Handwriting OCR Alternative: Leo for Historical Documents You Need to Trust

Handwriting OCR is a web service at handwritingocr.com that turns photos and scans of handwriting into text, built mainly for modern notes and forms. It names no model: its About page says only that it runs AI models trained on public-domain documents. Leo transcribes with its own model, ATR-1, fine-tuned specifically for historical documents, and keeps the writer's spelling on the page. Leo also lets you choose the best model for each part of the job and compare their readings side by side.

"Handwriting OCR" is also a generic term, used by several unrelated apps. This article is about the product at handwritingocr.com, made by Automatic Eye Ltd in London, as its own pages described it in October 2026. For the wider field, see our guide to the best handwriting transcription software.

Which model reads your documents?

The first question to ask any transcription tool is which model reads your documents, and whether it was built for historical material. Everything else, from spelling to the errors you will have to catch, follows from the answer.

Handwriting OCR. Its About page says: "we run AI models trained on hundreds of millions of public-domain documents". It names no model and gives no further description of it.

Leo. The answer is public. Leo transcribes with its own model, ATR-1, fine-tuned specifically for historical documents, handwritten and printed. It is highly accurate, it transcribes what is on the page without silently modernising, it reads the image at high resolution, and it is zero-shot, so there is nothing to train. Built-in failure detection hides suspect output, retries it, and refunds the credit if it cannot succeed.

What Handwriting OCR is

The service started in late 2023. Most of its site is about modern handwriting: it publishes guides to converting handwriting in GoodNotes, Notability, OneNote, reMarkable and on the iPad, and its homepage addresses educators, journal writers and businesses processing forms, timesheets and invoices. Its paid plans add table export to Excel and custom extractors that pull named fields from forms. There is a free trial of 5 pages, then pay-as-you-go credits from $15 per 100 pages or a monthly plan (as of October 2026).

Family history is one strand of its marketing, and that is where its own pages are worth reading closely.

What Handwriting OCR's own pages and reviewers show

In historical work, the transcription is evidence. A writer's spelling, an abbreviation or an unusual name can identify a person or place, and a tidied version loses it. This is the problem our guide to historical language normalization describes.

• The vendor's own demo shows normalisation. Its cursive-to-text page runs a 1940s wartime letter from the Australian War Memorial's archive beside the archive's human transcription. The vendor's notes say spelling "follows the document's language setting (British English here)". They list a name normalised from "Micky" to "Mickey", and the writer's "Your" changed to "You're".
• The AI can add text that is not on the page. Its troubleshooting help says the AI occasionally generates text that was not in the original, most often on faint, smudged or sparse pages. It advises the plain-text export over the "AI-enhanced" format, as "closer to a raw transcription".
• No confidence scores. Its API help says it returns no per-word confidence scores or bounding boxes, so nothing flags the doubtful words for you.
• Reviewers reported misread names and years. Diane Henriks, testing early-1900s Missouri death certificates in 2025, reported a wrong birth year, a wrong maiden name, a wrong medical term, and a birthplace abbreviation read as a place in Germany. A tester on the Airfield Research Group forum, working on a 1902 will and an admission register, reported misread names and capitals, with "Oates" read as "Dalt".
• There is no lasting workspace. Files are deleted automatically after 7 days by default, and at most 14, and the help centre notes that the transcription is deleted with them. You download the results and organise them somewhere else.

Why historians and genealogists choose Leo

Leo is built for one purpose: transcribing historical documents, and keeping that work where you can check, search and build on it. It is not one model but a workspace for managing transcription workflows.

One model, or the best model for each job

In Leo's transcription window, one menu holds dozens of models: Leo's own ATR-1, the recommended default for historical documents, alongside models from OpenAI, Google and Anthropic, and cloud OCR engines such as Google Cloud Vision, Azure AI Read and AWS Textract. Prices are per image, by model.

• Use the best model for each part of the job. ATR-1 reads historical handwriting and print. For clean, legible modern print, such as a typescript or a twentieth-century book, a small model or a cloud OCR engine costs a fraction of a credit, from 0.02 credits an image.
• Compare readings side by side. Each model's reading goes into its own tab on the same image, beside the original, and nothing is overwritten.
• Reconcile them. The Interpolate Transformation combines two readings into one reconciled version, in a new tab.

A highly accurate model, fine-tuned for historical documents

• It transcribes what is on the page. ATR-1 is trained not to normalise, over-correct or silently modernise. A writer's "Recieved", a phonetic place name, an abbreviation or a long s stays as written. General-purpose AI models tend to quietly modernise spelling, and in a historical document the spelling is part of the evidence.
• It reads the image at high resolution, where general models downsample it
• It is zero-shot. There is nothing to train and no sample pages to prepare: upload and transcribe.
• It checks its own work. Built-in failure detection hides output that looks wrong and retries it, and refunds the credit if it cannot succeed. The errors that do get through are the recoverable kind: a wrong character or word, which you check against the image.
• It improves with use. Every correction you make feeds back into training.

