The Spark

I have been passively learning Dutch for a few years. While I can read Dutch texts, listening comprehension remained a challenge—especially with fast speech or different accents. That’s the reality of language learning: real-world listening is key to fluent speaking.

I wanted to learn from movies I enjoy, not from generic content. Unable to find an app that fit this workflow cleanly, I built my own: an AI-powered Dutch learning partner that turns any movie into an interactive lesson. Each session is 5 minutes per clip—focused, progressive, and pressure-free. Progress only saves when you complete a session, which keeps you accountable.

This is designed for intermediate learners who want to improve listening and speaking in an engaging, authentic way.

---

The Blueprint

How do you transform a movie into a language course?

1. Input: A film (MKV) + its SRT subtitles
2. Pipeline: AI splits the movie, extracts vocabulary, generates quizzes and cloze tests, tracks progress
3. Output: Interactive Streamlit app—study vocab, watch clips, take quizzes, fill blanks, see mastery grow

Technical Architecture:

• Data Layer: JSON files for vocab, quizzes, and progress
• AI Layer: DeepSeek API for vocabulary extraction and quiz generation
• Automation Layer: FFmpeg for video clipping and subtitle conversion
• UI Layer: Streamlit orchestrates the workflow and tracks progress

The Flow:

┌─────────────┐
│   Prep      │  - Upload folder with .mkv + .srt
│             │  - Split into 5-min clips (FFmpeg)
│             │  - Generate vocab, quiz, cloze per segment (DeepSeek)
└──────┬──────┘
       │
       ▼
┌─────────────┐
│   Learn     │  1. Study: Review vocab (word, translation, example)
│             │  2. Watch: 5-min clip (subtitles OFF, toggle ON to rewatch)
│             │  3. Quiz: 10 MCQs (auto-generated, local grading)
└──────┬──────┘
       │
       ├── Score < 8/10 → Rewatch (subtitles ON) → Retry Quiz
       │
       ▼ Score ≥ 8/10
┌──────────────────┐
│ Semantic Mastery │  - Watch clip again
│  (Cloze Test)    │  - Fill 10 blanks from memory
│                  │  - 80% pass threshold
└──────┬───────────┘
       │
       ├── Score < 80% → Retry Cloze
       │
       ▼ Score ≥ 80%
┌─────────────┐
│  Next Clip  │  Progress advances
└─────────────┘

---

The Build

Constraint: Simple setup—just Python, FFmpeg, and an API key.

Tech Stack:

• Streamlit — UI and workflow orchestration
• DeepSeek API — Vocabulary extraction, quiz generation
• FFmpeg — Video clipping
• Python — Data handling, orchestration
• JSON — Progress and vocab storage

Visual Strategy:

• Native subtitle overlay (WebVTT)
• Segment-by-segment progress tracker
• Mastered/enhancement word lists

Lesson Creation Steps:

1. Clip movie into 5-min segments; convert subtitles to WebVTT
2. Extract vocab and generate quizzes via DeepSeek
3. Study vocab → watch clip (subtitles OFF/ON) → take quiz
4. Cloze test for semantic mastery
5. Track mastered words and completed segments

---

The MVP Release

Live Demo

---

Learning Retro

What Surprised Me

1. Real content drives motivation — Watching actual films made learning feel authentic and fun
2. AI handles the heavy lifting — DeepSeek extracts vocab and generates quizzes with minimal effort
3. Progress tracking keeps you accountable — Seeing mastered words and segments motivates continued learning
4. Simplicity wins — No complex setup, just Python, FFmpeg, and an API key

What Was Hard

1. Subtitle conversion — Ensuring SRT → WebVTT worked reliably for native overlays
2. Cloze data quality — Multi-word phrases sometimes didn’t appear verbatim in the transcript
3. Large video files — GitHub can’t host clips; only code and data are synced

Key Wins

✅ Built a real, usable language app
✅ Automated lesson creation from any film
✅ Progress tracking and mastery lists
✅ Low cost: LLM calls for a 90-min movie cost 0.25 RMB (€0.03)

---

Next Phase

Interested in expanding to other languages? Reach out and I can share the repo so you can continue building your own language trainer.