TranslatorAI — Offline Multilingual Translator (Raspberry Pi 5)

Fully offline OCR + Speech + Text translation across 12 Indian languages. Runs on-device: no cloud, no internet needed after setup.

Stack

• OCR: Tesseract
• NMT: IndicTrans2 (CTranslate2 INT8, indic↔en, pivot for indic↔indic)
• STT: faster-whisper (base)
• TTS: Piper (fallback: eSpeak-NG)
• Backend: FastAPI
• Frontend: static HTML/JS (static/index.html)

Hardware

Tested on Raspberry Pi 5 (8GB RAM recommended — 200M param models + Whisper base need headroom). 4GB may work but expect swap.

Quick Start

git clone https://github.com/<you>/translator-ai.git
cd translator-ai
bash setup_models.sh          # pulls models, voices, system deps — takes a while
uvicorn main:app --host 0.0.0.0 --port 8000

Visit http://<pi-ip>:8000.

Repo Layout

translator-ai/
├── main.py              # FastAPI backend — OCR, translate, TTS, image overlay
├── static/index.html    # frontend UI
├── requirements.txt      # python deps
├── setup_models.sh       # downloads/converts models (NOT in git — too big)
├── Dockerfile            # optional containerized run
├── .gitignore            # keeps models/voices out of git
└── README.md

Models (not committed — see setup_models.sh)

Model	Source	Purpose
indictrans2-indic-en-dist-200M	ai4bharat (HF) → CT2 int8	Indic → English
indictrans2-en-indic-dist-200M	ai4bharat (HF) → CT2 int8	English → Indic
whisper base	faster-whisper	Speech → text
Piper voices (9 langs)	rhasspy/piper-voices	Text → speech

Health Check

GET /health — reports which models loaded + which fonts found.

Debug

GET /debug/translate?text=Hello&src=eng_Latn&tgt=tam_Taml

License