Practice your work English out loud with a US coworker. Pay cents per practice.
AI: publik API or your key
publik verified: the founder reviewed this exact version for malware and approved it

Prices are for one typical use: one request of about 1,500 words sent and 375 words back. A higher quality score is better.
Not checked yet: publik has not read whether Dilo can run its AI on your computer.
| Model | Quality | Per 1,000 uses |
|---|---|---|
| On your computer (Ollama, 4-bit download size) | ||
| Granite 4.2 3B2.2 GB | 9 | $0 |
| Phi-4 Mini2.5 GB | 6 | $0 |
| Llama 3.1 8B4.9 GB | 7 | $0 |
| gpt-oss 20B14 GB | 9 | $0 |
| Gemma 4 31B20 GB | 19* | $0 |
| Qwen3.5 35B-A3B24 GB | 19* | $0 |
| publik API | ||
| publik-fastGLM-5.3 Flash · Dilo | 42 | $1.00 |
| publik-balancedMiMo-V2.6-Pro | 46 | $10.00 |
| publik-smartGPT-6 Sol | 48 | $18.00 |
Quality: Artificial Analysis Intelligence Index v4.3.2, read 2026-09-22 (publik API models 2026-09-25); * = estimated by Artificial Analysis. Sizes: the Ollama library, read 2026-09-22.

Practice your work English on a call with a native English speaker. Feedback on every phrase. Pay cents per practice, no subscription.
Dilo (Spanish for "say it") is a daily speaking practice for Spanish-speaking professionals who work or job-hunt in English. No subscription: you pay only the AI you use through publik, about $0.07 to $0.09 for a 10-minute practice. It focuses on work, where apps like Speak, Praktika and Cambly are general.
The interface comes in English, Spanish, Portuguese, French, Italian and German (a language menu at the top; the browser's language is picked the first time). The conversation is always in English. Each learner picks a level: Beginner (simpler English, slower voice, tips and translations in your language), Intermediate (tips in English, translations in your language) or Advanced (natural speed, English only). They also pick a practice partner: four photo avatars (AI-generated portraits of fictional people, made with Blotato), each with its own voice.
Requirements: Node.js 22.13 or newer (it uses the built-in node:sqlite), npm.
npm install
npm run dev
Open http://localhost:3000. With no environment variables the app runs in demo mode: the AI is a deterministic mock, the counterpart speaks with your browser's English voice, and the whole flow (practice, feedback, streak, phrases, history) works offline. A small "Demo mode: simulated AI" chip says so on the home screen.
Other commands:
| Command | What it does |
|---|---|
npm test | Run the test suite (Vitest) |
npm run test:coverage | Tests plus coverage report (thresholds: 80% lines, branches, functions, statements) |
npm run typecheck | TypeScript check |
npm run lint | ESLint |
npm run build / npm start | Production build and server |
To use the real AI, copy .env.example to .env.local and follow PUBLIK-SETUP.md. Never commit real values.
