Ramaiah Institute of Technology

(Autonomous Institute, Affiliated to VTU)

Department of Artificial Intelligence and Data Science

REPORT ON
Project Based Learning

GuardianVoice: An AI-Powered Real-

Time Scam Call Detection and Protection
System for Senior Citizens

SUBMITTED BY

Student Name Student USN

Rudrapratap Patil 1MS22AD047
Sannakki Swaroop 1MS22AD049
Srinidhi N S 1MS23AD402
Tarun Kumar B S 1MS23AD404

As part of the Course

Natural Language Processing– 22AD71
Supervised By
Dr. Vaneeta M
Sept 2025 –Dec 2025
1
 Evaluation Sheet

Sl. USN Name Demonstration Presentation Total

No of Project & Report Marks
(10M) (10M)
(20)
1 1MS22AD047 Rudrapratap Patil
2 1MS22AD049
Sannakki Swaroop
3 1MS23AD402
Srinidhi N S
4 1MS23AD404
Tarun Kumar B S

Evaluated By
Name: Dr. Vaneeta M
Designation: Associate
Professor Signature:

2
Topic No. Table of Contents Page No.

1. Introduction 4
2. Problem Statement 4
3. Project Objectives 5
4. Real time Applications 5
5. Proposed Systems 6
6. Methodology 8
7. Implementation 15
8. Result and Analysis 18
9. Conclusion 25
10. References 26

1
Sl. No. List of Figures Page No.

1. System Workflow 8
2. System Architecture 10
3. Working Explanation 13
4. Incident Dashboard 18
5. Selected incident Snapshot 18
6. Controls Snapshot 20
7. Confusion Matrix 21
8. ROC Curve 22
9. System Execution Output 23

2
 ABSTRACT
GuardianVoice is an intelligent real-time call-scam detection system designed to protect
vulnerable individuals, especially senior citizens, from phone-based financial fraud. The
system continuously listens to live calls, converts speech to text using lightweight speech
recognition, and analyzes the content using a hybrid model combining rule-based detection and
machine-learning classification. It identifies scam cues such as OTP requests, urgency, threats,
financial demands, and reward-based traps. When a high-risk pattern is detected, the system
instantly warns the user through synthesized speech and optionally notifies a guardian.
GuardianVoice also logs each incident for later review through a dashboard. Built with Python,
Streamlit, and a deployable architecture, this project demonstrates an effective practical
approach to enhancing personal digital safety. The solution is modular, privacy-focused, and
designed to run on edge devices, making real-time scam prevention accessible, proactive, and
reliable.

3
 1. INTRODUCTION
Phone-based scams have dramatically increased in recent years, exploiting human emotions
such as urgency, fear, and trust. Fraudsters often trick individuals into revealing sensitive
information like bank details, OTPs, or personal identity data. Despite growing awareness,
many people particularly senior citizens struggle to identify these evolving social-engineering
techniques. This highlights the need for automated systems that can detect suspicious language
during live calls.
GuardianVoice addresses this challenge by providing an intelligent, autonomous assistant that
monitors ongoing conversations. Unlike conventional fraud-detection systems that work after
financial transactions occur, GuardianVoice prevents scams at the earliest stage the phone call.
The system uses speech recognition, rule-based detection, and machine-learning techniques to
differentiate between genuine and fraudulent communications in real time.
The system further enhances safety by integrating text-to-speech alerts, logging mechanisms,
and an optional guardian-notification system. Its modular architecture, with separate
components for audio capture, transcription, detection, scoring, and UI visualization, ensures
scalability. By combining artificial intelligence with practical usability, GuardianVoice offers
a powerful technological safeguard against social-engineering attacks.

2. PROBLEM STATEMENT
Financial fraud over phone calls continues to grow at an alarming rate, with scammers using
increasingly sophisticated techniques. Many individuals fail to recognize scam patterns in real
time, especially when conversations involve pressure tactics or impersonation of trusted
authorities. Without timely intervention, victims may unknowingly share confidential
information leading to financial loss.
Existing solutions such as spam-call filters, bank alerts, and awareness campaigns do not
effectively detect scam language during active calls. Spam filters block unknown numbers but
cannot analyze the content of conversations. Meanwhile, machine-learning solutions deployed
by organizations work after the scam has occurred, not during the crucial interaction.
Hence, there is a need for a proactive, real-time, user-side detection system that listens to call
audio, identifies scam cues instantly, and warns the user before any irreversible damage occurs.
The system must be accurate, lightweight, interpretable, and easy to deploy across devices such
as desktops and mobile phones.

