EVA: Emotion-Aware Conversational AI with Personalized Voice for Adolescents
Abstract
Adolescent mental health is becoming a global issue, with many young people facing emotional distress and struggle to find caring and responsive support. Traditional conversational AI tools, like chatbots and voice assistants, often disappoint because it provides generic answers which lack in emotional depth and also use synthetic voices. To fill this gap, the proposed framework introduces an emotionally adaptive, personalized voice agent for conversation. This agent is designed to be a companion for ado-lescents. In addition, it explores subtle replication of individual personality traits through adaptive responses and voice changes allowing the system to be a more relatable peer-like companion. The model uses a multimodal approach that combines several advanced techniques in audio processing, emotion detection, natural language creation, and personalized speech synthesis. The proposed framework includes four different emotion agents: happy, sad, angry, and neutral. These agents adjust the tone and content of responses dynamically. This is backed by carefully training on a selected dataset and reviewed with user emotions, agent emotions, and response goals. The developed dataset was specifically curated to address the lack of emotionally annotated, dialogue-style materials developed for adolescents. The suggested method was compared with prior state-of-the-art methods based on assessment in ROUGE, perplexity and cross-entropy loss and it performed significantly better, validating its potential to provide more emotionally resonant and context-aware conversational support.