Embodied Multimodal Chatbot for Mental Health Support in Web-Based 3D Environments

This paper presents the design, implementation, and evaluation of a multimodal intelligent conversational agent embodied within a WebGL-accessible 3D virtual environment. The system integrates voice and text-based interaction with a personalized avatar capable of delivering emotionally attuned and context-aware responses. Developed using Unity WebGL, Flask microservices, and containerized with Docker, the architecture incorporates SpeechRecognition for local speech-to-text (STT), and a fine-tuned DeepSeek-r1 large language model deployed via the Ollama framework. The agent's dialogue engine has been specialized through domain-specific transfer learning focused on mental health support, enabling empathetic and contextually appropriate responses. The fine-tuned model achieves an average response time that is only 23.5 % of that of the original DeepSeek model, while also demonstrating superior semantic similarity performance. The personalized avatars are synchronized dynamically with the affective content of each interaction, enhancing user engagement. A usability study with 20 participants demonstrates high satisfaction, low-latency performance, and emotional resonance. This work contributes a modular framework for deploying browser-based embodied AI agents in sensitive domains such as digital mental health care.