This study aimed to examine speech learning effects among speakers of English as a foreign language (EFL) occurring after conversations with generative artificial intelligence (AI) and to explore effective speech learning methods using generative AI. As learners face difficulties in auditory perception during the early stages of foreign language learning, this study focused on perceptual learning effects obtained from English conversation practice with generative AI. This study conducted experiments using generative AI equipped with speech recognition and text-to-speech synthesis technologies and employed quantitative and qualitative data analyses. The participants were 31 first- and second-year undergraduate students at a university in Japan who were EFL learners at Common European Framework of Reference (CEFR) A1–B1 levels. They participated in two experimental sessions held three to four weeks apart. In each session, they engaged in four 5-minute conversations with ChatGPT in English using formulaic sequences (FSs) or fixed expressions. In each session, they engaged in four 5-minute verbal conversations with ChatGPT in English using formulaic sequences (FSs) or fixed expressions. Prompts written by one of the researchers were used to initiate four types of conversations guided by ChatGPT. Each session lasted 30 minutes. The mere-exposure experimental paradigm was used to indirectly measure perceptual fluency; the participants were presented with synthetic speech of the FSs in English, and differences in favorability toward these expressions before and after the sessions were measured. The results demonstrated that favorability toward the FSs increased significantly after the first session (BF = 400.511) but not after the second session (BF = 0.908). Furthermore, interviews were conducted with 26 participants, and the results suggested that the lack of a significant change in the second session was partly due to the increased complexity of conversational content following a ChatGPT version update, which made the conversations difficult to understand. All participants reported that generative AI use assisted their speech learning; however, no change in favorability toward the FSs was observed among 18 participants. Notably, eight out of the 28 participants who reported positive effects and three out of five participants who exhibited increases in favorability but did not report positive effects stated that the speech quality was good, and the speech was easy to understand. The clarity of the AI-generated speech was influenced by multiple factors, including pronunciation, accent, speech rate, pauses, rhythm, intonation, prosody, and sound quality. Many participants noted that the speech rate was fast; however, this did not affect understanding, suggesting that other factors influenced speech clarity. Moreover, the qualitative results indicated that positive feelings toward generative AI and proactive learning attitudes may have enhanced speech learning effects. These findings suggest that learners should use generative AI with high speech clarity to improve their perceptual fluency in English conversation. Moreover, learners require positive attitudes towards generative AI and learning.

Keywords: Speech learning effect, generative AI, perceptual learning effect, formulaic sequence, proactive learning attitude.