Voice-First AI Tutor

The language tutor
that learns from your voice

Invisible spaced repetition. Semantic knowledge graphs. Goal-seeking agents.
No flashcards. No tapping. Just conversation.

6 languages · 100% local option · Privacy-first

Three steps. Zero buttons.

01
🗣️

You speak

Converse naturally with a personality tutor — a Moscow intellectual for Russian, a Madrid journalist for Spanish. Each language has its own character and cultural depth.

02
🔍

It listens

Every word you say is analyzed in real-time — morphology, grammar, pronunciation confidence, response latency. No buttons. The system grades you from your voice.

03
🧠

It adapts

The Supervisor picks your next goal — remediation for struggles, new vocabulary, or grammar practice. FSRS schedules reviews. Semantic ripple reinforces connected words.

New Never seen
Learning S < 3 days
Review S +3 days
Relearning Lapsed
1 Again Relearning
2 Hard Short interval
3 Good Review
4 Easy Long interval
FSRS Engine

Spaced repetition that listens

Traditional SRS makes you tap "Hard, Good, Easy." LingLang grades you from your voice — pronunciation confidence, response latency, escalation level. The FSRS v5 algorithm maps these to stability and difficulty scores, scheduling reviews before you forget.

4Grade levels
90%Target retention
0Button taps
Semantic Ripple

Learning one word boosts the next

Every word has a 1024-dimensional embedding. When you master "яблоко" (apple), the system finds its 5 nearest semantic neighbors via pgvector — "груша" (pear), "фрукт" (fruit) — and applies a micro-boost to their stability. When you struggle, the ripple penalizes related words. Your knowledge graph heals itself.

1024Dimensions
5Neighbors
HNSWIndex type
👤
You speak
Processor
Analyzes every word
🎯
Supervisor
Sets goals, adapts plan
🗣️
Tutor
Converses naturally
Every 3-5 turns → SRS update → Goal check → Plan refresh
Goal-Seeking Agents

A tutor that plans for you

The Supervisor runs on a timer and signal accumulator. When you struggle, it sets a remediation goal. When you've mastered the current unit, it introduces the next vocabulary. The plan updates in real-time, injected into the tutor's context mid-conversation. You never choose what to study next.

3Goal types
Real-timePlan updates
0Manual choices
🔒

Privacy-first

Run the entire stack locally — STT, LLM, TTS — or use cloud services. Your conversation data stays on your machine.

🏠

Local AI stack

Ollama for inference, Qwen3-ASR for speech-to-text, MossTTS for synthesis. All run on your GPU. No data leaves your network.

🧬

pgvector knowledge

Vocabulary and progress live in PostgreSQL with pgvector HNSW indexes. Semantic search, ripple effects, all in your database.

Six languages.
Six personalities.

Each tutor is a character, not a chatbot. They have opinions, humor, and cultural depth.

🇷🇺

Russian

Маша — Moscow intellectual

"Ну вот, русский. Привет!"

80/20 immersive
🇪🇸

Spanish

Marta — Madrid journalist

"¡Hola! ¿Qué tal?"

80/20 immersive
🇫🇷

French

Professeur cultivé

"Bonjour — on commence?"

75/25 immersive
🇵🇹

Portuguese

European Portuguese

"Olá — vamos lá!"

80/20 immersive
🇸🇦

Arabic

Patient guide

"أهلاً! Ready to learn?"

70/30 assisted
🇬🇧

English

Power Vocabulary — advanced

"All right. Say one sentence."

100/0 immersive

Ready to talk?

The most advanced language tutor on the planet is waiting.