Construction and Optimization of Deception Detection Models Based on Audio Features

This project aims to construct and optimize a deception model based on audio feartures, aiming to distinguish between truthful and deceptive narratives. In the era of rapidly spreading digital information, the detection of deceptive audio content has become increasingly critical to combat misinformation. This research collected a dataset of 100 audio clips, each 60 seconds in duration, recorded by speakers from various countries in their native and non-native languages. Key audio features, including pitch, power, jitter, shimmer, and zero-crossing rate (ZCR), were extracted from the audio clips using the librosa library, a widely used tool for audio signal processing. To classify the audio data, several machine learning models were implemented, such as Histogram-based Gradient Boosting Decision Trees (HGBDT), Random Forest, Support Vector Classifier (SVC), and Logistic Regression. These models were chosen for their complementary strengths in handling different types of data and classification tasks. To further improve the models performance, an ensemble approach was adopted using the VotingClassifier, which combines the predictions of the individual models. The experimental results demonstrasted that the proposed model achieved an accuracy of 75%, successfully identifying subtle differences between truthful and deceptive speech patterns. This research highlights the effectiveness of audio-based deception detection in multilingual contexts and lays the groundwork for future studies.