anki_ai

Autogenerated Anki Decks for Language Introduction

A small, scriptable pipeline that generates beginner-friendly sentence cards with audio and exports them as an Anki deck. Example sentences are composed by LLM, while the information about each words in each card is based primarily on Wiktionary data, curated and adjusted via LLM. The current repo targets Tagalog, but the same process can be quickly retargeted to other languages with an LLM edit pass. For updating any single script to another language, you can just ask an AI (e.g., DeepSeek), to revise each given script, see e.g. L2change.md for a prompt.

What the pipeline does

The pipeline (1) uses gen_lists.txt with an LLM to produce frequency-based word buckets (WORD_BUCKETS) and a curated set of beginner grammar points (GRAMMAR_POINTS); (2) gen_sentence.py composes, based largely on Wiktionary data, validated example sentences (ensuring coverage of target words/points), adds short English notes, and writes a timestamped JSON; (3) gen_voice.py synthesizes per-item WAV audio via system TTS; and (4) gen_card.py builds an Anki .apkg with both directions (L2→EN and EN→L2), bundling audio and metadata.

Repo layout

• gen_lists.txt — LLM prompts to (re)build word/grammar lists (step 1).
• word_list.py, grammar_list.py — seed data generated by gen_lists.txt prompts.
• gen_sentence.py — generate json/tl_cards_*.json (step 2).
• gen_voice.py — synthesize WAV audio per card (step 3).
• gen_card.py — package the Anki deck from JSON + audio (step 4).
• L2change.md — example prompt for retargeting any single script to another language.

Requirements

• Python 3.9+
• Packages: genanki, requests, and an OpenAI-compatible client
• Windows + Balabolka (bal4web.exe) for TTS using Microsoft voices

Setup

pip install genanki requests openai

Provide an API key via OPENAI_API_KEY (env) or a local config.json. If needed, edit BAL4WEB and voice names in gen_voice.py.

Usage

1. Prepare lists (via gen_lists.txt) Use gen_lists.txt to run the three prompts and create:

• a ~500-item frequency word list and WORD_BUCKETS (greetings/expressions/grammar_drills),
• 40–80 grammar points (GRAMMAR_POINTS module). Save to word_list.py / grammar_list.py. (The scripts later import these modules at runtime.)

1. Generate cards (JSON)

python gen_sentence.py

This script imports your word_list.py / grammar_list.py, calls an OpenAI-compatible chat endpoint, and writes a timestamped JSON like json/tl_cards_YYYYMMDD_HHMM.json plus a small usage summary. A partial checkpoint JSON file is written during generation. For convenience, you can point later steps at the current file:

cp json/tl_cards_YYYYMMDD_HHMM.json json/tl_cards.json

Notes:

• The script expects WORD_BUCKETS and GRAMMAR_POINTS to be defined; it validates outputs and enforces that targets actually appear in the Tagalog sentences.
• There’s a small built-in Tagalog detector and Wiktionary lookups to enrich card metadata (IPA, word breakdown).
• TEST_MODE defaults to True to produce a tiny sample; set it to False for full generation.
• API usage is recorded to json/openai_usage_summary_*.json.

1. Synthesize audio

python gen_voice.py

Audio is currently generated via Balabolka’s bal4web.exe using Microsoft voices (Tagalog fil-PH), probing for Blessica and Angelo and saving one WAV per card per voice as readings/{id}_Blessica.wav and/or readings/{id}_Angelo.wav. This can be rather straightforwardly to modify to other TTS services and voices.

1. Build the Anki deck

python gen_card.py

This reads json/tl_cards.json, looks for matching audio in readings/, and creates an .apkg named like AITagalogIntro_YYYYMMDD_HHMM.apkg. Cards are added in both directions per 30-item block, tagged v1 (TL→EN) and v2 (EN→TL), and include dual-audio fields if present. Import into Anki via File → Import….

More details & caveats

• LLM stage (gen_sentence.py)

  • Loads config from config.json or OPENAI_API_KEY.
  • Imports word_list.py/grammar_list.py (must exist and define the expected symbols).
  • For each target, it prompts the LLM to produce a concise Tagalog sentence + English gloss, verifies the target’s presence, and writes structured JSON entries with optional IPA/word-breakdown from Wiktionary.
  • If you see only a tiny sample, disable TEST_MODE.

• Voice stage (gen_voice.py)

  • Windows-only as written (depends on Balabolka + Microsoft voices).
  • Probes voice names and synthesizes to readings/. You might need to adjust BAL4WEB, voice labels, or LANG_CODE. If your system doesn’t have Tagalog voices, either skip audio (the deck still builds) or switch to another TTS provider.

• Deck stage (gen_card.py)

  • Expects the non-timestamped json/tl_cards.json; if you generated a timestamped file, copy it to that path first.
  • Looks specifically for {id}_Angelo.wav / {id}_Blessica.wav (filenames matter). Missing audio is tolerated—the note is created without that channel.
  • Builds both directions and includes metadata (IPA, breakdown, stage, grammar id/title) on the back.

• Other languages / services

  • For LLM retargeting, re-run step 1’s prompts for your target language, then repeat steps 2–4.
  • For TTS on macOS/Linux or where Microsoft Tagalog voices aren’t available, swap out gen_voice.py for another service (e.g., Azure TTS, ElevenLabs, Coqui-TTS, gTTS) and keep the output naming convention (readings/{id}_VoiceName.wav). The deck builder will pick up any files that match those names.

Notes

• This repo is not actively maintained.
• Upload the script and ask any AI to explain more details.

License

GPL.

This site is open source. Improve this page.