[Link] Objectives
4
1. Detect scam-related language during live phone calls in real time.
2. Use speech-to-text conversion to analyze spoken content.
3. Identify scam patterns using rule-based and ML-based approaches.
4. Provide immediate audible warnings to the user through TTS.
5. Log all analyzed calls for later review.
6. Enable guardian notification for high-risk incidents.
7. Build a simple, user-friendly dashboard.
8. Ensure privacy by processing audio locally.

4. Real-Time Application
4.1 Protecting Senior Citizens from Scam Calls.

Senior citizens are increasingly becoming prime targets for phone-based scams due to factors
such as limited technical awareness, social isolation, and a higher tendency to trust unknown
callers. Scammers exploit these vulnerabilities through impersonation, emotional
manipulation, and urgency tactics to extract personal information, bank details, or direct
payments. As digital communication grows, fraudsters use advanced tools like AI-generated
voices, spoofed caller IDs, and automated robocalls, making scams even harder to detect
manually.

Traditional awareness programs and manual call-screening methods are often insufficient for
elderly individuals who may not understand modern fraud techniques. Many seniors live alone,
experience cognitive decline, or rely on memory-based judgment when answering calls. This
makes them more susceptible to threats like “Your bank account is blocked” or “Your
grandchild is in danger,” which trigger emotional decisions.

4.2 Preventing financial fraud in households.

Preventing financial fraud in households has become increasingly important as digital

communication and online banking continue to grow. Fraudsters frequently use deceptive
phone calls, impersonation of banks, fake government officials, and emotional manipulation to
trick individuals into sharing confidential information. These scams often appear convincing,
and families may find it difficult to differentiate real calls from fraudulent ones.

Households are particularly vulnerable because scammers exploit common weaknesses such
as lack of technical knowledge, panic-driven decision-making, and trust in official-sounding
callers. A single scam call requesting OTPs, bank details, or money transfers can lead to severe
financial damage, identity theft, and months of emotional stress. Traditional awareness
campaigns are not always enough, as fraud tactics continuously evolve.

To effectively prevent fraud, intelligent solutions such as AI-powered scam call detection
systems provide real-time protection. These systems analyze caller behavior, detect suspicious
keywords, and alert users during the call before any sensitive information is revealed. By

5
integrating such technology, households can significantly reduce their risk of being targeted
and ensure a safer financial environment for all family members.

4.3 Personal safety tools integrated into smartphones.

Personal safety tools integrated into smartphones have become essential in modern life, where
threats can arise unexpectedly from scam calls to physical emergencies. Today’s smartphones
include built-in features such as emergency SOS buttons, GPS tracking, call recording, and
fraud detection systems. These tools allow users to respond quickly to risky situations without
needing additional devices.

Such safety features are particularly valuable for vulnerable groups, including senior citizens,
children, and individuals living alone. Smartphones can automatically detect anomalies like
suspicious calls, sudden location changes, or high-stress voice patterns and alert trusted
contacts or emergency services. This significantly reduces response time during critical
moments and ensures that users are never completely isolated during emergencies.

With the integration of AI-powered safety technologies, smartphones are transitioning from
simple communication devices to comprehensive personal protection tools. Features like real-
time scam call analysis, voice-based threat detection, and automated alerts help users identify
dangers before harm occurs. As a result, smartphone-based safety systems have become a
reliable and accessible method for enhancing personal security in households and communities.

5. Proposed System
The proposed system introduces an AI-powered, real-time call monitoring and scam-detection
assistant designed to protect senior citizens from fraudulent phone calls. It listens to ongoing
calls, converts speech to text, analyses the content using rule-based detection and machine
learning, and alerts the user immediately if the conversation appears risky. It also provides
optional notifications to guardians for enhanced safety.

1. Real-Time Audio Capture

● Continuously records short audio segments during ongoing phone calls.

● Runs silently in the background without interrupting the conversation.
● Ensures constant monitoring without requiring user involvement.

2. Speech-to-Text Conversion

● Converts each recorded audio segment into text using speech recognition.
● Allows the system to understand context and detect harmful intent.
● Works even with noisy environments and mixed-language scenarios.

3. Rule-Based Scam Detection

● Uses predefined keywords and trigger patterns such as OTP, PIN, urgent payment,
account verification, or winning prizes.
● Identifies common scam techniques like fake bank calls, impersonation, financial
urgency, or lottery scams.

