BubbleTalk: A Child-Friendly Conversational Speech Large Model

CSC5051: Natural Language Processing
CUHK-Shenzhen | MAIR Group 25

🧒 Motivation

While voice-based AI has become increasingly popular, most systems remain adult-oriented and unsuitable for children. They often:

• Use overly complex language
• Lack engaging and expressive voice styles
• Fail to convey warmth or empathy

BubbleTalk aims to bridge this gap with a child-centric conversational system that delivers cognitively appropriate language, emotional expression, and consistent high-quality interactions.

🛠️ Methodology

BubbleTalk integrates several cutting-edge components:

• End-to-End Integration
  Seamless combination of:

  • Speech Recognition (ASR): Whisper
  • Large Language Model: Langchain
  • Speech Synthesis (TTS): CozyVoice

• Prompt Engineering for Children
  Specially designed prompts tailored to children's comprehension and interaction needs.

• Adaptive Multi-Emotion Expression
  Supports 8 emotions (e.g., happiness, sadness, curiosity) with dynamic tone control.

• Diverse Character Voices & Personalized Cloning
  Multiple cartoon voice options and user-uploaded (e.g., parental) voice cloning.

• Global Memory & Contextual Flexibility
  Handles repetitive questions and non-linear speech patterns.

• Low-Latency Streaming
  Real-time speech synthesis for smooth user experiences.

🧪 Results

A user study with 12 children (ages 4–8) and their caregivers showed:

• 83.3% of children preferred BubbleTalk’s expressive character voices.
• 87.5% of parents preferred BubbleTalk over traditional assistants for bedtime or educational use.

🎯 Features Summary

• Child speech-optimized interaction
• Emotionally rich responses
• Age-aware content filtering
• Real-time emotion analysis
• Voice personalization

📫 Contact

• Lin Zhiyu: 224040288@link.cuhk.edu.cn
• Yang Jingwen: 224040204@link.cuhk.edu.cn

🎬 Demo

Emotion-adaptive & Multi-Voice

Click