Appearance
Transkribus charges for transcription through prepaid credits, spent mainly when you run text recognition over pages — roughly one credit per page. Uploading, layout analysis, model training and correction in the editor do not consume recognition credits. That makes budgeting largely a matter of counting pages, confirming the current per-page rate, and adding a sensible contingency for re-runs. Here is how to plan a project without nasty surprises.
What exactly do credits pay for?
The key mental model: you pay to read pages, not to handle them. Spend and no-spend break down clearly.
| Action | Spends recognition credits? |
|---|---|
| Upload images | No |
| Layout / line detection | No |
| Run text recognition | Yes (~1 / page) |
| Correct text in the editor | No |
| Train a custom model | No (training is separate) |
| Re-run recognition after improving a model | Yes, again |
The trap is re-running recognition. Every time you reprocess pages with a new model, you pay again. So the cheapest workflow recognises the whole collection once, with a model you have already validated.
How do you estimate the cost of a collection?
Cost is close to linear in page count, which makes a back-of-envelope estimate reliable:
text
estimated credits = pages x credits_per_page x expected_recognition_runs
example: 12,000 pages x 1 x 1.15 (15% re-run buffer) = 13,800 creditsAlways plug in the current published rate rather than a remembered number, and treat the result as recognition spend only — correction labour is your separate, larger human cost.
Why run a pilot before buying bulk credits?
A pilot of 50–100 representative pages tells you two things that dominate the budget: whether a public model is accurate enough (avoiding a re-run after custom training) and how much human correction each page needs. Spending a handful of credits up front routinely saves thousands later by catching a model mismatch before you process 12,000 pages with the wrong one.
text
Pilot checklist
[ ] 50-100 pages across the worst and best hands
[ ] measure CER of candidate public models
[ ] estimate minutes-per-page correction
[ ] decide: public model now, or train first?How do you avoid paying for recognition twice?
- Validate the model first. Confirm CER on the pilot before the full run.
- Train before the big run, not after. Custom training does not consume recognition credits, so improve the model while it is cheap, then recognise once.
- Recognise the full collection in a single pass with the chosen model.
- Reserve corrections for the editor, which is free, rather than re-recognising.
Following this order, a 12,000-page collection is one recognition run plus correction, not three runs while you experiment on the whole set.
Do credits expire, and how should you buy them?
Purchased credits usually carry a validity window rather than lasting forever, and the free starter credits may differ from paid ones. Buy in tranches matched to your timeline — enough for the pilot, then a block sized to your estimate plus a 10–20% contingency. For large or grant-funded work, contact Transkribus about institutional or volume arrangements instead of buying small packs at list price.
A worked budget example
For a parish-register project of 12,000 pages:
| Line item | Quantity | Notes |
|---|---|---|
| Pilot recognition | ~100 credits | Validate model, measure correction load |
| Full recognition | 12,000 credits | One pass with chosen model |
| Contingency (15%) | ~1,800 credits | Re-runs, missed pages |
| Total credits | ~13,900 | Confirm at current rate |
| Human correction | separate | Often the real cost driver |
The headline lesson: recognition credits are predictable; the larger, looser cost is human correction time, so budget that explicitly too.
Key Takeaways
- Credits pay for text recognition (~1 per page); upload, layout, training and correction do not spend them.
- Cost scales almost linearly with page count — easy to estimate.
- Re-running recognition costs again; recognise the full collection only once.
- Run a small pilot to validate the model and gauge correction effort first.
- Train custom models before the big run, since training is not charged per page.
- Add a 10–20% contingency and buy credits in timeline-matched tranches.
- Budget human correction time separately — it usually dwarfs credit costs.
Frequently Asked Questions
How do Transkribus credits work?
Credits are a prepaid currency you spend on text recognition, charged roughly per page processed. Layout analysis, model training and correction in the web app do not consume recognition credits; only running a model over pages does.
How much does it cost to transcribe one page?
Recognising a page typically costs about one credit, so cost scales almost linearly with page count. Always confirm the current rate and any volume discounts on the official pricing page before budgeting, as rates change over time.
Do unused Transkribus credits expire?
Purchased credits generally remain valid for a defined period rather than indefinitely, so buy in tranches that match your project timeline. Check the current terms, because the free starter credits and paid credits may have different validity.
Is training a model free in Transkribus?
Training a custom model does not consume recognition credits in the same way page recognition does. Your main costs are the recognition runs over your collection and any re-runs after you improve a model.
How do I avoid surprise costs on a big collection?
Run a small pilot to confirm accuracy before committing, recognise the full collection only once with a good model, and avoid repeatedly re-running recognition. Budget a contingency of 10 to 20 percent for re-runs and corrections.
Are there discounts for institutions or large volumes?
Transkribus offers volume and institutional arrangements in some cases. For large or funded projects it is worth contacting them directly rather than buying small credit packs at list price.