I have an Anki deck of 1,247 pronunciation cards. It took me approximately 200 hours to build with Forvo, copy-paste, and manual IPA lookup. I started it three years ago.
Last week I built another deck of 95 cards in ten minutes. Better quality, fully automated, six card types, all calibrated to my CEFR level. This is how.
The two problems with manual pronunciation decks
If you've tried building one, you know the pain:
Problem 1: IPA lookup is slow. Wiktionary first, dictionary second, Forvo for audio, copy-paste each segment into Anki. Twenty seconds per word becomes ten minutes for thirty words.
Problem 2: One-card-type decks underperform. Most learners build "Word on front, translation + IPA on back." That tests recognition. It doesn't test production, sentence usage, or sound discrimination. After a year you can read the IPA but can't say the word.
A good pronunciation deck needs at least three card types per word. Building three cards manually means another three-fold slowdown.
What changed
IPAtics generates Anki cards directly from selected text. You highlight a word in any app, press Alt+Q, and the overlay gives you the IPA. From there, one click sends six card types to your running Anki app via AnkiConnect.
Six card types per word, AI-generated, CEFR-calibrated, language-specific grammar included. Ten seconds per word. No copy-paste.
Download IPAtics free for macOS and Windows to follow along.
The six card types
Each card targets a different recall axis:
| Card type | Front (you see) | Back (you recall) | What it trains | |---|---|---|---| | Vocabulary | Translation in your native language | The target word, its IPA, grammar, and an example sentence | Productive recall from meaning. | | Production | The target word + IPA | Translation, grammar, usage in context | Confirming meaning from form. | | Sentence Cloze | A sentence with a blank | The missing word, its IPA, and grammar notes | Word usage in context. | | IPA Reading | Raw IPA: /baɪˈʃpiːl/ | The written word, grammar, meaning | Reading phonetic notation fluently. | | Sentence Translation | A full sentence in the target language | Its translation and grammar breakdown | Connected meaning extraction. | | Minimal Pair | Two similar IPA forms: /ʃɪp/ vs /ʃiːp/ | The two different words, meanings, translations | Sound discrimination. |
Each card type tests a different skill. A learner who can do all six knows the word in a way a single-card-type deck can't measure.
The 10-minute session structure
Here's what I do, replicated every time:
Minute 0-1: Pick the source text. A YouTube subtitle file, a chapter of the book I'm reading, a transcribed podcast, a news article. Volume matters less than coherence — words from one context are easier to remember together.
Minute 1-5: Select and add. For each word I want to learn:
- Highlight it in the source.
- Alt+Q to get IPA.
- Hit the Anki icon in the overlay.
- IPAtics generates all six card types and sends them to AnkiConnect.
I aim for 15-25 words per session. More than that and the cards start blurring together.
Minute 5-7: Quick review of generated cards. I switch to Anki and skim the card preview. About 5% need a manual tweak — usually the example sentence is too long or the translation is awkward. The rest ship as-is.
Minute 7-10: First-pass review. I review the new cards immediately to lock in the initial encoding. Anki's default new-card flow handles the rest from tomorrow on.
That's it. Two weeks of consistent sessions = 200-300 cards, six types each = a deck that actually moves the needle.
CEFR calibration
The most important thing AI-generated cards bring is level-appropriate example sentences.
A B2 learner doesn't benefit from a sentence with one new vocabulary word surrounded by C2 grammar they can't parse. The example sentence has to be slightly above the learner's level — comprehensible with effort, not impossible.
When you set your CEFR level once in IPAtics preferences, every generated card uses that as the upper bound for sentence complexity. Cards for an A2 learner reading German will get simpler grammar than cards for a C1 learner reading the same text.
This isn't possible with manual deck-building. You'd have to write level-calibrated sentences yourself. Nobody does this for 100+ words.
Language-specific grammar that auto-fills
For each word, the generated card includes grammar relevant to the target language:
- German: article (der/die/das), plural form, separable prefix if relevant.
- French: gender, plural form, irregular feminine if relevant.
- Spanish: gender, plural, irregular verb stems.
- Russian: gender, case endings for nouns, aspect pair for verbs.
- Turkish: vowel harmony pattern, possessive forms.
- Japanese: kanji + furigana, particle usage, pitch accent.
- Korean: hanja + hangul, formality level, irregular conjugation.
If you've ever built a Korean Anki card and tried to remember whether 읽다 takes ㅂ-irregular conjugation, you know what time this saves.
What about minimal pairs specifically
The Minimal Pair card type is the one most learners skip when they build manually, because finding good minimal pairs is hard. AI generation handles this — it picks two phonetically similar words that contrast on one phoneme, then puts both IPA forms on the front.
This is the card type that fixes English speakers' Spanish /r/ vs /rr/ confusion. Pero /ˈpe.ɾo/ vs perro /ˈpe.ro/. Both shown side by side, both required for the answer. Three weeks of these and your ear separates the tap from the trill.
When this workflow breaks
Two cases:
Rare or domain-specific words. AI-generated cards work best on common-to-mid-frequency vocabulary. If you're studying legal Russian or medical Spanish, the example sentences will sometimes miss context. Manual tweak: 10% of the time.
Languages with strong dialect variation. IPA transcriptions default to the standard variety. If you're learning a regional dialect, the IPA will be the standard form, which may differ. You'll know when this matters.
For everything else — language learners, exam prep, daily reading vocabulary — the workflow works.
The deck I built last week
95 cards, German B2 vocabulary from a news article I was reading. Six card types each = 570 review items.
Total time: 11 minutes (including the minute I spent fixing one weird translation).
Download IPAtics free for macOS and Windows, connect it to your running Anki app, and try one session. The compounding starts immediately. (Here's how the automatic IPA and audio cards work.)
Related reading: IELTS pronunciation: the IPA sounds examiners listen for · From PDF to fluent: the 2-second pronunciation workflow · The Anki + IPA pronunciation workflow · Minimal pairs in English