Capture
Screenshots, source windows, Steam Deck overlays, browser text, subtitles, articles, lyrics, and scripts.
2025-present / Acorn Talk
Acorn Talk
A Japanese learning project I started in 2025. The goal is to make native material less frustrating: capture what you want to understand, save the words that matter, and review them with the context still attached.
Why I kept going
A personal project that stayed interesting
Acorn Talk is the first project I have worked on mostly by myself for this long. That has been surprisingly fun. I am building something I actually want as a Japanese learner, so the motivation is not abstract.
I was using WaniKani to learn kanji and vocabulary, and it was useful. But I was still frustrated by how hard it was to approach native material. Games, manga, shows, songs, and websites were the reason I wanted to learn, but getting through a single page or scene could feel like a slog.
I also had a lot of ideas about how recent AI tools could make language learning better. Not one giant idea, really. More like dozens of small ideas: better lookup, better audio, better context, better review, and less friction between the thing I wanted to read and the words I needed to learn.
The native material gap
One early measurement stuck with me. The first episode of Spy x Family had roughly 600 unique words. Out of about 2,000 words I knew from textbooks and WaniKani-style study, only about 100 showed up in that episode. The generic path was not obviously taking me where I wanted to go.
The Loop
Native content becomes study material
The workflow I keep coming back to is: find Japanese you care about, inspect it, save the useful pieces, and review them later. The hard part is making that feel light enough that you keep doing it.
Understand
OCR, Japanese tokenization, deconjugation, dictionary matching, LLM-assisted context, translation, and ambiguity recovery.
Save
Selected words keep their sentence, screenshot, reading, sense, audio state, source pack, and occurrence metadata.
Review
Tango materializes notes, items, contexts, and cards so learners build quick kanji and word recognition.
Tango
The memory layer
Tango is organized around packs. A pack might be a game, a manga volume, an anime episode, a song, an article, or a pile of Lens captures. The pack keeps motivation local: study this because it helps you understand that.
Review sessions train kanji, readings, words, and context. The goal is not to replace native material with flashcards. The goal is to make coming back to native material easier.
Lens
The capture layer
Lens began as the missing piece for Japanese games on Steam Deck. I wanted to stay in the game, inspect a line, hear pronunciation, save a word, and keep playing.
Later, Desktop Lens brought the same idea to any window on Windows, macOS, and Linux. Lens can analyze a lot of text, but only the words you choose become study material.
Product Surface
A few parts of the product
The product grew into a few related surfaces. They are all trying to keep the same loop intact: understand something real, save the useful parts, and make review feel connected to the source.
Create
Paste or upload Japanese text, inspect detected words line by line, then save the pieces worth reviewing.
Review
Study sessions keep readings, context sentences, kanji breakdowns, and audio close to the answer.
Audio
Native speaker recordings are preferred, synthetic fallback is labeled, and bad generated audio can be detected and replaced.
Engineering
A few practical constraints shaped the system
The original Lens pipeline was designed for batch screenshots: OCR, LLM classification, token refinement, JMDict matching, disambiguation, extraction writing, pack inclusion, and materialization. It was useful, but it was too slow for a desktop experience where a dictionary popup has to appear while the pointer moves.
The desktop direction moved the model out of the hot path. OCR runs once, character boxes and offsets support hit testing, a local dictionary artifact handles fast lookup, and the server only materializes the words the user saves.
Screenshot or text
Window capture, Steam Deck overlay, subtitles, lyrics, scripts, articles, or manual text.
OCR and Japanese NLP
Vision OCR, bounding boxes, Sudachi, dictionary forms, deconjugation, and ambiguity handling.
Words in context
Surface forms, selected senses, readings, sentence context, screenshots, audio, and source packs.
SRS materialization
Notes, items, contexts, kanji cards, word cards, review history, and pack-specific progress.
Milestones
Some milestones
The motivation problem
Generic study lists were not getting me closer to the specific Japanese I wanted to read, watch, and play.
Tango takes shape
Packs became the core abstraction: a way to keep vocabulary, kanji, sentences, audio, and progress tied to one real source.
Open beta and Lens for Steam Deck
Tango opened to everyone, and Lens gave Japanese games a read-and-save overlay on Steam Deck.
Lens for Desktop
Lens expanded to Windows, macOS, and Linux with hotkey capture, OCR, dictionary lookup, audio, and saving into Tango.
One year in
The project had grown from personal study tooling into a product suite, a corpus, a desktop app, a Steam Deck plugin, and a learning workflow I kept using.
Gallery / Memories
Pieces of the Acorn Talk system
A few pieces of the product surface: desktop lookup, Steam Deck lookup, Tango study, kanji feedback, audio quality work, and the gallery where captures become packs.
Media preview
Lens desktop gallery
LensCaptures are organized into packs so a game, manga volume, show, or article can become a study project.
Gallery item 1 of 8