I’ve been studying Japanese for a few years pretty seriously (anki decks, textbooks, online tutor) but I always felt that the speaking practice pipeline could be optimized. With a real tutor, my mistakes didn’t get enough specific reps, they oftentimes weren’t pedantic enough, and practice in multiple contexts was a problem. (I also subconsciously avoid saying things I know I might mess up on to not sound dumb to a real human)
I’m pretty bearish on most other apps, they seem mostly for casual learners with streaks/gamifications and canned scenarios with shallow feedback, even the recent stuff using AI. I for one hate using mobile apps to study (small screen, notification driven, gamified, etc) and I really wished I could deep dive/rabbit hole into my mistakes and drill them specifically.
I built Lingle as a desktop-first tool explicitly for learners who already have a serious study routine and want to work on deliberate practice for speaking with rigorous feedback and detailed explanations and more pedantic feedback to bridge the gap between understandable and natural.
Details:
-Voice input with real-time transcription fed into a conversation model prompted to behave as a native speaker in a user-defined context
-A second pass evaluates utterances for naturalness, register, and pragmatic appropriateness, not just grammatical correctness
-Corrections surface inline without interrupting conversation flow, allows you to retry if you made a suboptimal response
Users define their own contexts entirely — no preset scenarios, you tell it what you're watching, reading, or which grammar structures you want to drill
watching, reading, or which grammar structures you want to drill
The hardest problem has been calibrating correction sensitivity - getting the model to catch subtle pragmatic errors without overcorrecting or flagging legitimate stylistic variation. Still actively working on this and would genuinely love input from anyone who's thought about this problem. Also thinking about adding longer term memory and integrating with other study tools (Anki, etc) to make it more fullstack. Maybe one day replacing my italki tutor entirely with AI + memory and textbook integration or something.
Currently Supports Japanese, Spanish, French, Mandarin, German, Korean, and Portuguese. Free to try while I validate whether this is a real problem worth building further.
Happy to get into the technical architecture or prompting approach in the comments.
– Andrew
0 comments
—
Show HN: Desktop conversation practice tool for serious language learners · HackerTrans
I’m pretty bearish on most other apps, they seem mostly for casual learners with streaks/gamifications and canned scenarios with shallow feedback, even the recent stuff using AI. I for one hate using mobile apps to study (small screen, notification driven, gamified, etc) and I really wished I could deep dive/rabbit hole into my mistakes and drill them specifically.
I built Lingle as a desktop-first tool explicitly for learners who already have a serious study routine and want to work on deliberate practice for speaking with rigorous feedback and detailed explanations and more pedantic feedback to bridge the gap between understandable and natural.
Details:
-Voice input with real-time transcription fed into a conversation model prompted to behave as a native speaker in a user-defined context -A second pass evaluates utterances for naturalness, register, and pragmatic appropriateness, not just grammatical correctness -Corrections surface inline without interrupting conversation flow, allows you to retry if you made a suboptimal response Users define their own contexts entirely — no preset scenarios, you tell it what you're watching, reading, or which grammar structures you want to drill watching, reading, or which grammar structures you want to drill
The hardest problem has been calibrating correction sensitivity - getting the model to catch subtle pragmatic errors without overcorrecting or flagging legitimate stylistic variation. Still actively working on this and would genuinely love input from anyone who's thought about this problem. Also thinking about adding longer term memory and integrating with other study tools (Anki, etc) to make it more fullstack. Maybe one day replacing my italki tutor entirely with AI + memory and textbook integration or something.
Currently Supports Japanese, Spanish, French, Mandarin, German, Korean, and Portuguese. Free to try while I validate whether this is a real problem worth building further. Happy to get into the technical architecture or prompting approach in the comments.
– Andrew