多元的閱讀活動能提昇幼兒的學習意願，透過聊天機器人的對話引導將能吸引幼兒注意力，增加閱讀的樂趣。市面上幼兒陪伴的機器人產品缺少能直接以自然語言互動的模式，而這些產品在面對幼兒不同的情緒反應時，將無法適時給予回饋。本研究提出能以自然語言互動的幼兒學習人機對話系統，為一個語音對話的機器人故事機。藉由一來一回的問答聊天，引導幼兒對童話故事的閱讀及理解。系統架構包含語音辨識、語意理解和文字轉語音，先以弱監督式學習的Whisper模型識別語音對話，再使用大型預訓練語言模型GPT-3.5以文字回應對話內容，最後使用Amazon Polly將文字回應轉換成自然的人類語音輸出。本研究驗證了基於多重AI引擎的幼兒互動學習對話框架，藉此提供後續幼兒學習應用系統的開發基礎。;Diverse readings can enhance young children′s willingness to learn. Engaging in conversations with chatbots can attract their attention and increase the enjoyment of reading. Robot products available in the market for young children lack the ability to interact directly in natural language, and they are unable to provide appropriate feedback with various reactions from young children. In this study, we propose a natural language interactive human-machine dialogue system for early childhood learning, specifically a voice-based story-telling chatbot. The system guides young children in reading by back-and-forth conversations. The system performs speech recognition, semantic understanding, and text-to-speech conversion. We use Whisper model in speech recognition, and generate text responses based on GPT-3.5, and finally convert the text responses into speech through Amazon Polly. We demonstrate effectiveness of a multi-AI engine-based framework for interactive dialogue systems in early childhood learning, which can serve as a foundation for future applications.
