The researchers developed a new system for real-time debate training which uses Large Language Models (LLMs) to function as three different systems that compete with users while teaching them privately and detecting multiple types of abuse through one system during live voice interactions. The system provides ongoing feedback during live debates because it differs from Debatrix [1] which only evaluates completed debates through automated judging. The system uses WebRTC-based voice transport [16] to deliver audio through its dual-speech recognition system which combines Web Speech API with Vosk WASM fallback and uses Socket.IO to manage real-time events while the Groq API performs LLM inference. The system uses a dual-layer dynamic health and credibility scoring model which replaces traditional static verdicts to create an educational experience that resembles a video game. The system uses a three-strike moderation system to oversee both live voice transcripts and chat messages with the help of the Hugging Face unbiased-toxic-roberta model [7]. The evaluation process shows that the system can perform real-time adversarial rebuttals while providing private coaching during every spoken sentence and enforcing multimodal safety measures which no other existing debate AI system can accomplish.