Intelligent Spam Call Detection using Natural Language Processing and DistilBERT

Abstract

With the rapid growth of mobile communication, spam and fraudulent calls have become a significant threat to users, leading to financial loss and privacy breaches. Traditional spam detection methods that rely only on call logs or blacklisted numbers are insufficient to detect content-based scam attempts. A content-based spam call detection system based on DistilBERT is presented for detecting fraudulent and spam calls using speech-to-text transcription and transformer-based natural language processing. The system converts call audio into text using speech-to-text processing, after which the transcribed content is analysed using DistilBERT for contextual understanding and classification. Unlike traditional machine learning models, DistilBERT captures semantic meaning, tone, and hidden patterns within the conversation, enabling more accurate spam detection. The model is trained and evaluated using accuracy, precision, recall, and F1-score. Experimental results demonstrate improved detection capability for phishing and scam calls, achieving up to 98% accuracy and a 95% F1-score on 6,000 labelled samples. The proposed system enhances real-time security and can be integrated into mobile or telecom platforms for intelligent spam call prevention.