May contain hallucinations!

Realtime voice translate

Minimal full-screen mobile voice translator.

Uses ElevenLabs Scribe v2 Realtime and OpenAI GPT-5.4-mini

This app and README was written using OpenAI Codex GPT-5.5.

What It Does

• Uses ElevenLabs Scribe v2 Realtime through the official client-side token flow.
• Keeps the ElevenLabs API key on the server.
• Uses VAD and partial transcripts.
• Detects the source language from the script of each partial transcript.
• Sends each partial transcript to OpenAI for translation without waiting for earlier partials.
• Uses gpt-5.4-mini with reasoning effort none for translation.
• Displays only the latest available translation for the latest segment.
• Bottom-aligns the full latest translation and clips older overflowing lines off-screen.
• Shows ??? if script-based language detection fails for the latest partial.
• Supports English/Hindi and English/Russian language pairs from a dropdown.
• Shows a Settings button during translation that disconnects Scribe and resets the current conversation.
• Clears the current conversation when the page is hidden, closed, or unloaded.
• Ignores stale translation responses after a reset.
• Requests a screen wake lock while translating when the browser supports it.
• Optimizes the display for mobile use, with a horizontal phone recommendation and Add to Home Screen note for fullscreen mode.

Run

For local/manual testing after dependencies and tokens are present:

npm install
npm run build
source /root/all_tokens.sh
HTTP_PORT=80 HTTPS_PORT=443 \
TLS_CERT=/etc/letsencrypt/live/translate.samuelshadrach.com/fullchain.pem \
TLS_KEY=/etc/letsencrypt/live/translate.samuelshadrach.com/privkey.pem \
node server.js

Fresh Server Deployment

These steps are intended to be enough for a new OpenAI Codex instance or human operator to host the app on a fresh Ubuntu/Debian server.

Assumptions:

• Domain: translate.samuelshadrach.com
• App directory: /root/realtime-voice-translate
• Public repo: https://github.com/samuel-da-shadrach/realtime-voice-translate.git
• Runtime user: root
• HTTP/HTTPS ports: 80 and 443
• TLS: Let's Encrypt certificate at /etc/letsencrypt/live/translate.samuelshadrach.com/

1. DNS

Create an A record:

translate.samuelshadrach.com -> SERVER_IPV4_ADDRESS

For initial Let's Encrypt setup, use DNS-only mode in Cloudflare if possible. After the certificate works, Cloudflare proxying can be enabled with SSL/TLS mode set to Full or Full strict.

2. System Packages

Install Node.js, npm, git, and certbot:

apt update
apt install -y nodejs npm git certbot

If the distro Node.js is too old for Vite/React tooling, install a current LTS Node.js release and then rerun npm install.

3. Clone

Clone this repo into the path expected by the included service files:

cd /root
git clone https://github.com/samuel-da-shadrach/realtime-voice-translate.git realtime-voice-translate
cd /root/realtime-voice-translate

If you clone elsewhere, update both start.sh and realtime-voice-translate.service.

4. Tokens

Create /root/all_tokens.sh:

cat >/root/all_tokens.sh <<'EOF'
export ELEVENLABS_API_KEY='replace-with-elevenlabs-api-key'
export OPENAI_API_KEY='replace-with-openai-api-key'
EOF
chmod 600 /root/all_tokens.sh

Do not commit this file. It is intentionally outside the repo.

5. Build

cd /root/realtime-voice-translate
npm install
npm run build

6. TLS Certificate

Make sure nothing is already listening on port 80, then issue the certificate:

certbot certonly --standalone -d translate.samuelshadrach.com

Expected files:

/etc/letsencrypt/live/translate.samuelshadrach.com/fullchain.pem
/etc/letsencrypt/live/translate.samuelshadrach.com/privkey.pem

The app can run HTTP-only if these files are absent, but microphone access in browsers requires HTTPS for normal use.

7. Install Service

cd /root/realtime-voice-translate
chmod +x start.sh
cp realtime-voice-translate.service /etc/systemd/system/realtime-voice-translate.service
systemctl daemon-reload
systemctl enable realtime-voice-translate.service
systemctl restart realtime-voice-translate.service

8. Verify

Check service status:

systemctl status realtime-voice-translate.service

Check logs:

journalctl -u realtime-voice-translate.service -f

Check health:

curl -sS http://127.0.0.1/health
curl -sS https://translate.samuelshadrach.com/health

Open:

https://translate.samuelshadrach.com

9. Firewall

Ensure inbound TCP 80 and 443 are allowed by the server firewall and any cloud firewall.

10. Common Failure Points

• ELEVENLABS_API_KEY is not set: /root/all_tokens.sh is missing, unreadable, or does not export ELEVENLABS_API_KEY.
• OPENAI_API_KEY is not set: /root/all_tokens.sh is missing, unreadable, or does not export OPENAI_API_KEY.
• Browser microphone does not work: use HTTPS, not plain HTTP.
• HTTPS does not start: check the Let's Encrypt cert/key paths in realtime-voice-translate.service.
• Service starts but domain fails: check DNS, Cloudflare SSL mode, firewall, and whether ports 80/443 are already in use.
• Certbot standalone fails: stop anything using port 80, verify DNS points to this server, and rerun certbot.

Service

The included realtime-voice-translate.service is the systemd unit used on the server.