Sveiki
Latvian language learning platform built on the Birkenbihl method — a four-step approach that prioritizes understanding over memorisation: (1) Decoding — every word translated literally in source word order, creating a strange-but-readable interlinear text; (2) Active Listening — audio plays alongside the decoded text with word-level highlighting; (3) Passive Listening — the same audio loops in the background while you do other things; (4) further activities like shadowing and free production.
Content as data
The guiding architecture principle is content is data, code is a player. All lesson material lives as validated JSON under content/<languageCode>/ — no HTML, no layout, no colour decisions inside the content layer. Adding a new target language means adding a content folder; it never means rewriting application logic. The Zod schemas in src/lib/content/schema.ts are the single source of truth for what valid lesson data looks like, and CI enforces them on every push.
Running validation
npm install
npm run validate # checks schema, tokenization, and dictionary coverage for every lesson
npm run test # vitest unit tests for the validator
npm run lint # ESLint
npx tsc --noEmit # TypeScript type check
The validator reports a coloured PASS/FAIL per lesson and exits with code 1 on any failure, making it safe to use as a CI gate.
Dialogue audio
Each lesson is synthesized in one Text-to-Dialogue request so ElevenLabs sees
the complete conversation and can keep voices, pacing, and prosody coherent.
Generated source files live together under
content/<lang>/audio/; npm run sync-audio copies only
the lesson MP3 to the public directory used by the app.
npm run audio lv-a1-03 # regenerate the complete lesson take
npm run timings lv-a1-03 # review/correct sentence boundaries
npm run timings lv-a1-03 --insert-silence-after s1 --seconds 0.5
npm run sync-audio # copy lesson MP3s into public/audio
The generator refuses scripts above ElevenLabs' recommended 2,000-character
limit for a single reliable dialogue request. Timing corrections remain visible
after regeneration, but must be reviewed because they refer to the previous take.
Decoding view
The heart of the app is the lesson page at /lessons/<lessonId>. Each sentence is displayed interlinear-style: the Latvian word sits above its word-for-word gloss, pairs flowing left-to-right and wrapping as units on narrow screens so a gloss is never orphaned from its word.
Three reading modes — Decode (full interlinear), Natural (Latvian + idiomatic translation), Latvian only (target text alone for self-testing) — switch from a pill toggle at the top of the page. All three draw from the same loaded lesson data; no refetch on mode change. Tokens with grammatical notes reveal them on tap (mobile) or hover/focus (desktop) via an accessible disclosure. Natural translations are collapsed by default in Decode mode.
[Screenshot placeholder — to be added after native-speaker review]
Stack
Next.js 16 (App Router) · TypeScript · Tailwind CSS · Zod · deployed on Vercel