6
 ● Quickly detects known fraudulent behaviors.

4. Machine Learning Classifier

● A trained ML model evaluates the transcribed text for scam probability.

● Understands conversational tone, emotional pressure, and suspicious requests.
● Produces a scam-likelihood score for every conversation chunk.

5. Risk Assessment Engine

● Combines rule-based alerts with ML probability to form a unified risk score.

● Classifies calls into safe, suspicious, or high risk.
● Supports adaptive decision-making for user safety.

6. Real-Time Audio Warning

● Plays an immediate voice alert when a suspicious pattern is detected.

● Warns the user with messages like: “This call appears suspicious. Do not share OTP or
bank details.”
● Helps stop the user from giving away sensitive information under pressure.

7. Logging and Alert Generation

● Stores details such as timestamp, transcript, triggered flags, and risk score.
● Generates alert files for high-risk situations.
● Enables future integration with SMS or push notifications to guardians.

8. Guardian/Family Notification Support

● Sends alerts to trusted family members when risk exceeds a threshold.

● Allows guardians to monitor potential threats.
● Helps prevent fraud through timely intervention.

9. Simple Dashboard Interface

● Provides an easy-to-use dashboard for viewing alerts, transcripts, and risk analysis.
● Helps guardians track suspicious patterns and support senior citizens more effectively.

10. Lightweight and Device-Friendly

● Designed to work efficiently on standard smartphones and low-end systems.

● Optimized for low power consumption and smooth performance.
● Ensures reliable operation without affecting device usability.

[Link]
6.1 System Workflow

7
 Fig:1

The diagram shows how GuardianVoice processes each short piece of audio (a 5–6 second
chunk) and decides whether the call is suspicious or safe. It illustrates how different
components of the system communicate step-by-step.

1. Audio Capture (Mic → Guardian)

● The microphone records a small audio chunk.

● This chunk is immediately sent to the GuardianVoice system for processing.
● This recording–sending cycle repeats continuously during the call.

2. Speech-to-Text Conversion (Guardian → STT)

● The captured audio chunk is forwarded to the Speech-to-Text (STT) module.

● The STT module (Whisper/Tiny) listens to the audio and converts it into text.
● The converted text is then sent back to the Guardian.

This is the raw conversation text used for scam detection.

3. Parallel Risk Analysis (Guardian → Detector + Scorer)

8
The Guardian sends the text to two components at the same time:

(a) Rule-Based Detector

Checks for red-flag keywords such as:

● “OTP”
● “urgent”
● “account blocked”
● “transfer money”
● “lottery”
● threats or pressure

It returns the risk flags.

(b) ML Scorer

The ML model (TF-IDF + Logistic Regression) analyses the whole sentence.

It returns an ML scam probability (between 0 and 1).

4. Decision Making (Guardian)

The Guardian combines:

● Rule-based flags
● ML probability

A final risk score is computed.

This decides what action should be taken.

5. Alert Mechanism (Guardian → TTS)

Two possible paths:

a) Risk Detected (High Risk)

When the risk is above the danger threshold:

● The Guardian triggers Text-to-Speech (TTS).

● The system speaks a warning to the user in real time:
“This call seems dangerous. Please do not share any OTP or bank details.”
● The TTS module confirms that the message was spoken.

b) No Risk (Safe)

If the score is low:

9
 ● No sound is played.
● The system stays silent and waits for the next chunk.

6. Logging & Dashboard Update (Guardian → Dashboard)

Regardless of whether the call was risky or safe:

● The transcript
● Flags
● Risk score
● Any warnings given are all logged automatically

The dashboard uses these logs to show:

● A timeline of calls
● Risk heatmaps
● Transcript history
● Alerts for guardians or family members

6.2 System Architecture Diagram

Fig :2

This diagram shows how the entire GuardianVoice system works internally — from recording
audio to detecting scams, warning the user, and updating the dashboard. Each block represents
a component, and arrows show how data flows through the system.

1. Microphone / Audio Listener

10
 ● This is where everything starts.
● The microphone continuously listens to the call.
● It captures small audio chunks (5–6 seconds each) and sends them for processing.

2. Speech-to-Text (STT) Module

● Converts spoken audio into written text.

● Uses Whisper/Tiny or any STT engine configured.
● The output is the exact conversation transcript.

3. Preprocess and Normalize

● Cleans the text so it can be analyzed properly.

● Steps include:
o Lowercasing
o Removing noise or extra punctuation
o Normalizing numbers or special words
● This prepares the text for both rule-based and ML models.

