A local-first, multi-language learning system — learn a language from your own real material. Cards · spaced repetition · word-pick card creation · built-in reader · AI tutor · local pronunciation coaching
It doesn't teach you a language. It trains you on your own material.
Mainstream language apps teach their curriculum. langlab trains you on yours — work contracts, industry memos, foreign series, ebooks. Select a word to make a card, cards get scheduled automatically for review, and a virtual tutor listens to you read and points out where you stumbled.
All data stays on your own machine: nothing uploaded, no subscription, no lock-in.
Aloysius Luo — full-stack AI development and server operations.
- GitHub: @tttworks
- Contact: [email protected]
- Issues and PRs welcome (please read CLA.md before contributing)
A change in my work meant I had to bring my English up to the level of reading professional documents and sitting in negotiations — within six months.
I did it the old way and memorised a few thousand words. Then came the frustrating part: in the meeting room, none of them would come when called. Later I recorded myself speaking and listened back — nearly half of the recording was silence; and the words I misspelled were almost all guessed from pronunciation. What I had stored was sound, not spelling.
That's when it clicked: my bottleneck wasn't "not knowing" — it was "not being able to retrieve".
And every tool on the market was solving the first problem. They teach you to order coffee and talk about the weather; the situations I actually needed weren't covered at all. More to the point, I wasn't short of material — I had plenty of real English documents from my own field. What I lacked was a system that turns that material into training — and because the material is confidential, it had to run on my own machine.
So langlab exists. Every design decision in it traces back to one of those specific sticking points.
📖 The full story: why langlab exists
- Professionals with a large body of real foreign-language material to absorb (legal / investment / medical / engineering …)
- People who care about data sovereignty and won't accept a cloud account
- Learners who already "know a bit" but can't retrieve it or speak it fluently
It is not for people starting a language from zero — build a foundation with Duolingo or Babbel first.
Every feature maps to a specific published finding. Every citation can be looked up.
| # | Feature | Theory | Author | Source |
|---|---|---|---|---|
| 1 | Material comes only from your own real texts | Comprehensible Input Hypothesis (i+1) | Stephen Krashen Professor Emeritus, USC |
Krashen, S. D. (1985). The Input Hypothesis: Issues and Implications. Longman. |
| 2 | Spaced-repetition scheduling for cards | Forgetting curve & spaced repetition | Hermann Ebbinghaus Piotr Woźniak |
Ebbinghaus, H. (1885). Über das Gedächtnis. Woźniak, P. A., & Gorzelańczyk, E. J. (1994). Optimization of Repetition Spacing in the Practice of Learning. Acta Neurobiologiae Experimentalis, 54(1), 59–62. |
| 3 | Review log recorded (preparing for FSRS) | Spaced-repetition scheduling optimisation | Ye Junyao / Su Jingyong / Cao Yilong | Ye, J., Su, J., & Cao, Y. (2022). A Stochastic Shortest Path Algorithm for Optimizing Spaced Repetition Scheduling. KDD '22, 4381–4390. |
| 4 | Production mode (Chinese → say it in English) | Testing effect / retrieval practice | Henry L. Roediger III Jeffrey D. Karpicke |
Roediger, H. L., & Karpicke, J. D. (2006). Test-Enhanced Learning: Taking Memory Tests Improves Long-Term Retention. Psychological Science, 17(3), 249–255. |
| 5 | Framework retelling + recorded self-review | Output Hypothesis | Merrill Swain University of Toronto |
Swain, M. (1995). Three Functions of Output in Second Language Learning. In Principle and Practice in Applied Linguistics. Oxford University Press. |
| 6 | Phonetic transcription + syllable split per card | Dual Coding Theory | Allan Paivio University of Western Ontario |
Paivio, A. (1986). Mental Representations: A Dual Coding Approach. Oxford University Press. |
| 7 | Collocations + source sentence | Lexical Approach | Michael Lewis Paul Nation |
Lewis, M. (1993). The Lexical Approach. Language Teaching Publications. Nation, I. S. P. (2001). Learning Vocabulary in Another Language. Cambridge University Press. |
| 8 | Determiner highlighting / staged release | Cognitive Load Theory | John Sweller UNSW Sydney |
Sweller, J. (1988). Cognitive Load During Problem Solving: Effects on Learning. Cognitive Science, 12(2), 257–285. |
| 9 | Immediate, targeted feedback in pronunciation coaching | Deliberate Practice | K. Anders Ericsson | Ericsson, K. A., Krampe, R. T., & Tesch-Römer, C. (1993). The Role of Deliberate Practice in the Acquisition of Expert Performance. Psychological Review, 100(3), 363–406. |
| 10 | One course = one domain | Narrow reading | Stephen Krashen | Krashen, S. (2004). The Case for Narrow Reading. Language Magazine, 3(5), 17–19. |
| 11 | Select a word → card, immediately | — | Common practice in immersion-learning communities (AJATT / Antimoon) |
⚠️ Item 11 is flagged deliberately: sentence mining is community practice, not backed by a paper. Folding it into "theoretical grounding" would be dishonest — but it works, so it stays, clearly marked.
