LearnChinese AI helps young and heritage-language learners study Mandarin through accurate human-recorded pronunciation, visual vocabulary, stroke-order practice, bilingual guidance, and immediate interactive feedback in one browser-based lesson.
Inspiration
Young Chinese learners often need several kinds of support at the same time: accurate pronunciation, visual context, stroke order, repeated practice, immediate feedback, and clear English guidance. In many classrooms these resources are scattered across a printed textbook, separate recordings, worksheets, and unrelated websites.
LearnChinese AI was inspired by a practical teaching need: make one lesson easier for children to understand and practise independently, while preserving the teacher's voice and the structure of the original curriculum. The goal is not to replace the teacher. It is to turn carefully prepared teaching material into a coherent, child-friendly learning experience.
What it does
The Build Week demo transforms Lesson 6 of a Chinese language course into an interactive bilingual experience covering pages 47–53.
Learners can:
- explore illustrated Chinese characters with pinyin and English meanings;
- hear Chinese and English vocabulary audio;
- watch stroke-order animation and use tracing practice;
- study all 21 Mandarin initials and 21 initial-plus-vowel examples;
- play 66 independent, locally stored pronunciation clips, including 24 tone-practice items;
- complete image-based character recognition and classroom exercises;
- read a short classical Chinese passage with bilingual source and meaning explanations;
- reset activities and stop audio cleanly without overlapping playback.
The demo is offline-first. Its core lesson, images, fonts, stroke data, and pronunciation audio are stored locally, so the learning experience does not depend on browser text-to-speech or a continuous network connection.
How we built it
LearnChinese AI is a lightweight static web application built with HTML, CSS, and JavaScript. Curriculum content is stored in JSON, and a small local audio engine maps lesson items to human-recorded MP3 clips. Hanzi Writer provides stroke-order rendering using locally stored character data. The interface is bilingual and designed for children who may not yet recognize instructional Chinese labels.
The implementation process combined educator-led content decisions with AI-assisted software development:
- the educator selected the curriculum scope, recorded the Mandarin audio, checked pronunciation, and performed the final visual and listening tests;
- Codex and GPT-5.6 helped inspect the existing project, implement and refine the interface, integrate local media, diagnose encoding and playback problems, improve reset behavior, validate resource paths, and prepare the submission repository and documentation;
- repeated human testing identified issues that automated checks could not reliably judge, especially audio segmentation quality and child-facing language clarity. Xiaomi MiMo contributed to the earlier prototype and foundational lesson implementation, while Codex and GPT-5.6 supported the Build Week refinement, debugging, audio integration, testing, and submission preparation. Important clarification: Codex and GPT-5.6 were used to build, debug, test, and package the project. The submitted demo itself runs locally and does not require a live OpenAI API call at runtime.
Challenges we ran into
The hardest challenge was audio correctness. Automatic segmentation based only on waveform changes produced clips that were too short or captured instructional speech instead of the intended pronunciation. Because pronunciation quality is central to language learning, the solution was to record three new structured source files: 21 initials, 21 spelling examples, and 24 tone-practice syllables, with deliberate pauses between items. These were then divided into 66 independent local clips and checked by a human listener.
We also encountered character-encoding problems after moving the project through Windows ZIP workflows. The visible result was mojibake in Chinese interface text. We corrected the affected files and treated browser rendering as a required acceptance test rather than relying only on syntax checks.
A third challenge was audio state management. Several playback paths could run simultaneously, and the reset button initially changed only the visual state. We unified playback behavior so that starting a new item stops the previous one and reset also stops active audio.
Finally, bilingual usability required more than translating the lesson text. Instructional controls such as Stroke Order, Tracing Practice, Reset, Read Aloud, Guess the Character, Try Again, and Reset Matching also needed clear English labels for children who are still learning Chinese.
Accomplishments that we're proud of
- Converted a real printed lesson into a coherent interactive learning flow.
- Replaced unreliable browser pinyin speech with 66 teacher-recorded local MP3 clips.
- Integrated all 21 initials, 21 spelling examples, and 24 tone-practice items.
- Added child-friendly bilingual controls and explanations across pages 47–53.
- Fixed overlapping playback and made reset stop both activity state and audio.
- Preserved offline operation and kept the submission lightweight and easy to run.
- Completed iterative human listening and browser tests on the final demo.
- Produced a clean, documented GitHub repository for independent evaluation.
What we learned
The project reinforced that successful educational AI work depends on domain judgment as much as code. Automated tools can verify that an audio file exists and loads, but a teacher must still decide whether the pronunciation, timing, meaning, and instructional context are correct.
We also learned that accessibility begins with small interface decisions. A control can function perfectly and still fail a child if its label is not understandable. Bilingual labels, consistent layout, clear stopping behavior, and local audio all materially improved the learning experience.
Most importantly, the strongest workflow was not “AI builds everything.” It was an explicit division of responsibility: AI accelerated implementation and diagnosis, while the educator retained authority over language, pedagogy, and final acceptance.
What's next for LearnChinese AI
The next step is to turn the Lesson 6 prototype into a reusable lesson-production pipeline. Planned work includes:
- a structured content-authoring workflow for additional textbook lessons;
- teacher-facing tools for recording, segmenting, reviewing, and replacing audio;
- learner progress tracking that remains privacy-conscious and offline-capable;
- improved accessibility, responsive layouts, and tablet support;
- classroom pilots with children and teachers;
- optional AI-assisted feedback features that preserve teacher control and clearly separate generated guidance from authoritative curriculum content.
Log in or sign up for Devpost to join the conversation.