Intelligent Spam Call Classification Using Machine Learning

The widespread adoption of automated telephony and Voice over Internet Protocol (VoIP) services has
generated an unprecedented surge in malicious call traffic, placing mobile users at risk of financial
exploitation and privacy violations. Conventional call-screening tools—which rely on static, number-based
blacklist databases—are structurally incapable of countering contemporary threats such as AI-synthesized
voice impersonation (deepfakes) and Caller ID spoofing. This study introduces a novel real-time machine
learning framework designed for automated spam call detection, functioning by concurrently evaluating
the network-level provenance of an incoming call and the acoustic characteristics of the transmitted voice
payload. The framework is realized as a native Android application built on a Kotlin Coroutine
concurrency model, enabling simultaneous lookups across multiple global spam repositories (Truecaller,
Tellows, and ListaSpam) combined with on-device STIR/SHAKEN protocol validation. When a call is
flagged as suspicious, the corresponding audio stream is forwarded to a Python FastAPI backend, where
40 Mel-Frequency Cepstral Coefficients (MFCCs) are computed to encode the timbral signature of the
voice. A purpose-built PyTorch feedforward neural network subsequently classifies the audio as either
authentic human speech or a synthetic artifact, achieving an overall test accuracy of 94.2%. By unifying
caller provenance verification with deep acoustic fingerprinting into a single hybrid decision engine, the
proposed system substantially lowers false positive rates and processing latency, yielding a scalable and
resilient countermeasure against emerging telecommunication threats.