Speech-based analysis of aggression in social content: resistance of Large Language Models to signal degradation