4. Rule Engine (Keyword & Pattern Detection)

● Looks for suspicious keywords or phrases such as:

o “OTP”, “verification code”
o “urgent payment”
o “lottery”
o “account blocked”
o threats
● Flags each type of scam indicator.
● Works instantly and gives direct risk signals.

5. Machine Learning Classifier

● A trained model using TF-IDF + Logistic Regression.

● Reads the entire sentence or phrase.
● Predicts the probability of scam based on patterns learned from training data.
● Works even when scammers use tricky or indirect language.

6. Risk Combiner

This module merges information from:

● Rule Engine flags

● ML classifier probability

It produces a final risk score (0 to 1).

For example:

● 0.2 → Safe
● 0.6 → Suspicious
● 0.9 → Highly dangerous

11
The risk score decides what to do next.

7. Action Dispatcher

Based on the final risk score, this module chooses what action should happen:

Low risk

● No warning.
● Only dashboard update.

Medium risk

● Gentle audio warning (“This call seems suspicious…”).

● Log saved for review.

High risk

● Strong real-time warning (“Danger. Do not share any codes or money.”)

● Alert generated for guardians.
● Log entry created.

The dispatcher is the decision maker of the system.

8. Text-to-Speech Warning

● Converts the warning message into audio.

● Plays it in real time so the user hears it immediately.
● Helps prevent mistakes during scam calls.

9. Logger

● Stores every processed call chunk:

o transcript
o risk score
o flags
o any warnings issued
o timestamp

This is useful for analysis, auditing, and improving the model.

10. Alerts

● When a high-risk scam is detected:

o An alert file or message is created.
o The guardian dashboard can display it.
o Could be extended to SMS, notifications, etc.

12
11. Streamlit User Interface (Dashboard)

● Provides a visual dashboard for guardians or users.

● Shows:
o call history
o risk graph
o alerts
o detailed transcript
o timestamps

It stays updated with new logs sent by the Logger and Action Dispatcher.

6.3 Working Explanation

Fig:3

The system detects possible scam behaviour during a live phone call. It listens to the call,
converts speech into text, analyzes the text for suspicious content, and warns the user in real
time. It uses both rule-based checks and machine learning models to ensure accurate
identification of scam-like activity.

1. Audio Input Handling

The system starts by capturing the caller’s voice through the microphone. Instead of processing
the entire call at once, the audio is divided into smaller chunks, usually 6 or 15 seconds long.
Each chunk is processed separately to allow fast and continuous analysis throughout the call.

13
2. Coordination by the Guardian Module

The Guardian module controls and coordinates the whole workflow. For each audio chunk, it
sends the data to other components, collects their outputs, and makes the final risk decision. It
functions as the central controller, ensuring that all subsystems work together smoothly.

3. Speech-to-Text Conversion

Each audio chunk is sent to the Speech-to-Text component. This converts spoken words into
text. The text output becomes the basis for further analysis because both the rule-based and
machine learning components work on textual data.

4. Parallel Risk Analysis

4.1 Rule-Based Detection

This component checks the text for specific signs of scams. These include requests for OTP,
threats, urgent messages, banking-related terms, or common fraud keywords. Such predefined
rules help catch known scam patterns quickly.

4.2 Machine Learning Risk Scoring

A trained machine learning model evaluates the text and generates a risk score. This helps
detect scam behaviour even when it does not match predefined rules. The model is trained
using both scam and non-scam examples so it can recognize patterns in conversation.

Both results are sent back to the Guardian.

5. Decision Making

The Guardian combines the rule-based results and the machine learning score. If either
indicates high risk, or if both give moderate risk, the system labels the chunk as suspicious. If
the risk is low, it considers the conversation safe.

6. User Alert Through Text-to-Speech

When a risk is detected, the Guardian tells the Text-to-Speech module to speak a warning
message to the user. The warning is private and only the user hears it. If the conversation is
safe, no audio warning is generated.

7. Logging and Dashboard Updates: All processed chunks, including transcripts, risk levels,
and warnings, are stored in logs. The dashboard shows these updates in real time. This helps
the user or admin review what happened during the call.

7. IMPLEMENTATION
Technologies, Libraries & Packages Used

14
To build the Scam Call Detection System, multiple Python libraries and tools were used. Each
library serves a specific purpose in the workflow from capturing audio, converting it to text,
analyzing risk, and generating warnings. Below is a clear explanation of each technology used:

Python Libraries

1. sound device

● Used for recording audio from the microphone in real-time.

