformer models (i.e.
users, but its reliance on predefined patterns BERT, RoBERTa) for enhanced emotion
embedded in the chatbot limits the opportunity for classification. However, open-source research
more nuanced dialogues about complex emotional chatbots have experienced challenges in crisis
states. Replika Emotion AI V2 (2023) presented a detection, privacy for users, and long-term use to
friend-simulated avatar that offered a more engage users.
conversational experience, but lacked robust crisis 5. the open-source mental health chatbots explored
detection features. Academic prototypes have indicate the effectiveness of AI-mediated emotional
explored multilingual and culturally adaptive chatbot assistance but underscore the gaps in response to
designs. As an example, Kang and Hong (2025) adaptability, cultural localization, or user's data
presented a ChatGPT-4.0–based mental health privacy. In light of this, the design of the AI Mental
chatbot for Korean users, using mixed- Health Buddy is reflected in the gaps discussed
above and will draw on: emergent transformer-based

ISSN : 2581-7175 ©IJSRED: All Rights are Reserved Page 947

International Journal of Scientific Research and Engineering Development-– Volume 8 Issue 5, Sep-Oct 2025
Available at [Link]
emotion detection, voice/text interaction, culturally- Natural Language Processing (NLP) pipeline. The
inclusive content, and a privacy-focused Emotion Detection Engine will determine emotional
architectural response to the above challenges to states using one or more transformer-based models,
assist in a user-centered mental health experience. while the Response Generator will create customized
and personalized response interventions. Finally, the
Analytics Module tracks and aggregates mood data,
IV PROBLEM STATEMENT allowing the client side to visualize easy and intuitive
Nowadays, many individuals experience data visualization.
stress, anxiety, and depression, particularly college
students and the workforce. While mental health is 5.2 Core Functions
equally important as physical health, seeking help is • Real-time emotion detection: Once user input is
often not an easy path, especially with social stigma, received, the system will analyze linguistic cues and
economic cost, and limited access to doctors. There prosodic features to detect emotions, such as sadness,
are a few chatbots and online tools available but there anxiety, calm, or happiness.
is very little evidence that they work. They don't truly • Personalized Support Intervention Recs: Based on
empathize, don't act well in crisis situations, and user-detected mood, the system will select from a
often don't personalize or protect privacy. library of actions and/or interventions—motivational
quotes, playlists, breathing guided slides, prompted
Given this evidence, it's time to look for a better journaling, etc. to provide empathetic and supportive
answer, a simpler, private, and easy-to-use approach response.
to explore one's feelings with spontaneous helpful • Vocal and Text Interactive Chat: User can speak
support in real-time. An AI-based mental health into a microphone or type via keyboard and converse
assistant uses text or voice input to capture emotions naturally. Users will receive continuous transcription
and provide motivational quotes, soothing music, of their voice, and provide instantaneous textual
deep breathing exercises, and tracking of one's mood. responses.
This type of system provides comfort, private • Mood Chart Dashboard: Displays daily and weekly
protection, and support for emotional well-being. trends in emotions using simple line and bar charts.
User will be able to see patterns and if they are
V PROPOSED SYSTEM making progress.
AI Mental Health Buddy is designed as a smart,
web-based virtual companion that provides 5.3 System Requirements
emotional support in real-time interaction via voice • Front-End: Up-to-date web browser with
and text. The system proposes to use state-of-the-art permission to use the microphone.
natural language processing, speech recognition, and • Back-End: [Link] server that runs RESTful APIs
personalized response generation, collecting data to for Speech-to-Text, NLP inference (HuggingFace
create a user-centered platform to monitor and Transformers), and data analytics.
improve mental health well-being. • Data Store: Local Storage for mood entries ensuring
that no private information is sent to offsite storage.
5.1 system overview and architecture
The system is designed with a web client that 5.4 Benefits Compared To Existing Systems:
will be utilized with as the front-end, AI components AI Mental Health Buddy uses pretrained transformer
for the back-end, and a local data store for private models that were fine-tuned on datasets with
mood and emotion tracking. User input, provided as emotional use cases and behaviors, allowing it to
either speech or text, will be sent to the Speech to interact dynamically rather than through rules-based
Text module (for speech), or sent directly to the chatbots to nuanced responses from users. Its dual
modality (voice and text) allows multiple user

