Until the recent emergence of AI-based tools on the pedagogical market, classic spellcheckers and grammar correctors were commonly used by L1 and L2 writers to automatically assess, correct, and improve their writing, both at the orthographic and at the grammatical levels. Thanks to such tools, L2 learners have been able to improve their writing skills in an autonomous manner through various activities. Speaking skills training, on the other hand, remains difficult. One of the reasons behind the lack of available tool for autonomous L2 speech assessment is linked to the variability between the phonetic, the phonological and the lexical levels of analysis that may be attached to the forms produced by the learner. This becomes even more challenging for automatic speech processing systems when the L2 contains cohorts of phonological neighbors (e.g., mangent [mɑ̃ʒə] vs mangeons [mɑ̃ʒɔ̃]) and homophones (e.g., mangé vs manger both as [mɑ̃ʒe]). In the case of Japanese learners of French, the phonetic errors induced by interphonological features in learners' speech may lead to linguistic ambiguities that standard automatic speech recognition systems may not be able to solve (e.g. does [ʒɛmleʃokola] correspond to “J'aime le chocolat” (‘I like chocolate') with a phonetic substitution between [e] and [ə] or “J'aime les chocolats” (instead of “J'aime le chocolat”) with a morphophonological substitution between the plural and the singular form of “le chocolat”?). In this presentation we describe an attempt to tackle this challenge by using a computer-interfaced translation task using a double output modality (spoken and written) to assess the possibility of designing methodological guidelines to bypass the technical hurdle. In our study, 12 sentences containing typical morphophonological learning difficulties for Japanese learners of French were selected, and 40 learners were asked to translate them from Japanese to French, both in a spoken and in a written form, in a constrained double-task recording protocol. The collected data was then automatically (for the written production) or manually (by human evaluators) assessed and the results were compared to determine whether such method could help to build L2 speech learning applications allowing for translation-based automatic L2 spoken lexical form evaluation. Our results support our approach since our protocol allows us to identify the nature of the substitution between the target element and the learner's production, i.e., whether they are phonetic or morpholexical substitutions. This constitutes a methodological basis for pedagogical developments in L2 speech learning applications, with automatic L2 speech assessment at the lexical level. References De Fino, V., Fontan, L., Pinquier, J., Ferrané, I. & Detey, S. (2022). Prediction of L2 speech proficiency based on multi-level linguistic features. Proceedings of Interspeech2022, 4043-4047. Detey, S., Fontan, L., Le Coz, M., Barcat, C. & Kawaguchi, Y. (2022). Identifying segmental substitutions in spontaneous speech of L3-French/L1-Japanese learners: a corpus-based pilot study. Journal of the Phonetic Society of Japan 26(3): 124-134. Detey, S., Fontan, L., Le Coz, M. & Jmel, S. (2020). Computer-assisted assessment of phonetic fluency in a second language: a longitudinal study of Japanese learners of French. Speech Communication 125: 69-79.