Sarcasm Detection in Monolingual Speech Using Deep Learning–Driven Sentiment Analysis

Abstract

Sarcasm detection in spoken language is a challenging task due to the implicit and often contradictory relationship between literal expression and intended sentiment. In audio-based communication, sarcasm is primarily conveyed through acoustic and prosodic cues such as pitch modulation, intonation patterns, speech rate, and energy variation rather than explicit lexical indicators. This study presents a deep learning–based approach for sarcasm detection using sentiment analysis of a monolingual audio corpus. The proposed framework focuses on extracting sentiment-aware acoustic features, including Mel-frequency cepstral coefficients, prosodic features, and spectral characteristics, which are then modeled using deep neural architectures to capture both spatial and temporal dependencies in speech signals. By exploiting sentiment incongruity between vocal expression and underlying intent, the system aims to distinguish sarcastic utterances from non-sarcastic ones more effectively. Experimental evaluation demonstrates that integrating sentiment-oriented features with deep learning models significantly enhances sarcasm detection performance, highlighting the potential of audio-based sentiment analysis for improving speech-driven intelligent systems.