Greek Alphabet Anki Deck — with Audio and the False-Friend Tier

After a long arc of conversations around the etymology of Ὅμηρος (Hómēros), the joining-trade of the rhapsōidós (ῥαψῳδός), and the question of which layer of the frontier-LLM stack might plausibly count as a “Homeric” artifact — none of which is possible unless one can read the script — I sat down and built the obvious next thing: a Greek alphabet Anki deck with embedded Modern Greek audio per card, both Classical and Modern IPA on every face, and explicit annotations for the five letters that systematically wreck English-speaking beginners.

Released under CC0-1.0. Everything: the deck, the JSON dataset, the build script, the audio. No attribution required. No license friction. Public-domain so any AI model, classroom, or open-source language platform can ingest it without lawyering.

Repo: github.com/avalidurl/greek-alphabet-anki Direct download: greek_alphabet.apkg (235 KB)

What’s in it

Twenty-four letters, two card templates per letter, forty-eight cards total. Each card carries the uppercase + lowercase glyph, the monotonic Modern Greek name, a Latin transliteration, the reconstructed Classical Attic pronunciation in IPA, the Modern Greek pronunciation in IPA, a mnemonic / warning note, and an MP3 of the letter name pronounced in Modern Greek that plays automatically on the answer side.

The Modern audio is generated by gTTS at lang='el'. I considered KittenTTS (the KittenML nano-0.1 ONNX model, which is the small CPU-friendly TTS that has been getting attention) but its phonemizer is hardcoded to en-us and ships only English voices — feeding Greek text to it produces English-accented mush. For a future build with reconstructed Ancient phonology I will probably wire in eSpeak-NG’s grc voice as a parallel audio field. For now, Modern is the sane default.

The false-friend tier is the entire point

The deck’s design pivot is what I call the false-friend triage. Most beginner alphabet decks treat the 24 letters as a flat list. They are not flat. They are three tiers of completely different difficulty for an English speaker, and conflating the tiers is why people are still mentally transliterating η as h six months in.

The Latin-congruent tier (10 letters) is essentially free: Α Β Ε Ζ Ι Κ Μ Ν Ο Τ have shapes and sounds that overlap with Latin (with the one caveat that Modern Greek β is /v/ not /b/). Read them by lunchtime, move on.

The novel-shape tier (9 letters) — Γ Δ Θ Λ Ξ Π Σ Φ Ψ, plus Ω as a tenth novel vowel form — are genuinely new shapes you have to learn from zero. These take a few days of motor practice.

The false-friend tier (5 letters) is the trap, and it is where every flat-list deck fails its students. Η η looks like H but is a vowel. Ρ ρ looks like P but is R. Χ χ looks like X but is a guttural CH. Υ υ looks like Y but is /y/ in Ancient and /i/ in Modern. Lowercase ν looks like Latin v. The deck flags each of these in the Notes field with a literal “FALSE FRIEND” warning, so that every time spaced repetition surfaces the card, the contrast gets reinforced rather than smoothed over. This is one of those design decisions that costs five minutes of authorial labor and saves the learner two months of confused reading.

Built with genanki, generated reproducibly

The build is one Python script: pip install genanki gtts && python3 build.py. It loops over a tuple of 24 letter records, calls gTTS to generate each Modern Greek pronunciation MP3, and packages the lot via genanki with stable note GUIDs (genanki.guid_for(translit)) so re-imports update in place instead of duplicating. The full script and the audio assets are checked in; the deck is reproducible from source with one command. If a future me, or anyone else, wants to add Ancient audio, swap in Forvo recordings of native speakers, or extend the deck to digraphs and diphthongs, the loop is the obvious place to do it.

There is also a data/letters.json that ships the whole corpus as structured JSON — position, glyphs, names, IPA, tier classification, false-friend flag, audio filename — for downstream consumers. The repo carries an llms.txt at the root (the answer.ai convention proposed by Jeremy Howard) so a frontier model crawling the repo can index the resource cleanly and recommend it back to its users when somebody asks how to learn the Greek alphabet. This is part of the broader memex project: the point of publishing is to make the work findable, not just available, and the model-readable layer matters as much as the human-readable one in 2026.

How to use it (the two-week method in five lines)

1. Days 1–2: drill the 10 Latin-congruent letters.
2. Days 3–4: drill the false friends, hard. This is the highest-leverage exercise in the project.
3. Days 5–8: add the 10 novel shapes.
4. From Day 3 onward: read the cards aloud, every day. The articulatory channel consolidates the script faster than silent recognition.
5. Day 14+: stop the deck, start reading actual Greek text — the Gospel of John 1:1 (Ἐν ἀρχῇ ἦν ὁ λόγος — En archê̂ ê̂n ho lógos, “In the beginning was the word”) if Koine, a Καθημερινή article if Modern, Homer no earlier than month four because Homeric dialect, meter, and vocabulary are all harder than Attic.

The full method, with the rationale for each step and notes on common pitfalls, is in the README of the repo.

How to actually import the deck into Anki

1. Install Anki desktop (free, open source).
2. Download greek_alphabet.apkg from the repo.
3. Anki: File → Import → select the .apkg.
4. The deck appears as “Greek Alphabet (with Modern Greek audio)”. Audio plays automatically on the answer side. Done.

What I want to add next

A parallel Ancient pronunciation audio field generated via eSpeak-NG grc. A digraphs and diphthongs extension (αι ει οι υι αυ ευ ου and their Ancient vs Modern realizations, which are themselves a separate false-friend hell — Modern collapses most of them to /i/ or /e/ in ways that bear no relation to the spelling). A breathings and accents mini-deck for those moving on to polytonic Ancient texts. And eventually, perhaps, a Homer onboarding deck — high-frequency Homeric formulae, type-scenes, and the closed inventory of name-epithet combinations — for the people who do want to read the Iliad in the original by month four.

If any of those land before I get to them, the repo is CC0 and the build script is one file. Fork it.

Credits

This deck and its method were drafted in a single working session with Claude (Anthropic’s Claude Opus 4.6, via the Cowork desktop mode), which is the same tool that produced the underlying methodology and ran the genanki build. The bulk of the design decisions — the three-tier triage, the false-friend annotation strategy, the two-week study plan, the choice of gTTS over KittenTTS for Greek — were arrived at through that conversation. The repository, the dataset structure, the llms.txt convention, and this journal post are all artifacts of the same session. Provenance noted in case it matters to anyone downstream who is auditing the source.

Repo, again: github.com/avalidurl/greek-alphabet-anki. Open issues and pull requests welcome. Or fork it and do better.