| Course apps (Duolingo et al.) |
Flashcard tools (Anki et al.) |
langlab | |
|---|---|---|---|
| Content source | The platform's generic curriculum | Cards you build yourself | Your own real material |
| Memory scheduling | Weak | Strongest | Yes (SM-2, FSRS data already recorded) |
| Production training | Limited | Depends on how you make cards | Built-in production mode |
| Speech feedback | Basic recognition | None | Local pronunciation coaching |
| Reading integration | None | None | Built-in reader + word-pick cards |
| Data ownership | Cloud account | Local | Local |
| Setup cost | Low (open and learn) | High (assemble your own toolchain) | Medium (bring your own material) |
In one line each:
- Course apps solve "I don't know what to learn" — they give you content;
- Flashcard tools solve "I need to memorise this" — but you build the cards and find the material;
- langlab solves "I have a pile of material to get through" — it chains reading, card creation, review and pronunciation feedback into one flow.
⚠️ To be clear: the above is design rationale, not a guarantee that using it will make you fluent. There is no silver bullet in language acquisition.
Six card types (word / term / chunk / sentence pattern / concept / annotation); both receptive and productive modes; shuffle, random draw, read-aloud, mark-as-mastered.
EPUB / PDF rendered in their original layout, or an intensive reading mode rendered sentence by sentence with a parallel translation. When you hit something you don't know: select it → one click makes a card, automatically carrying its source and context.
Material can be uploaded as files, registered by local path, pasted as text, or fetched from a URL:
Multi-role conversation (English / Japanese / Thai), with the avatar switching per language; session history is saved automatically so you can pick up where you left off. More importantly — it actually listens to you read: the recording is analysed locally with Whisper, which points out where you mispronounced, slurred or stalled, then gives one targeted piece of feedback. The audio never leaves your machine.
Japanese · Aoi
Thai · Ploy
Core metrics, stage-by-stage comparison, and a 182-day heatmap.
General dictionary and domain-terminology dictionary (one word base, two views); material supports file upload, local path registration, pasted text and URL fetching.
Requirements: PHP 8.2+, Node 18+; (optional) Python 3.11+ for pronunciation coaching and speech synthesis.
# 1) Backend dependencies
cd backend && composer install
# 2) Configuration (every key is optional — it runs with none of them filled in)
cp data/.env.example data/.env
# 3) Database — these three steps must run in this order
php scripts/init.php # create tables (⚠️ drops and recreates — only on an empty DB)
php scripts/migrate.php # add columns introduced later (idempotent, safe to re-run)
php scripts/seed_demo.php # optional: import demo cards to see it working
# 4) Frontend
cd ../frontend && npm install && npm run devThen open http://localhost:5174.
On Windows you can also just double-click
start.batin the repository root (requiresphpandnpmto be on PATH).
The repository ships with no corpus — you start with an empty database and fill it yourself.
There are four ways in, all on the Material library (#/materials) page:
- Open Material library → click "Upload files", or drag files onto the page
- Supports EPUB / PDF / DOCX / TXT / SRT / subtitle scripts
- Choose the course and language, then submit
- Text is extracted, segmented and indexed automatically (large files are chunked, with visible progress)
Useful when you have many books and don't want two copies on disk.
- Enter an absolute path (a file or a whole directory) under "Register local path"
- Only the path is registered — your original files are neither copied nor modified
- Directories are scanned recursively and grouped by course
For fragments — a paragraph from an email, a news item — just paste it in.
Enter a URL and the body text is extracted with tags stripped. If it fails, you'll be prompted to use "Paste" instead (many sites block scraping).
- Read: open from the material library → intensive mode renders sentence by sentence, with optional parallel translation and read-aloud
- Make cards: select any text → "Save and create card" in the popover, optionally letting AI fill in definitions and phonetics
- Review: Card learner → "Due today", scheduled by SM-2
- Film & series: importing
SRT + videotogether creates a three-level browser: category → title·season → episode
⚠️ Only import material you have the right to use. Users are responsible for the legality of the content they import.
backend/data/demo_cards.json holds 31 hand-written sample cards
(19 contract sentence skeletons + 12 legal concepts), so that anyone who clones the repo can see
it working immediately.
- Derived from no real contract, client file or commercial document; contains no institution names, amounts or commercial arrangements
- Licensed CC0 (public domain) — use it, change it, no strings
- Import with
php scripts/seed_demo.php(idempotent, safe to re-run) - After import the cards appear under the course "Contract English (sample)"
Want to swap in your own? Edit demo_cards.json directly, or load real material using the
methods above.
- No curriculum — it doesn't provide content from zero to fluency; you bring the material
- No cloud sync — data lives on your machine; using it in several places is up to you
- Collects nothing — no telemetry, no account, no phone-home
Except for the AI conversation feature, everything works offline (cards, reading, review and pronunciation coaching all run locally).
| Subject | License |
|---|---|
| Code | MIT |
| Documentation | CC BY 4.0 |
| Demo data | CC0 (public domain) |
| Character assets | With the code, MIT |
Third-party component licences are listed in THIRD-PARTY-NOTICES.md. Please read CLA.md before contributing code.