Lang Tutor

Lang Tutor teaches by comparison. Instead of drilling rules in isolation, every topic starts from what your own language already does — and then shows exactly where the new one diverges. A German lesson opens by pointing out that English barely conjugates at all, and that this is the habit you will have to unlearn.

A lesson is five to ten minutes: a short explanation, a run of exercises, the vocabulary that came up, and a summary. Everything you get wrong is scheduled for review, and the course adapts — a placement test sets the starting point, and a pass-out exam lets you skip what you already know.

The screens below show the app configured with an English interface teaching German, one of the language pairs it supports.

Screens are shown with sample content.

• Why your answer was wrong

  When an exercise is failed, gemini-2.5-flash-lite writes the explanation — not a generic rule, but why your specific answer did not work, anchored back to your own language.

• Marking free-form answers

  When an answer cannot be matched offline, the model grades it by meaning, so a correct sentence phrased differently still counts.

• Live spoken conversation

  The voice lesson opens a duplex WebSocket session with gemini-3.1-flash-live-preview: PCM audio in and out, both sides transcribed as subtitles, interruptible mid-sentence. Silence sensitivity is deliberately relaxed — learners need time to think. The model ends the session itself through a tool call.

• A cheaper turn-by-turn mode

  For everyday practice there is a second mode: the model writes one short reply, on-device recognition handles your speech so no audio leaves the phone, and the correction arrives as a separate note rather than the character breaking role to correct you.

• Pronunciation feedback

  Recorded speech is sent inline for a breakdown of how it was pronounced, and a closing report lists goals met, phrases that worked, and mistakes to carry into review.

• Deterministic where it should be

  Progress tracking, weak-spot detection, spaced repetition and level estimation are plain code, not model calls. Extra practice is assembled from a prepared pool. The model is used where judgement is needed — not to decide what you know.

Lang Tutor is not yet available for download. When it ships, the store link will appear here.