ISSN : 2581-7175 ©IJSRED: All Rights are Reserved Page 948

International Journal of Scientific Research and Engineering Development-– Volume 8 Issue 5, Sep-Oct 2025
Available at [Link]
preferences and accessibility options. The design's ❖ Conducting a literature review related to AI
focus on privacy, including not requiring registration, chatbots and emotion detection will help with
prevents barriers to use and reduces stigma, while identifying design specifications.
using Local Storage for mood data allows the user to
control their information. 2. Architectural Design:
❖ Front-end web client: A [Link] application will
5.5 Workflow Sequence provide voice and text inputs, present the
1. The user opens the web application and chooses conversation, and visualize moods.
to input via voice or text. ❖ Back-end services: A [Link] server will host
2. The Speech-to-Text module converts the audio (if RESTful APIs that handle speech-to-text
applicable) and sends the transcribed text to the NLP processes, connect to the emotion detection
pipeline. systems, generate responses, and interact with
3. The Emotional Detection Engine classifies the some analytics functionality.
user’s mood in milliseconds. ❖ Local data store: A browser Local Storage will
4. The Response Generator selects an intervention store mood entries and user settings.
from the repository and sends it to the user.
5. The mood entry is saved locally for the user to 6.2 Module Creation
visualize on the dashboard. The core functionalities are contained within a
6. The user can review the mood entries and dedicated module for each of the following
visualize their emotional trends to reflect on their capabilities:
progress.
1. Module for Speech-To-Text.
These functions are developed into the architecture Use the Web Speech API to transcribe speech in
to deliver empathetic, accessible and private mental real time. Preprocess audio data to reduce noise and
health services that meet users’ emotional needs and normalize audio levels.
their resources.
2. Module For Natural Language Processing (NLP).
VI METHODOLOGY Utilize spaCy for tokenizing and lemmatizing.
The methodology for AI Mental Health Buddy Use a fine-tuned version of RoBERTa that has been
consists of an overall system design phase, trained on labeled emotional data to classify emotion
development phase, and evaluation phase for robust and sentiment.
functionality, a user-centered design, and to provide
empirical evidence on the effectiveness of the 3. Module For Response Generation.
chatbot. ❖ A rule-based selector will match detected
emotions to appropriate interventions.
6.1 System Design Phase: ❖ Retrieval of content from the library of
The design phase focuses on specifying the overall interventions (quotes, music playlists, breathing
system architecture and component interactions: scripts).
❖ Format the responses to be used in the text-to-
1. Requirements Gathering speech synthesis or simply displayed on screen.
❖ Stakeholder interviews with mental health
professionals and potential users will identify 4. Module For Mood Tracking And Analytics
notable features and highlight privacy Use Local Storage to log mood labels with
considerations. timestamps. Play audio-data and calculate daily
aggregated data. Play average weekly, weekly, and

ISSN : 2581-7175 ©IJSRED: All Rights are Reserved Page 949

International Journal of Scientific Research and Engineering Development-– Volume 8 Issue 5, Sep-Oct 2025
Available at [Link]
same day of week analytics, and also play audio-data. ❖ Carry out a security review to check that no
Use client-side charts with [Link]. personally identifiable information moves off the
user’s device.
❖ Review Local Storage encryption routines for
6.3 Implementation mood logs.
1. Development Environment
❖ Source Control: GitHub repository along with 6.5 Iterative Refinement
❖ CI/CD pipeline. Containerization: Docker Feedback and the data from each stage provide
images for the back-end services. fodder for use in subsequent development sprints so
that conversation quality, relevance of responses,
2. API Integration and user satisfaction can be iteratively refined.
Expose API endpoints on transcribe, analyze,
respond, and analytics. Implement rate limiting and VII. NATURAL LANGUAGE PROCESSING
input validation for reliability. METHODOLOGIES:

