An interpreter
in your pocket

Each person speaks their own language into one phone. Ink 2 listens and notices when each of you stops talking. An LLM translates each turn, and then Sonic 3.6 says it out loud in the other language.

Deploy to Vercel View on GitHub

The desktop layout, English on the left and Spanish on the right, with the status, larger-text, Flip and End controls above

Screenshot. The template running locally.

Models: ink-2, sonic-3.6
Languages: Any two of English, Spanish, French, Hindi, Japanese
License: Apache-2.0
Stack: Next.js 16React 19AI SDKWeb Audioshadcn/ui
Cartesia products: Ink 2Sonic 3.6

What it does

The screen splits so each person reads the conversation in their own language, and one half can turn upside down for a phone lying flat between you. Under each translated turn, a small line shows the time to the first translated word and to the first audio.

Run it yourself

1
Install
Run the commands below with Node 22.22 or newer. pnpm works too. Fill .env.local from the table below. If a key is missing, the page says which.
npx degit cartesia-ai/cookbook/examples/pocket-interpreter pocket-interpreter cd pocket-interpreter # add CARTESIA_API_KEY, LLM_MODEL and the provider key cp .env.example .env.local npm install npm run dev
Environment variables
CARTESIA_API_KEY
play.cartesia.ai/keys
LLM_MODEL
provider/model, for example openai/gpt-5.4-mini
OPENAI_API_KEY
If LLM_MODEL starts with openai/
ANTHROPIC_API_KEY
If LLM_MODEL starts with anthropic/
AI_GATEWAY_API_KEY
For any other provider
APP_PASSWORD
Optional locally. Required on Vercel, or none to make the app public
TRUST_PROXY
Optional. 1 behind your own reverse proxy
Start talking
Open the URL that npm run dev prints. Pick the two languages and press Start talking. Then take turns speaking, with nothing to press in between. A phone needs HTTPS for the microphone, so deploy it or use a tunnel such as cloudflared to try it on one.

Build it with your coding agent

One prompt that tells your coding agent where to start and what to read first.

Prompt

Claude Code Cursor Codex

Build a two-person voice translation app with Cartesia. Start from https://github.com/cartesia-ai/cookbook/tree/main/examples/pocket-interpreter and follow its AGENTS.md for the commands, files and rules. The server mints a 60 second access token with only the stt and tts grants. The browser streams the microphone to ink-2 on /stt/turns/websocket with @cartesia/cartesia-js, sends each turn to an AI SDK route that replies with the language code and the translation, and speaks it with sonic-3.6 over the TTS WebSocket. Keep the mute gate that sends silence to Ink 2 while Sonic plays. If I don't have a Cartesia API key yet, follow https://www.cartesia.ai/agent-quickstart.md to get one.

How it works

MicSends silence to Ink 2 while a translation plays

Ink 2Live captions, then a signal when you stop talking

LLMNames the language, then translates

Sonic 3.6Speaks each sentence in a voice for the target language

ScreenEach half shows the conversation in its own language

Why it works this way

Translation starts before the turn ends

When you pause, Ink 2 sends an early guess that you may be done, and the page starts translating right away. If you were done, it uses that translation. If you keep talking, it cancels it and waits for the full sentence.

The LLM says who spoke

Ink 2 doesn't report which language it heard. So the translation reply starts with a language code on its own line. The page uses that code to pick the voice and the side of the screen.

The app mutes itself while it speaks

While a translation plays, the microphone sends silence to Ink 2. Otherwise the app would hear its own voice and translate it back, forever.

Make it yours

Use a different voice

The voice for each language is set in lib/languages.ts. Pick replacements from the Cartesia voice library, or use your own clone, made only with the speaker's permission.

Change how it translates

Edit lib/prompt.ts to make translations formal or casual. Keep the language code on the first line of the reply, because the app reads it to pick the voice.

Tune turn detection

Ink 2 decides when a turn ends. To make it wait longer or end sooner, set turn_end_timeout_ms when the app opens the speech connection. That call is in lib/audio/listener.ts.

Join the teams making the switch to Cartesia

Ranked #1

in Speech Arena leaderboard & Speech to Text leaderboard by Artificial Analysis

Questions

Is my Cartesia API key safe?

Yes. The key stays on the server. The browser gets a token that lasts 60 seconds and can only turn speech into text and text into speech.

Where does the conversation go?

Both people's speech goes to Cartesia, and the text of each turn goes to your LLM provider for translation. Tell the other person before you start.

Can both people talk at the same time?

No. There is one microphone and one turn at a time, so wait for a translation to finish playing. You don't press anything between turns, because Ink 2 can tell a pause from the end of a sentence.

Which languages does it support?

Any two of English, Spanish, French, Hindi and Japanese. Those are the languages Ink 2 can hear. Adding another language needs a different speech-to-text model.

Does it work on a phone?

The layout is built for a phone first. It was tested in desktop Chrome with a recorded voice as the microphone, not on a real phone.

Does it need headphones?

No. The app mutes its own microphone while it speaks and for a third of a second after. In a loud room, background voices can still start a stray turn.

Is there a step-by-step guide to this template?

Yes. The recipe walks through the code, the README covers setup, environment variables and deployment, and the Ink 2 turn events docs explain the Cartesia side.