Toward pedagogically grounded design and evaluation of generative artificial intelligence speaking tools in English as a foreign language education

Generative artificial intelligence has accelerated adoption of AI-enabled speaking tools in English as a Foreign Language (EFL) education, but the field still lacks a shared way to interpret results across highly heterogeneous software systems. This perspective paper argues that progress now depends on treating generative AI speaking tools as educational software artifacts rather than as generic model access, because learning opportunities are shaped by design decisions that determine speaking tasks, feedback behaviour, and classroom orchestration. The paper identifies why the current evidence base is difficult to trust and difficult to cumulate, with recurring problems related to intervention heterogeneity, outcome operationalizations that privilege rehearsed formats over transfer to spontaneous speaking, and limited reporting of feedback validity, speech recognition reliability, and failure modes. A compact three-layer framework is proposed to guide design and evaluation, linking the speaking construct targeted, the feedback model implemented, and orchestration and governance conditions for responsible classroom use. Building on this framework, the paper outlines minimum standards for credible claims, emphasizing transparent reporting of tool configuration, routine documentation of reliability and breakdown patterns when voice and scoring pipelines are involved, outcome designs that include transfer and durability, and comparative studies that isolate feedback and orchestration variables. Practical implications are articulated for developers, educators, and researchers.

Hali tarjima qilinmagan