● Allows the system to continuously listen to the call and capture short audio chunks.

2. sound file

● Helps in saving the recorded audio into WAV format.

● Ensures compatibility with speech-to-text models.

3. whisper / speech recognition

● Converts the recorded audio into text.

● Whisper provides accurate transcription using deep-learning models.
● Text is required for fraud detection analysis.

4. scikit-learn

● Used to train a Machine Learning classification model.

● TF-IDF converts text into numerical form.
● Logistic Regression predicts if the call sounds suspicious.

5. joblib

● Saves and loads ML models efficiently.

● Helps avoid re-training the model every time the app runs.

6. numpy, pandas

● Used for data preprocessing and dataset management.

● Helps clean text, remove noise, and structure training data.

7. pyttsx3

● Text-to-speech library that generates voice warnings.

● Alerts the user immediately when a scam likelihood is detected.

8. streamlit

● Used to build a clean and interactive dashboard.

● Shows risk scores, call history, and real-time alerts.

15
9. matplotlib

● Generates charts and graphs for visual analysis.

● Helps in showing scam probability trends in the dashboard.

10. os, pathlib, json

● Used for file handling tasks like saving call logs, accessing models, etc.
●

Code Snippet - Core Pipeline :

Below is a simplified version of the main logic used in the system:

from src.audio_listener import record_chunk

from [Link] import transcribe_file

from [Link] import rule_checks

from [Link] import get_ml_prob, combine_scores

from src.tts_warn import speak_simple

audio_file = record_chunk(duration=6)

text = transcribe_file(audio_file)

flags = rule_checks(text)

ml_prob = get_ml_prob(text)

risk = combine_scores(flags, ml_prob)

if risk > 0.85:

speak_simple("Danger. This call seems fraudulent.", severity=3)

elif risk > 0.6:

speak_simple("This call appears suspicious.", severity=2)

Explanation of the Pipeline

1. Audio Recording

16
 ● The system listens constantly and records 6-second audio chunks.
● This prevents memory overload and allows fast processing.

2. Speech to Text Conversion

● The recorded audio is passed to the Speech-to-Text module.

● The output is clean, readable text used for analysis.

3. Rule-Based Checks

● The text is scanned for scam keywords and patterns such as:
o “urgent”
o “your bank account is blocked”
o “share OTP”
o “lottery”
● If such phrases appear, flags are generated.

4. Machine Learning Prediction

● The text is sent to the ML model.

● The model outputs a probability score (0 to 1) indicating scam likelihood.

5. Score Combination

● Rule-based flags and ML prediction are combined.

● The final score gives a more accurate risk estimation.

6. Warning System

● If risk is above 0.85, a strong warning is spoken.

● If risk is 0.60 - 0.85, a moderate suspicion alert is given.
● pyttsx3 converts the warning into audio so the user hears it immediately.

8. RESULTS AND ANALYSIS

17
 Fig :4

Fig: 5

1. Dashboard

The GuardianVoice dashboard provides a real-time view of scam-detection activities. It shows:

1.1 Threshold Controls

Two sliders allow the user to control how sensitive the system is:

● Alert Threshold – When the risk score goes above this value, the system gives a
warning.
● Escalate Threshold – When the risk score is even higher, the system marks the incident
as highly dangerous.

These thresholds can be adjusted easily and [Link] helps the system adapt to different
environments (quiet/noisy rooms, different languages, etc.).

18
1.2 Incident Summary

The dashboard displays:

● Total incidents detected

● List of recent incidents
● Risk score
● Detected action (alert or escalate)
● Transcript of what was spoken
● Audio file stored

This gives a quick idea of what kind of voice activity is happening and whether any scam-like
behavior occurred.

2. Detailed Incident View

The “Selected Incident” panel shows in-depth information about any incident:

2.1 Transcript

The exact sentence recognized from the audio is displayed.

Example:
“this is your bank verification team your account will be permanently blocked…”
This makes it easy to understand why the system triggered a warning.

2.2 Risk Score

The system assigns a score between 0 and 1.

Higher score → greater chance of the call being a scam.

Example shown:
Risk Score = 0.851
This indicates a high-risk message.

19
 Fig: 6

2.3 Detected Action

The system chooses between:

● alert
● escalate

In this case, the system selected: escalate

This means the message is very suspicious and needs attention.

20
2.4 Flags / Normalized Data

The system breaks down the message into flags such as:

● OTP request
● Threat
● Urgency
● Money request
● Explicit OTP request

Example:

● otp_request : true
● otp_explicit_request : true
This means the message clearly asked for an OTP a common scam behavior.

