yt2anki

Turn a YouTube video into an Anki deck for learning a language. Each card is one sentence: the source-language sentence plus its audio clip on the front, and a translation into your language on the back. Each card also gets a short grammar/usage note on the back (the key grammar point plus the dictionary form of an inflected verb/adjective/adverb). Japanese sources additionally get furigana readings over the kanji, and with --color-words the matching words in the source and translation are shown in the same color. Optionally (--vocab) it also builds a separate vocabulary deck of the unique content words, each with its reading, meaning, and the example sentence it came from.

It's a single Python script that chains together a few command-line tools: download the video, transcribe the speech, split it into sentences, translate them, and cut the audio into per-sentence clips.

What you need

The script calls a few external command-line tools (it checks for them on startup and tells you how to install any that are missing). Most are Python packages installed from requirements.txt into the venv; only ffmpeg is a separate system binary:

Tool	What it does	Install
genanki	builds the .apkg deck	requirements.txt
yt-dlp	downloads the video & audio	requirements.txt
whisper	transcribes the speech with word-level timing	requirements.txt (pulls in torch — large)
ffmpeg	cuts the per-sentence audio clips	apt install ffmpeg / brew install ffmpeg

yt-dlp and whisper are found in the venv automatically — no need to activate it or install them globally. ffmpeg stays a system dependency (it isn't pip-installable, and whisper needs it at runtime too).

You also need an Anthropic API key for the translation step.

Setup

1. Install the Python dependencies in a virtual environment (this includes whisper's torch, so it's a large download the first time):

   python3 -m venv .venv
   .venv/bin/pip install -r requirements.txt

   Then install ffmpeg via your system package manager (apt install ffmpeg / brew install ffmpeg).

2. Put your API key in a .env file next to the script:

   ANTHROPIC_API_KEY=sk-ant-...
   

   (.env is gitignored. You can also just export the variable in your shell — an exported value takes precedence over .env.)

Usage

Run with the venv's Python. --source-lang (the spoken language of the video) is required; --user-lang defaults to en:

.venv/bin/python yt2anki.py --source-lang ja "https://www.youtube.com/watch?v=..."

It runs straight through the five steps. The deck is named after the YouTube video; the title is shown as an editable prompt so you can tweak it.

Handy options:

Option	Default	Meaning
--source-lang	required	spoken language of the video (e.g. ja, fr)
--user-lang	en	the language to translate into
--yes	off	run unattended: reuse existing output and skip the title prompt
--title	video title	name the deck yourself (skips the title prompt)
--clip-pad	0.50	seconds of audio padding added to each side of a clip
--color-words	off	color-code matching words across the source and translation (extra Claude pass)
--screenshots	off	add a video frame (at each sentence start) to the card front (larger .apkg)
--shot-offset	0.3	seconds after the sentence start to grab the screenshot (0 = exact start, risks scene-cut frames)
--vocab	off	also build a vocabulary deck (<video-id>.vocab.apkg) of unique content words (anthropic backend only)
--vocab-only	off	build only the vocabulary deck, skip the sentence deck (implies --vocab)
--model	small	Whisper model size (tiny/base/small/medium/large) — bigger is more accurate but slower
--translate-model	claude-sonnet-4-6	which Claude model translates
--workdir	./work	where intermediate files are saved

Everything for a video is saved under work/<video-id>/. If a step's output already exists, it asks whether to reuse it or recreate it, so it's safe to stop and start again. Choosing recreate at one step recreates every later step too; --yes always reuses without asking.

What you get

When it finishes you'll have, inside work/<video-id>/:

• <video-id>.apkg — a ready-to-import Anki deck with the audio bundled in (imports as youtube::<title>::sentences)
• <video-id>.vocab.apkg — (with --vocab) a separate vocabulary deck of the unique content words. Each card shows the word (dictionary form) on the front; the back adds the part of speech (for Japanese, the verb class u-/ru- or adjective class i-/na-), reading, meaning, a per-kanji meaning breakdown, and the example sentence with its audio. Words are deduplicated by dictionary form, keeping the first sentence each appeared in.

Importing into Anki

Double-click the .apkg, or in Anki choose File > Import and select it. The note type, cards, and audio are all included — no manual media copying needed.

How it works

The pipeline runs in five steps:

1. Download the video and extract the audio (yt-dlp).
2. Transcribe the audio with word-level timing (Whisper).
3. Split the transcript into one card per sentence (on sentence punctuation, falling back to Whisper's segment boundaries when none is present).
4. Translate each sentence into your language (Claude). The same call also writes a short grammar/usage note (and, for Japanese, furigana readings), and with --color-words it tags which source words match which translation words so they can be colored the same.
5. Build the deck — cut a tight audio clip per sentence (ffmpeg, padded by --clip-pad), optionally grab a video frame per sentence (--screenshots), and package the sentences, translations, clips, and frames into a single .apkg (genanki).

With --vocab, step 4 makes one extra Claude pass that extracts the content words (nouns/verbs/adjectives/adverbs) per sentence, and step 5 deduplicates them by dictionary form and packages them into a separate <video-id>.vocab.apkg, reusing each word's first-occurrence sentence clip for the audio. --vocab-only skips building the sentence deck. The vocabulary deck is anthropic-only; reading is filled for non-phonetic scripts (kana for Japanese) and left empty for languages already written phonetically.