Speech-to-Speech Translation MVP (Local)

A near-zero-cost speech translation system that runs entirely locally. Supports microphone input (for your voice) and system audio capture (for meetings).

Languages: English <-> Japanese Stack: FastAPI (Backend), Next.js (Frontend), Faster-Whisper (STT), NLLB (Translation), Piper (TTS).

Prerequisites

• Docker & Docker Compose
• Nvidia GPU (Optional, defaults to CPU)
• ~2GB of Disk Space for models

Quick Start

1. Start the application

   docker-compose up --build

   Note: First run will download models (~1-2GB), so be patient.

2. Access the UI Open http://localhost:3001

3. Usage

   • Mode Selection:
     • Microphone: Use this to translate your own voice.
     • System Audio: Use this to translate others in a meeting (Zoom/Teams). Browser will ask to share a tab/screen - select the meeting window and check "Share System Audio".
   • Start: Click Start.
   • Language: Toggle between EN <-> JA.

Development

Backend:

cd backend
pip install -r requirements.txt
python download_models.py
uvicorn main:app --reload

Frontend:

cd web-app
npm install
npm run dev

Troubleshooting

• No Audio: Ensure browser permissions are granted.
• Slow Performance: CPU inference is used by default. For faster results, uncomment GPU settings in backend/config.py if you have an Nvidia card.