3. User Interface The AI Mental Health Buddy utilizes several well-

Responsive layout that works on desktop and established Natural Language Processing (NLP)
mobile browsers. Accessible design with a high- methods to preprocess user text, identify linguistic
contrast theme and keyboard navigation. characteristics, and accurately classify emotional
content. The present methodology consists of the
following:
4.4 Evaluation Strategy
i. Tokenization
1. Performance Testing The process of splitting n raw text into
❖ Assess speech-to-text latency and transcription meaningful segments (tokens) consisting of words or
accuracy in noisy conditions. subwords. This first process assists with subsequent
❖ Assess emotion classification accuracy on a processing by separating out the smallest semantic
held-out dataset. units of the sentence.

2. Usability Testing ii. Stop-Word Removal

❖ Engage 20 students and professionals in a The process of removing common words (e.g.
usability trial lasting two weeks. “the,” “and,” “is”) from the text that convey minimal
❖ Collect quantitative metrics (task time, error semantic value, in order to decrease noise and focus
rates) and qualitative feedback (System Usability on identifying emotionally relevant terms.
Scale).
iii. Stemming/Lemmatization
3. User Experience Study ❖ Stemming pertains to the reduction of the word
❖ Conduct pre- and post-study surveys on their to the root form (truncating the suffixes of the
perceived level of emotional support and comfort word) (e.g. “ happy, ” “ happiness, ”
in sharing their feelings. and“ happily”→ “ happy” ).
❖ Analyze changes in self-reported levels of stress ❖ Lemmatization describes the conversion of the
and anxiety. word to the canonical dictionary form,
contingent on the part-of-speech (e.g., “running”
4. Privacy Assurance
→“run” ). Lemmatization improves linguistic
consistency in analysis.

ISSN : 2581-7175 ©IJSRED: All Rights are Reserved Page 950

International Journal of Scientific Research and Engineering Development-– Volume 8 Issue 5, Sep-Oct 2025
Available at [Link]
iv. Part-Of-Speech (POS) Tagging expressions and the intervention prompt for
The process of assigning grammatical contextually relevant healing suggestions (quotes or
classifications to each token (e.g. noun, verb, exercises, etc.).
adjective) to help with syntactic parsing and more
accurately identify adjectives and adverbs, which xi. Multilingualism
frequently contain emotional meaning. Integrate language detection and gather
transformer variants to apply multilingual processing
v. Dependency Parsing of user input, as appropriate to provide cultural
The process of analyzing grammatical sensitivity inclusivity.
relationships to establish the manner in which words
are connected, in order to better understand the Together, these nlp techniques form a robust
grammatical structure of a sentence (e.g. subject- pipeline—transforming raw user input into rich,
verb-object structure), and therefore capture structured representations—allowing ai mental
complex emotional verbal expressions. health buddy to deliver empathetic, context-aware
emotional support.
vi. Named Entity Recognition (NER)
The ability to identify authors, people, locations, VIII. ACCURACY COMPARISON
and other entities so that the chatbot can detect when
a user is speaking specifically about a person or
context in a way that could affect their emotional
state.

vii. Transformer-Based Word Embeddings

This entails the use of pretrained transformer
models (bert, roberta, etc.) to develop contextualized
word embeddings that capture fine-grained semantic
and syntactic information (e.g., polysemy or subtle
different meanings.)

viii. Fine-Tuned Emotion Classification

A roberta model is fine-tuned on labeled emotion IX. ACCURACY ANALYSIS AND
datasets to classify user input into discrete emotions PERFORMANCE METRICS
(i.e., sad, anxious, calm, happy) achieving
reasonable accuracy on real-world conversational
data.

ix. Sentiment Analysis

Classifying input with sentiment polarity (positive,
negative, neutral) as an additional signal to emotion
classification, understanding that sentiment can
provide additional insight into qualitative user
experience.

x. Text Vectorization And Similarity Matching

Converting text to numerical vectors and
calculating cosine similarity between the user’s

