Robust Video Plagiarism Detection Using Word Embeddings from Audio Transcriptions

Abstract

Video piracy presents a significant challenge in the digital era, requiring effective detection methods to protect intellectual property. This paper proposes a novel approach for detecting video plagiarism by leveraging word embeddings derived from audio transcriptions. Our method begins by extracting audio streams from videos and transcribing the audio content. We then generate semantic embeddings, storing these embeddings in a vector store for efficient similarity searches. To identify potential plagiarism, query videos are processed through the same pipeline, and their embeddings are compared against reference embeddings. A Euclidean distance below a predefined threshold indicates possible plagiarism, enabling accurate classification and identification of plagiarized videos. Experimental evaluations demonstrate the method’s scalability and efficiency, particularly in detecting complete video copies with explicit English speech content. This approach offers a robust and scalable solution against joint video manipulations, providing a practical framework for combating video piracy in large-scale content environments.