Voice Translator

Speak in English (like a housekeeping preshift). The app translates the meaning into the language you pick (Spanish, French, Portuguese, Chinese, Arabic, Hindi, Filipino, Korean, Russian, Ukrainian or Turkish) and plays it through your Bluetooth speaker in your own cloned voice.

Live at: https://translate.kenndoesdev.com

How it works

1. Speak: your phone's built-in voice typing turns your English into text (free). You can fix any mistakes before translating.
2. Translate: Claude translates the meaning, not word for word, split into short paragraphs. Each paragraph has a "What this says in English" check, and anything that sounded misheard gets flagged.
3. Play: ElevenLabs speaks each paragraph in your cloned voice. The audio is saved on your phone, so replays cost nothing.

Until your ElevenLabs voice is added, the app plays the Spanish with your phone's built-in Spanish voice (free), so you can start using it right away.

Works in Chrome on Android and Safari on iPhone. You can add it to your home screen like an app.

Setup

1. Anthropic (translation)

• Log in at https://console.anthropic.com (or create a new account).
• Billing → add credits ($5 lasts a long time; each preshift costs a few cents at most).
• API Keys → Create key → copy it.

2. ElevenLabs (your voice)

You can add your voice right in the app: tap ⚙ → Your ElevenLabs voice, paste the Voice ID and API key, tap Save, then Test my voice. They're saved only on that phone. (Or put them in Vercel as below so every phone uses them.)

• Subscribe to Starter (Instant Voice Clone) or Creator (Professional Voice Clone, sounds more like you, more credits).
• Voices → Add voice → Instant Voice Clone: upload 1–3 minutes of clean recordings of you talking (quiet room, no music).
• Open your new voice → copy the Voice ID.
• Developers → API Keys → create a key → copy it.

3. Vercel

• Add New → Project → import this GitHub repo. Framework preset: Other. No build command needed.
• Settings → Environment Variables, add:

Name	Value
APP_PASSCODE	A passcode you choose (you'll type it once per phone)
ANTHROPIC_API_KEY	From step 1
ELEVENLABS_API_KEY	From step 2
ELEVENLABS_VOICE_ID	From step 2

Optional: ANTHROPIC_MODEL (default claude-sonnet-5-5), ELEVENLABS_MODEL (default eleven_multilingual_v2; eleven_flash_v2_5 uses about half the credits).

• Redeploy after adding variables.
• In the app, tap the ⚙ settings button to confirm everything shows ✓.

4. Move the subdomain (shoe → translate)

1. In Vercel, open the old project that uses shoe.kenndoesdev.com → Settings → Domains → remove it.
2. Open this project → Settings → Domains → add translate.kenndoesdev.com.
3. Vercel shows the DNS record it needs (usually a CNAME pointing to a vercel-dns address). In Squarespace → Domains → kenndoesdev.com → DNS:
   • Delete the shoe CNAME record.
   • Add a CNAME with host translate and the value Vercel shows.
4. Wait a few minutes until Vercel shows the domain as Valid.

Adding a language

Edit public/languages.js. Copy an entry and change the name, voice tag and the short guidance line (which dialect, formal or informal "you", etc.). Stick to languages ElevenLabs' multilingual model supports so your cloned voice can speak them.

Live rooms (Speak → Live room)

Your team listens on their own phones with earbuds and hears you in their language a few seconds after each pause.

1. On your phone: Speak → Live room → Start a room. You get a room code and a QR code.
2. Team members open the site, tap Listen (or scan the QR), pick their language, and tap the big button. The Listen screen shows its words in the language they pick.
3. Tap the mic and talk. Each pause sends what you just said; listeners hear it in your cloned voice (or their phone's voice if ElevenLabs isn't set up) with captions.
4. End room when you're done. Rooms also expire on their own after 4 hours.

Notes:

• Up to 4 languages per room. Each language costs its own ElevenLabs credits; the number of listeners doesn't matter.
• Live rooms use faster models: claude-haiku-5-5 and ElevenLabs eleven_flash_v2_5 (about half the credits). Change with ANTHROPIC_LIVE_MODEL / ELEVENLABS_LIVE_MODEL.
• Listeners must keep the page open with the screen on (the page tries to keep it awake).

Setup: Upstash Redis (free)

Live rooms need a small database to pass sentences to listeners.

1. Vercel → your translator project → Storage → Create Database (or Browse Marketplace) → Upstash → Redis.
2. Pick the free plan and a region near you (e.g. us-east-1), create it, and connect it to the translator project.
3. Vercel adds the connection settings automatically (KV_REST_API_URL / KV_REST_API_TOKEN or UPSTASH_REDIS_REST_URL / UPSTASH_REDIS_REST_TOKEN; either works).
4. Redeploy the project.

Editing the translation

• Housekeeping words: edit lib/glossary.js. Each term has a meaning (used for every language) and optional exact wording per language.
• Tone and style rules: edit buildSystemPrompt() in lib/translate.js.

Using it at preshift

• Some Android phones beep when voice typing starts. If the beep plays through the Bluetooth speaker, turn the speaker on after you finish speaking in English.
• Tap the big button, talk, tap again to stop. Fix anything misheard, then tap Translate.
• Glance at the yellow "double-check" box, if there is one, then tap Play in Spanish.
• If a preshift goes long, the app keeps listening. It restarts by itself if the browser pauses after silence.

Future: user accounts

All keys are read in one place, lib/config.js. To let other managers use their own ElevenLabs account, add sign-in and look up that user's saved key and voice ID there. Nothing else needs to change.

Development

npm test