ISSN : 2581-7175 ©IJSRED: All Rights are Reserved Page 951

International Journal of Scientific Research and Engineering Development-– Volume 8 Issue 5, Sep-Oct 2025
Available at [Link]
simple charts helpful for identifying trends in their
mood. Specific reasons to appreciate the system were
the strong privacy measures, including the ability to
use the tracker without an account and the ability to
store the mood data on their own device. Overall,
these results suggest that a companion built around a
friendly, private, and artificial intelligence- based
interface can support people in making a significant
difference in daily mental health, although future
work could focus on improving the ability to identify
more complicated or mixed emotions.

XI. CONCLUSION
The AI Mental Health Buddy project proves that
an AI-powered chatbot can provide individuals with
emotional support in a private, accessible way. With
rapid speech-to-text transcription, accurate
transformer-based emotion detection, and
personalized interventions the AI Mental Health
Buddy is able to provide individuals with timely
comfort and guidance. Usability testing with the first
prototype of the Buddy, indicated that users did find
the Buddy's interface easy to use, and enjoyed the
option to use speech or text without signing in or
X. RESULTS AND DISCUSSION providing the Buddy any personal information. By
The assessment of AI Mental Health Buddy checking in with simple mood tracking and
concluded that the system is capable of quickly and personalized recommendations, users reported
accurately recognizing and responding to users’ measurable decreases in stress from pre- to post-
emotions. The speech-to-text capability produced intervention. Overall, AI Mental Health Buddy
spoken words into text with an average delay of just shows us that we can offer a way for individuals to
beyond half a second at 90% accuracy and the check in on their mental health each day
emotion classifier detected moods less than 200 meaningfully, while offering a private and stigma-
milliseconds after the sound was created, correctly free alternative.
identifying mood almost nine times out of ten. In
usability testing with 20 volunteers, the application XII. REFERENCES
rated high for ease of use, and all participants
completed the essential tasks of starting a chat, [Link]
Fitzpatrick, K.K., Darcy, A., and Vierhile , M. (2017). Providing cognitive
therapy to young adults with depression and anxiety symptoms
requesting a breathing exercise, and displaying mood through a fully automated conversational agent (Woebot): a randomized
trends in less than thirty seconds without errors. controlled trial. JMIR Mental Health, 4(2), e19.
During the two-week study, participants indicated 2. Replika, (2023). Replika Emotion AI V2. Luka, Inc.
experiencing lower levels of stress by reporting an 3. Kang, S., & Hong, J. (2025). Development and Evaluation of a Mental
average of moderate stress levels down to low stress Health Chatbot Using ChatGPT-4.0: Mixed Methods User Experience Study
levels. The majority of participants used the mood with Korean Users. JMIR Medical Informatics.
tracker consistently and remarked that they found the

ISSN : 2581-7175 ©IJSRED: All Rights are Reserved Page 952

International Journal of Scientific Research and Engineering Development-– Volume 8 Issue 5, Sep-Oct 2025
Available at [Link]
4. Sehgal, N., Kumar, A., & Sharma, P. (2025). Exploring Socio-Cultural
Challenges and Opportunities in Designing Mental Health Chatbots for
Adolescents in [Link].08562.

5. IIIT-Hyderabad. MindBot: A Voice-Enabled Emotion-Specific Mental

Health Chatbot. 2025.

6. IIT Delhi & NIMHANS. Leora: A Multilingual Mental Health Companion

with Emotion Radar 2025.

7. AIIMS Delhi. Kaya Beta: Mood Journaling and Guided Breathing Chatbot
2024.

8. Honnibal, M., & Montani, I. (2021). spaCy: Industrial-strength Natural

Language Processing in Python.

9. Burns, P. J. LatinCy: Synthetic Trained Pipelines for Latin NLP. 2023.

10. Eyre, H., et al. Launching into Clinical Space with medspaCy: A New
Clinical Text Processing Toolkit in Python.2021.

ISSN : 2581-7175 ©IJSRED: All Rights are Reserved Page 953