인공지능 챗봇 음성인식의 오류분석 기반 유용도 탐색

Abstract

With the advent of Artificial Intelligence (AI) technology utilized in English classrooms, the present case study attempts to explore the strengths and limitations of Automatic Speech Recognition (ASR) supported by AI chatbots. This study selected 20 experimental sentences, which were found to be difficult for a Dialogflow-driven chatbot to recognize in EFL classes. Then, the number of successful speech recognition cases was measured on the basis of the speech uttered by three groups (i.e., native English speakers, Korean-English bilinguals, and EFL students). It showed that the chatbot found it difficult to recognize fast speech phenomena including Tapification and/or Contraction. Moreover, the chatbot recognized bilingual speakers' English pronunciation most accurately, followed by native English speakers and EFL students, which might be attributed to the ASR algorithm based on the spoken English as an International Language (EIL). An additional analysis employing the Praat revealed that this was due to the duration of speech and pauses and that this sophisticated technology did not appear to be fully functional in recognizing unstressed sounds with a low level of duration. Finally, the present study suggests some ideas for designing and utilizing chatbots to achieve a higher level of efficacy in TEFL.