About BIQE AI HTR SaaS
A practical solution to a stubborn problem: making historical manuscripts legible at scale.
The problem
European archives contain millions of pages of handwriting from before 1950. They've been scanned digitally, but are not searchable. Anyone looking for something in a register from 1894 has to read through it by hand.
Manual transcription is expensive — roughly €5–€15 per page, depending on quality and length. For large collections that means budgets that quickly run into the millions. The result: much material remains uncatalogued.
Automated HTR (Handwritten Text Recognition) can do this more cheaply, but the accuracy of the best open-source models sits around 8–15% error rate at the line level. For publication or serious research, that's not good enough.
Our approach
We add one step to the standard HTR pipeline: a correction layer based on a large language model that places the raw output in context.
A detection model finds the text lines on the page and determines the reading order.
An HTR model reads each text line character by character. Output is raw — often with typographical errors that are obviously wrong to a human reader.
Our layer sends image and raw transcription to an LLM that has contextual knowledge of language, period and document type. The result is substantially more accurate.
Example: a 19th-century administrative document that comes out raw as "Straff ge van genis te Goes voor het" becomes, after correction, "Strafgevangenis te Goes voor het" (a Dutch word for prison). The error is fixed because the LLM knows "strafgevangenis" is a valid word and the fragmented version is not.
Technical principles
Correction, not recreation
The LLM receives both the image and the raw transcription. It acts as a proofreader, not as a re-translator. This minimises hallucinations.
Coordinates preserved
We correct text within the existing PageXML structure. Bounding boxes, reading order and line IDs remain unchanged — you can export searchable PDFs or ALTO files directly.
Prompt presets per document type
The prompt driving the LLM is specialised per document type. A 17th-century register gets different instructions than a 20th-century typescript. We add new presets quickly for specific client projects.
Model-agnostic
We route each tier to the best-fitting model (Gemini, Claude, GPT-4o, DeepSeek and others). If one provider goes down or a new, better model becomes available, nothing changes for our clients.
Measured results
On a test corpus of 19th-century Dutch administrative documents we achieve the following error rates (Character Error Rate, at line level):
| Step | Error rate (CER) | Note |
|---|---|---|
| HTR only (no correction) | ~8–12% | Standard HTR output with republic-model |
| With BIQE correction layer — Balanced | ~2–3% | Gemini 3 Flash |
| With BIQE correction layer — Best | ~1–2% | Claude Sonnet 4.6. Note: not always better than Balanced on Dutch text — see pilot. |
Results vary per material. During the pilot we test on a small selection of your own documents — then we see directly which tier fits best.
About the developer
BIQE AI HTR SaaS was developed by Jannes Hoekman, active in the digitisation of historical material. The software is based on open-source components (we use publicly available models for layout analysis and HTR); the correction layer and orchestration are our own work.
For questions, pilot requests or collaboration: [email protected].