2.5 Quick Actions

User can:

● Mark incident as reviewed

● Delete the incident

This helps in managing many alerts during long-term monitoring.

3. Machine Learning Results

3.1 Confusion Matrix

Fig: 7

The confusion matrix shows:

● 118 benign sentences correctly classified

● 122 scam sentences correctly classified
● 0 misclassifications
21
This means:

● No false positives
● No false negatives
● 100% accuracy on the test dataset

This is a very strong performance, showing that the classifier distinguishes clearly between
scam and non-scam phrases.

3.2 ROC Curve

Fig:8

The ROC curve shows:

● AUC = 1.0

AUC of 1.0 means:

● Perfect separation between scam and non-scam classes.

● The model is extremely good at detecting scams.

Normally, most real-world models have AUC between 0.8–0.95.

A perfect AUC means the dataset and model matched very well.

4. System Execution Output

22
 Fig:9

The terminal output shows the real-time behavior during audio monitoring.

4.1 Audio Recording

The system records short audio chunks (like 15 seconds).

Each chunk is stored in the tmp_audio folder.

4.2 Speech-to-Text Output

If the audio is unclear or silent, the Google Speech Recognition API returns:

● “could not understand audio”

This is normal because:

● background noise
● silence
● unclear speech
can cause empty transcription.

4.3 System Reactions

When transcription is empty:

● The system ignores it and waits for the next audio chunk
● No false alarms are [Link] shows the system is stable and does not react to random
sounds.

5. Overall Analysis

23
5.1 Strengths of the System

● Accurate scam detection: Perfect model performance in tests.

● Live monitoring: Records audio continuously.
● Real-time warnings: Immediate alerts when scam patterns appear.
● Transparent dashboard: Shows all details (risk score, flags, transcript).
● Customizable sensitivity: Threshold sliders allow tuning based on user needs.

5.2 Practical Usefulness

The system effectively identifies classic scam patterns such as:

● Asking for OTP

● Threatening to block accounts
● Creating urgency
This can be very useful for:
● Elderly protection..etc

24
 9. CONCLUSION
The Scam Call Detection System (GuardianVoice) successfully demonstrates how artificial
intelligence can be used to enhance personal safety and protect users from fraudulent phone
calls. By combining speech-to-text transcription, rule-based keyword analysis, and machine
learning classification, the system provides an efficient and reliable method for identifying
scam-like conversations in real time.

The hybrid approach improves accuracy and reduces false positives, ensuring that legitimate
conversations are not interrupted unnecessarily. Real-time processing ensures quick warnings,
which is crucial during active scam attempts. The dashboard further adds value by visualizing
risk scores, transcripts, and previous alerts, helping users understand the nature of detected
threats.

Overall, the project achieves its objective of providing a fast, practical, and user-friendly tool
to help individuals especially vulnerable groups like senior citizens identify and avoid phone-
based scams. With future enhancements such as mobile integration and cloud-based updates,
the system can evolve into a comprehensive safety solution for households and personal
devices.

25
 [Link]
1. OpenAI. Whisper: Robust Speech Recognition. GitHub Repository.
[Link]
2. Pedregosa, F., et al. (2011). Scikit-learn: Machine Learning in Python. Journal of
Machine Learning Research, 12, 2825–2830.
3. Jurafsky, D., & Martin, J. H. (2023). Speech and Language Processing (3rd Edition
Draft). Pearson.
4. Python Software Foundation. Python Language Reference.
[Link]
5. Streamlit Inc. Streamlit Documentation — Web Apps for Machine Learning.
[Link]
6. Deng, L., & Li, X. (2013). Machine Learning Paradigms for Speech Recognition: An
Overview. IEEE Transactions on Audio, Speech, and Language Processing.
7. Chen, L., & Ngai, G. (2020). Detecting Phone Scams Using Machine Learning
Techniques. IEEE Symposium on Security and Privacy.
8. Federal Trade Commission (FTC). Consumer Fraud and Scam Reports.
[Link]
9. Bishop, C. M. (2006). Pattern Recognition and Machine Learning. Springer.
10. Aggarwal, C. (2018). Machine Learning for Text. Springer.
11. Sculley, D., & Pasanek, B. (2008). Mining Text Data Using TF-IDF. Journal of Digital
Humanities.
12. Symantec Cybersecurity. Phone Fraud Trends and Scam Detection Guidelines.
[Link]

26