# deckdrill

> Bring-your-own-keys studio that turns flashcards into Anki decks and long Pimsleur- and Glossika-style audio drills — optionally in your own cloned voice. Keys stay in the visitor's browser; a stateless relay forwards each request to one provider and stores nothing.

Website: https://www.deckdrill.app
Studio: https://www.deckdrill.app/studio · Your voice: https://www.deckdrill.app/voice · Demo: https://www.deckdrill.app/demo
MCP server (read-only, Streamable HTTP): https://www.deckdrill.app/api/mcp — tools `get_methodology`, `break_down_sentence`, `estimate_drill`, `build_drill_script`, `list_demo_decks`, `get_demo_cards`, `get_faq`.

## Drill styles

### Pimsleur-style
Each new card: English cue → the sentence → a backward build-up from the end (each chunk with a pause to repeat) → slow, normal, slow, normal. Then the card comes back as a recall prompt (English cue, a pause to say it yourself, the answer) 1, 4 and 10 cards later, and once more in a final recall pass.
Built on retrieval practice and spacing, the two most robust findings in memory research. Building up from the end keeps the sentence's natural final intonation on every repetition.

### Glossika-style
Two passes through the deck. Each card: English cue → the sentence → slow → the sentence again, each with a pause to repeat. No breakdown.
High volume of whole sentences at natural speed — good for listening fluency and rhythm once the pieces are familiar.

Pauses: Repeat pause = clip length + 0.3 s. Recall pause = 0.8 × answer length + 1.2 s. Slow versions play at 0.8× speed.

## Backchaining example

"Si hubiera vendido la granja aquel primer invierno, hoy estaría en la ciudad." → "en la ciudad" → "hoy estaría en la ciudad" → "la granja aquel primer invierno, hoy estaría en la ciudad" → full sentence

## Research

- Spacing: distributed review beats massed review (Cepeda et al., 2006 meta-analysis).
- Retrieval: trying to produce the answer beats re-hearing it (Roediger & Karpicke, 2006) — which is why the recall pause matters more than the repetition.
- Expanding vs. equal intervals is less settled than Pimsleur claimed (Karpicke & Roediger, 2007); what matters most is a short first delay and long delays eventually.
- Pure parroting without meaning transferred poorly in the old Audio-Lingual method — drill meaningful sentences with recall, not lists.

## Providers

- Card writing (visitor's key): Anthropic (default claude-sonnet-5-5), OpenAI (default gpt-5-mini), Google Gemini (default gemini-2.5-flash)
- Voices (visitor's key): ElevenLabs — Most natural voices, and the only option for cloning your own voice (Starter plan or higher). Azure Speech — Microsoft's neural voices (es-MX Dalia, Jorge…). Free tier covers about 500,000 characters a month.
- No key: System voices (Web Speech API) for live playback, and Anki's {{tts}} tag for decks — no key needed.

## Demo decks (A Wonderful Life, Mexican Spanish)

- Vocabulary: 100 cards
- Coulda Woulda Shoulda: 60 cards · audio previews https://www.deckdrill.app/demo/coulda-woulda-shoulda-pimsleur.mp3, https://www.deckdrill.app/demo/coulda-woulda-shoulda-glossika.mp3
- Quedar: 30 cards · audio previews https://www.deckdrill.app/demo/quedar-pimsleur.mp3, https://www.deckdrill.app/demo/quedar-glossika.mp3
- Recetas: 30 cards · audio previews https://www.deckdrill.app/demo/recetas-pimsleur.mp3, https://www.deckdrill.app/demo/recetas-glossika.mp3

Example: a Pimsleur-style drill of the 30-card Quedar deck runs about 16 minutes and needs about 1,188 characters of target-language speech.

## FAQ

### Do you store my API keys?

No. Keys stay in your browser (this tab only, unless you tick “remember on this device”). Each request sends the key to our stateless relay, which forwards it to that one provider and keeps nothing.

### Can I use it without any keys?

Yes. The demo decks and audio are free to browse. Drills can play live with your device's built-in voices, and exported decks can use Anki's own text-to-speech. Saving drill MP3s needs an ElevenLabs or Azure Speech key; writing new cards needs an Anthropic, OpenAI or Gemini key.

### How do I use my own voice?

On the Your voice page, record about two minutes of clear speech (or upload a file), confirm it's your voice, and it's cloned into your own ElevenLabs account (Starter plan or higher). Then pick it as the main speaker in the studio — for example on 75% of cards, with another voice on the rest.

### What does a drill cost?

You pay the TTS provider directly, per character. A 30-card deck of full sentences is roughly 4,000–7,000 characters for the target-language speaker, depending on sentence length, plus about a third of that for the English cues. Both styles share the same lines — each sentence, its slow version and its chunks are made once — and the studio shows the exact count before you render. Rendered lines are cached in your browser, so re-rendering or exporting the deck afterwards costs nothing extra.

### Will it work with my existing Anki decks?

Yes — import any .apkg (including the newer compressed format). The studio reads the front and the first line of the back of each note.

### Which languages?

Card generation and drills work for Spanish (Mexico), Spanish (US), Spanish (Spain), Portuguese (Brazil), French, Italian, German, Polish. English is the cue language.