Handwriting to Text
AI transcription for handwriting — measured on a real manuscript, not a demo.
VisionParse transcribes handwritten documents with its AI engine — the one capability where classical OCR is not weaker but unusable.
We measured both engines on a real 19th-century handwritten manuscript page: classical OCR read it at 43.9% — garbage — while the AI engine read it at 95.3%, opening lines letter-perfect. That is why this tool runs on the AI tier at one credit a document: it is the only honest way to offer handwriting at all.
- AI tier · 1 credit a document
- JPG · PNG · WEBP · PDF
- 3 MB · 3 pages
- Deleted within 1h
What this tool does, in numbers.
- ACCURACY
A real 19th-century handwritten manuscript page read at 95.3% character accuracy through the AI engine, opening lines letter-perfect.
- LIMIT
The same page through classical OCR scored 43.9% — unusable. Handwriting is why this tool is AI-tier: there is no free engine that can do this.
- SPEED
A handwritten page transcribes in roughly 2 seconds, at about a tenth of a cent of model cost.
- LANGUAGES
Script-independent: the measured page was right-to-left connected Arabic — a hard case — and mixed-language pages are handled.
- ACCESS
Needs an account with AI credits: every paid plan includes them monthly, and a one-time $15+ credit pack unlocks the capability without a subscription.
- PRIVACY
Illegible spans come back marked [illegible] — the model is instructed never to invent plausible text for what it cannot read, and it names the passages it is unsure of.
- RETENTION
Originals are deleted within one hour of processing, transcriptions within 30 days, and nothing trains any model.
Where the 95.3% on a real manuscript comes from.
Every figure on this page traces to a dated run on hardware we name. Nothing here is a vendor estimate.
Measured 2026-08-03
- Claim
- Real handwriting is readable — by the right engine: 95.3% on a 19th-century manuscript where classical OCR managed 43.9%.
- Method
- VisionParse standard recognition tier, single-threaded. Full methodology.
- Sample
- A genuine 19th-century handwritten naskh manuscript page (Gotha collection), hand-transcribed ground truth, both engines on the identical image, scored character-by-character with manuscript-fair normalization.
- Measured
2026-08-03on Contabo Cloud VPS 8 — 8 vCPU AMD EPYC, 24 GB RAM, Ubuntu 24.04.4- Limitations
- One real page is evidence, not a corpus — hands, inks and centuries vary, and the figure carries roughly ±3 points of transcription-judgement uncertainty. For a collection, we measure your material first.
- Contact
- VisionParse@senithu.lk
What this tool gets wrong
Connected cursive script — the hardest case — read at 95.3%. The measured page is right-to-left, fully joined handwriting from another century. Modern print-style handwriting on forms is the easier case, not the harder one.
Illegible is illegible — and the tool says so instead of guessing. Faded ink, crossed-out words and truly ambiguous letters come back marked [illegible], with uncertain passages named. A transcription that invents plausible text where it cannot read is worse than a gap, so this one never does.
One page measured, not a statistic. Our figure comes from one real manuscript page plus a degraded printed book scan (94.5% → 99.2% on the same engine). Your grandmother’s letters or your archive’s registers deserve their own measurement — send samples and we measure before you spend.
Three steps, no account.
Drop the handwritten page
A letter, a form, a manuscript photo — PNG, JPG, WebP or PDF up to 25 pages.
The AI engine reads it
One credit a document, roughly 2 seconds a page. No classical-OCR pre-pass: on handwriting its output is noise.
Take the transcription
UTF-8 text with line breaks preserved, illegible spans marked, uncertain passages named. Copy it or download a .txt.
What this tool accepts.
| Property | Free tool |
|---|---|
| Engine | AI tier (the extraction model), 1 credit per document |
| Measured accuracy | 95.3% on a real 19th-century manuscript (2026-08-03) |
| Classical OCR on the same page | 43.9% — the reason this tool is AI-tier |
| Input formats | JPG, PNG, WEBP, PDF — up to 25 pages |
| Output | Plain text, UTF-8, [illegible] markers, uncertain passages listed |
| Access | Account with AI credits — any paid plan, or a one-time $15+ pack |
| File retention | Originals deleted within 1 hour; results within 30 days |
| Processing location | Frankfurt, Germany |
When to use the API instead.
This page converts one file now. To convert files from your own software, the same engine runs over HTTP from Starter upward.
# same engine, from your code curl -X POST https://api.visionparse.app/v1/ocr \ -H "x-api-key: $VISIONPARSE_KEY" \ -F "file=@image.jpg" → 202 { "jobId": "c0df8e2c…" }
Use the API when
You convert more than a few files, convert on a schedule, or need results inside another system. Batch, webhooks and bounding boxes are API-only.
Frequently asked
Why is this not free like the other tools?
Because free OCR engines cannot read handwriting — we measured 43.9% on a real page, which is garbage. The AI engine that can costs real money per call, so it runs on credits: included monthly with every paid plan, or unlocked by a one-time $15 pack.
What kinds of handwriting work?
The measured cases are the hard end: a 19th-century joined manuscript (95.3%) and a degraded printed book page (99.2%). Modern handwriting on forms and letters is generally easier than both. Truly illegible writing comes back marked as such rather than guessed.
Can it transcribe old manuscripts and archives?
That is exactly the measured case, and archive projects get measured first: send three representative pages and we benchmark them through the real pipeline before you commit to anything.
Does it invent text where it cannot read?
No — this is the design rule that matters most. Illegible spans return as [illegible] and the model names passages it is unsure of, because invented plausible text in an archive is worse than an honest gap.
Do you keep my documents?
Originals are deleted within an hour of processing, transcriptions within 30 days, and nothing you upload ever trains a model.