Browser (Next.js client components) Server (Next.js route handlers, Node runtime)
----------------------------------- ----------------------------------------------
MediaRecorder (push to talk) --audio--> POST /api/practice/:id/turn
1. /audio/transcriptions (whisper-1)
2. /chat/completions (publik-balanced) as the counterpart -> text shown at once
POST /api/practice/:id/speech { turnIndex }
/audio/speech (tts-1-hd, the avatar's voice), cached in the page
Audio element or speechSynthesis <--base64 mp3 or null (browser voice fallback)
POST /api/practice/:id/help ("Help me", publik-fast)
POST /api/practice/:id/finish
/chat/completions (publik-balanced, retry on publik-fast) -> JSON feedback
POST /api/phrases/:id/drill ("Say it": whisper-1 with word timings, clarity, pace, 1/3/7-day schedule)
POST /api/pronounce (sound check: CMU dictionary, no AI call)
GET /api/progress (streak, last 7 days, history, phrases, due today)
GET/POST /api/gate (Sign in with publik)
POST /api/telemetry (app_opened, active_session)
Layers (each file is small and has its own tests):
| Folder | Role |
|---|---|
src/domain/ | Pure logic: scenarios and industries, prompts, conversation rules, feedback parsing (zod), streak math, sound rules for Spanish speakers (sounds.ts) and clarity and pace (clarity.ts) |
src/server/ai/ | The AI port (AiProvider), an OpenAI-compatible client used for publik API and for a user's own key, and the offline mock |
src/server/publik/ | publik.ts (installs, wallet, gate words, telemetry; adapted from publik's reference module) and gate.ts (Sign in with publik per browser session) |
src/server/db/ | The Store interface and its SQLite implementation |
src/server/ | Config (validated env), errors with friendly messages, sealed keys (AES-256-GCM), session cookie, rate limit, use cases (practice-service.ts) |
src/app/api/ | Thin route handlers |
src/components/, src/lib/ | UI, typed API client, browser hooks (recorder, speaker, gate polling, telemetry) |
pk_) and a user's own OpenAI key are stored sealed with AES-256-GCM (APP_SECRET) and never sent to the browser.AI_PROVIDER=publik). Model tiers only (publik-fast, publik-balanced), never vendor model names, for publik.gate.ts gives each browser session its own install (POST /api/v1/installs), shows the same gate screens with publik's exact words, and sends app_opened and active_session with that install's key. This adaptation is documented as an open question for publik in STATUS.md.src/server/errors.ts) to friendly messages. A 402 from publik shows its message and exactly one link (the top-up URL), as publik asks.node:sqlite)Store interface, so moving to a hosted database (for example libSQL/Turso or Postgres) means one new file, not a rewrite.Trade-off: it needs a server with a persistent disk (a container with a volume or a small VM). Serverless hosts without a disk would lose the data. See the deploy notes in PUBLIK-SETUP.md.
v1 is push to talk: record, send, transcribe (whisper-1, files up to about 4 MB, so recordings stop at 90 seconds), reply, speak.
v2 (documented, not built): publik's realtime voice over wss://realtime.publikhq.com/v1/realtime (OpenAI Realtime events, Gemini Live, 16-bit mono PCM at 24 kHz). publik's docs say a browser cannot set the Authorization header and must send the key in a subprotocol, "only with a key that belongs to the person at the browser". To keep keys on the server, v2 should relay the socket through our server (browser to our WebSocket, our server to publik with the install key in the header). publik lists the price at about $0.0225 per minute where both sides talk, and a session holds $0.225 of balance when it opens.
Production runs on Vercel with a Turso (libSQL) database: https://dilo-mocha.vercel.app, listed on publik at https://publikhq.com/dilo. Set DATABASE_URL and DATABASE_AUTH_TOKEN (or the Turso integration's TURSO_DATABASE_URL and TURSO_AUTH_TOKEN) and the app uses the hosted database instead of the local SQLite file. Steps and variables: PUBLIK-SETUP.md.
Palette, type, logo, voice and how the brand carries to publik and Instagram: BRAND.md. Price positioning and competitor prices: COMPETITIVE-ANALYSIS.md, section 6.
WCAG 2.1 AA target: color tokens checked for 4.5:1 contrast in light and dark mode, visible focus rings, labels on every field, a skip link, one h1 per screen, lang="en" on English text, icons from Lucide (no emoji icons), 44px+ touch targets, aria-live announcements during the conversation, space bar to talk, Escape to cancel, and prefers-reduced-motion respected. Mobile first from 375px.
Audio is sent for transcription and not stored. Transcripts, feedback and phrases are stored per browser session in the app's database so streaks and history work. publik routes AI requests to its providers (see ai_providers on publik's GET /wallet).
MIT. See LICENSE.
Pronunciations come from the CMU Pronouncing Dictionary through the cmu-pronouncing-dictionary package (ISC). Partner portraits are AI-generated images of fictional people.