Building on recent advances in Natural Language Processing (NLP), this work addresses the task of fallacy detection in political debates using a multimodal approach combining text and audio, as well as text-only and audio-only approaches. Although the multimodal setup is novel, results show that text-based models consistently outperform both audio-only and multimodal models, confirming that textual information remains the most effective for this task. Transformer-based and few-shot architectures were used to detect fallacies. While fine-tuned language models demonstrate strong performance, challenges such as data imbalance, audio processing, and limited dataset size persist.