Audio-Based Cross-Language Asset Recognition for a Scalable CTV Future

The Connected Television (CTV) ecosystem relies on diverse media pipelines across regions, making content identification increasingly complex. Traditional fingerprinting detects exact duplicates but struggles with near duplicates, especially when content is modified or lacks semantic awareness. This paper presents a hybrid audio-based identification pipeline, integrating audio fingerprinting for sound-based recognition and automatic speech recognition (ASR) for semantic analysis. Leveraging OpenAI's Whisper and deep learning models, the system enhances redundancy detection even across transformed media. Evaluation results show that full transcription with fingerprinting ensures high accuracy for exact and near-duplicate media, while summarized transcriptions improve efficiency at the cost of accuracy. This scalable, multimodal approach optimizes large-scale multimedia storage while maintaining precise content identification.