G2G: Have you used Transkribus or other Handwritten Text Recognition software?

Question

Hi all!

Does anyone here have experience with Transkribus (or other Handwritten Text Recognition software)? I am asking because I am doing research into this topic, specifically relating to it's use in heritage collections in libraries and archives, for my library and information services diploma and just general interest. I would love to hear about your experiences with the software, especially if you have used it in the course of your genealogy or history research.

Thanks!

Answer 1

Hi Oscar,

I work at a university library and we have used both Transkribus and other HTR software.

It is not for genealogy or history research we use it at the moment. If you still would like to know about it, you're welcome to contact me using private message on my profile.

Kind regards

Maria

Answer 2

I stumbled upon Transkribus and found that it was pretty accurate for some old court case files.  Has anyone found something that works well without converting pdf into an image format or something that could transcribe several files at once?  My thought is that this could save me lots of time with respect to opening each file to see what it says.

Comments

Transkribus works with pdf files.

https://help.transkribus.org/uploading-files-to-transkribus

"2. PDF
When uploading a PDF, each page of the PDF is extracted and uploaded as a page of the document. At this time, only one PDF can be uploaded per submission, with a file size limit of 200 MB or a total of 3,000 pages."

Answer 3

Interesting essays on HTR here: 

Discovering the Power of AI for Reading Medieval Manuscripts: HTR Winter School 2022

Handwritten Text Recognition for Documentary Medieval Manuscripts

Cremma Medieval | Share data | HTR-United

and 

Sergio Torres Aguilar‬ - ‪Google Scholar‬

For German manuscripts:
[2201.07661] Open Source Handwritten Text Recognition on Medieval Manuscripts using Mixed Models and Document-Specific Finetuning

Answer 4

I do some transcribing at the National Archives (US) website. They use some automated system, I don't know the details of it, but the results vary quite a bit depending on the writing. I'm looking specifically at 1870's US documents. Some pages it's about 80-90% accurate, other pages maybe 20%.

It only looks at the words individually, it doesn't try to make sense of the whole page or whole set of pages together. Like if a person's name is clear in one spot and hard to read in another, it won't cross-reference the name. It also doesn't try to narrow down word choices based on the topic of the page.