---
title: BilinguaScript HTR API
emoji: ✍️
colorFrom: blue
colorTo: green
sdk: docker
pinned: true
app_port: 7860
---

# BilinguaScript Model API

FastAPI service that serves fine-tuned TrOCR models for bilingual (EN/FR) handwritten text recognition.

## Endpoints

| Method | Path | Description |
|--------|------|-------------|
| GET | `/health` | Health check |
| POST | `/api/v1/transcribe` | Transcribe uploaded image line |

## Environment Variables

| Variable | Default | Description |
|----------|---------|-------------|
| `EN_MODEL_ID` | `unixio/trocr-cameroon-english-lora` | Fine-tuned English model (Cameroonian-adapted, 18.62% CER on held-out Cameroonian writers) |
| `FR_MODEL_ID` | `unixio/trocr-rimes-french-lora` | Fine-tuned French model (11.29% CER on RIMES-2011) |
| `HF_TOKEN` | — | **Required**: read token for the private model repos (set as a Space secret) |
| `NUM_BEAMS` | `4` | Beam search width |
| `MAX_LENGTH` | `160` | Max generated tokens per line |

## Request parameters

`POST /api/v1/transcribe` (multipart form):

| Field | Default | Description |
|-------|---------|-------------|
| `file` | — | Image (PNG/JPEG) |
| `language` | `auto` | `en`/`english`, `fr`/`french`; `auto` falls back to English |
| `segment` | `true` | `true`: split a full page into lines first; `false`: treat the upload as one pre-cropped line (what the Django backend sends) |