You can correct everything

In Leo the transcription is an editable document that sits beside the image. You check a name against the page, fix it, and the fixed version is what you search, translate and export. Your images and corrections stay in your account.

Full flexibility for advanced workflows

• Multiple transcription tabs per image. Re-running a model creates a new tab, never a destructive overwrite.
• Transformations with editable prompts. Correct, Modernize, Interpolate (which reconciles two readings), Translate and Custom produce new versions of a transcription. Summarize, Classify, Generate glossary, Extract named entities and Custom produce annotations. Each writes to a new tab and leaves the base transcription untouched, and you can edit the prompts or write your own.
• Batch work. Transcribe or transform every image in a document at once.
• Search across everything, with global and fuzzy search over documents, transcriptions and annotations, to catch the variant spellings of a name
• Organisation. Nested folders and structured metadata, such as creator, date, archive and identifier, for each document.
• Export and sharing. Export to PDF, Word, HTML or TEI (XML), or send a public read-only link that needs no account to view.
• The Leo API and MCP server (in beta). Scripts and AI agents can work on your documents directly.

Simple pricing

Leo has a free plan with 10 credits each month. One credit is one image, and ATR-1 costs 1 credit an image.

Handwriting OCR vs Leo, side by side

• The model: Handwriting OCR names no model; its About page says only that it runs "AI models trained on hundreds of millions of public-domain documents". Leo uses its own model, ATR-1, highly accurate and fine-tuned specifically for historical documents.
• Choice of model: Handwriting OCR uses its own models. Leo offers ATR-1 plus dozens of OpenAI, Google and Anthropic models and cloud OCR engines, each reading into its own tab to compare.
• Built for: Handwriting OCR is built mainly for modern notes, forms and business documents. Leo is built for historical documents.
• Original spelling: Handwriting OCR's own letter demo shows spelling following the language setting and a name normalised. ATR-1 is trained to transcribe what is on the page, without silent modernising.
• Added text: Handwriting OCR's help centre says its AI occasionally generates text that was not on the page. ATR-1's failure detection hides suspect output, retries it and refunds the credit if it cannot succeed.
• Keeping your work: Handwriting OCR deletes files and their transcriptions after 7 days by default. In Leo the editable transcription sits beside the image, in folders, for as long as you keep it.
• Working with the text: Handwriting OCR offers translation. Leo offers Translate, Modernize, Correct, Interpolate, Summarize, Classify, glossaries, named entities and custom prompts, each in its own tab.
• Price to start: Handwriting OCR offers a 5-page trial, then from $15 per 100 pages. Leo has a free plan with 10 credits every month.

How to test both on your own documents

Run the same small sample through each:

1. Pick five images that represent the job, including the hardest: a faint page, a crowded register entry, an unfamiliar hand
2. Check names, dates, numbers and places first. One wrong character in a surname or a year does the most damage.
3. Check the spelling against the image. Mark every modernised word, expanded abbreviation or corrected name.
4. Ask where the transcription will live in a year, and whether you can still correct, search and export it

Our guide to how to test AI transcription on your own documents sets out a fuller method, and verifying transcription accuracy covers the checking. For family historians, our guide to genealogy record transcription covers the method from register to family tree.

Leo's free plan gives you 10 credits every month, enough to try ATR-1 on your own documents today.

Frequently Asked Questions

What AI model does Handwriting OCR use?

It does not name one. Its About page says only that it runs "AI models trained on hundreds of millions of public-domain documents". Leo publishes its answer: it transcribes with its own model, ATR-1, fine-tuned specifically for historical documents.

Does Handwriting OCR keep the original spelling?

Not always, by the vendor's own example. Its letter demo notes that spelling follows the document's language setting, and that a name was normalised from "Micky" to "Mickey" and "Your" changed to "You're". Leo's ATR-1 is trained to transcribe what is on the page, without silently modernising it.

Can I correct and keep my transcriptions?

In Handwriting OCR you can edit a transcription, but files and their transcriptions are deleted after 7 days by default, so you must download them first. In Leo the editable transcription sits beside the image in your account, in folders, searchable, and exportable to PDF, Word, HTML or TEI.

Can I compare different AI models in Leo?

Yes. Leo's model menu holds ATR-1, the default for historical documents, plus dozens of OpenAI, Google and Anthropic models and cloud OCR engines. Each reading goes into its own tab on the same image, and the Interpolate Transformation reconciles two readings into one. Nothing is overwritten.

Is Handwriting OCR free?

No. It offers a free trial of 5 pages, and after that it charges per page, from $15 per 100 pages pay-as-you-go or through a monthly plan (as of October 2026). Leo has a free plan with 10 credits every month, and one credit transcribes one image with ATR-1.

What is the best Handwriting OCR alternative for historical documents?

Leo. Its own model, ATR-1, is highly accurate and fine-tuned specifically for historical documents, and it transcribes what is on the page rather than normalising it. It reads the image at high resolution, needs no training, and hides, retries and refunds output that looks wrong. Around it, Leo gives you an editable transcription beside the image, dozens of other models to compare in their own tabs, Transformations, search, export to TEI and an API in beta, with a free plan of 10 credits each month.