LUTHER is a cross-platform mobile application for language learning through audio-only
spoken conversation. The app is designed for learners who want to improve their speaking skills
in a safe, personal, and non-judgmental environment. Unlike most language-learning
applications, which rely mainly on text, quizzes, translations, or gamified drills, LUTHER
focuses on the part many learners avoid most: speaking out loud.
The motivation behind LUTHER comes from a common gap in language learning. Many
learners can read, understand, or passively recognize words, but still struggle to speak naturally.
Fear of making mistakes, embarrassment, lack of patience from others, and limited access to
tutors often prevents learners from practicing real conversation. LUTHER addresses this by
acting as a personal AI language tutor that listens to the user, analyzes their spoken sentences,
and responds with short, clear spoken feedback. The goal is to create a practice experience that
feels closer to a real tutor/friend conversation, while remaining available anytime and anywhere.
LUTHER implements a complete end-to-end AI pipeline. The user presses a microphone button
and speaks in the target language. The recorded audio is sent to a FastAPI backend, where it is
transcribed using OpenAI Whisper. The transcript is then processed by a GPT-4.1-nano
language-coaching engine, which analyzes grammar, vocabulary, and sentence structure. The
model produces structured feedback using structured Outputs, allowing the system to reliably
extract a corrected sentence, key error explanations, error categories, severity levels, and a
follow-up question. Finally, the response is converted back into natural speech using
ElevenLabs text-to-speech, and the user hears the correction and explanation as audio. The
frontend is implemented in Flutter, supporting Android, iOS, and web. The interface includes a
microphone-based interaction loop, listening/recording/processing indicators, language and
level settings, and a friendly visual identity based around the AI tutor avatar, Shauli.
LUTHER currently targets beginner-to-intermediate learners, approximately A1–B2 CEFR
levels. The feedback focuses on common learner mistakes such as articles, verb tense,
prepositions, word order, and vocabulary choice. It also considers language-specific challenges
such as gender, cases, verb placement, and separable verbs.
A key feature of LUTHER is personalized learning through error analytics. Each user interaction
is stored in a local SQLite database, including detected error types, corrections, and recurring
mistakes. This allows the system to identify repeated weaknesses over time and adapt future
feedback accordingly. In addition, the app includes learning-support features inspired by
successful educational apps, such as progress tracking, streaks, medals, points, tips and tricks,
and structured lessons. The project also includes AI-generated lesson support.