Using AI to OCR and Transcribe a 1935 typed travel diary

In 1935 my maternal grandfather, Reginald Boston (1909-1970), at the age of 25, sailed on the R.M.S. “Orontes” from London, England to Fremantle, Western Australia, to be with his fiance and to start a new life there.

He was a professional “corrector of the press”, i.e. a copy writer, and so was great writer.

I recently came across his typed 24 page account of the voyage.

Here’s how I went from those 24 pages, to this Google document presentation (below) which includes a map, AI-generated images, and a photo of the man himself and his wife-to-be.

Please note that:

  • I’m not going to spend much time explaining what DIDN’T work, and why.
  • Given the rate at which AI is improving, I don’t doubt that there will be faster, easier ways in the future – probably already!

Here we go:

  1. Epson ET-2850 multifunction printer/scanner. I used the Epson 2 scanner software of my Windows 11 Pro PC to scan all the pages into a single PDF, which you can see here: At this stage the pages are just a sequence of images in the PDF, one image per page.
  2. Thanks to the Facebook group called Genealogy and Artificial Intelligence (AI) I got referred to an AI-powerd OCR (optical character recognition) online service with a generous free allowed called Transkribus. I uploaded my PDF here.

From the ‘Jobs’ page I could see my uploaded document:

Click on the Title, and now I can see all the iindividual page. Select them all, then click on ‘Recognize’

Many different ‘Language Models’ become available at this stage, each one optimised for different specialities. I narrowed my search down to ‘English Printed’, chose ‘Transkribus Print M1’, then clicked ‘Start Recognition.

This created another ‘Job’ with a type ‘Text Recognition’. Open that up, select all, click on the 3 horizontal dots and choosed ‘Export’

I chose to export to ‘Docx files’, but I’m guessing that ‘Text Files’ might have worked just as well.

When it’s done, Transkribus sends you and email with a link to download the document (as as .zip file).

  1. I opened ChatGPT 4 (the paid version), and copied this prompt (as suggested by someone else in the Facebook group), followed by a full cut & paste of all the text from the .docx file.

    You are an expert transcriptionist and editor. Your goal is to create great writing. Find below the raw OCR text from 24 pages of text; the OCR quality was poor; the file needs proofreading. Act as a smart OCR chatbot; convert the bad input text below as a rough OCR scan; correct the raw text to instead reflect the most likely original text. As your conversation goes further, consider the context of the OCR text in your considerations. Correct and proofread the entire journal segment.
  2. ChatGPT didn’t want to do this all in one go, so I had to click on ‘Continue Generating’ multiple times.
  3. After ChatGPT had worked its magic, I cut & pasted the (amazing!) text into a Google doc.
  4. In the same ChatGPT chat as the above, I asked the following:
    For each of the days listed in the diary, please create an appropriate graphic using a style consistent with the 1930s.
  5. ChatGPT wasn’t able to obey the above command, but it did give me text like this:
  1. In an new ChatGPT chat, I then asked ChatGPT to generate the images, as per the suggested text (above):
  1. And that, ladies and gentlemen, is about it! Oh, except for one more thing: the map!

The map was a little more involved. In brief, I manually created a spreadsheet with the key places and dates, cut & pasted that sheet into ChatGPT, and asked it to create a map, with labels and lines joining each place. It took a bit of backwards and forwards chat to get it right. The end result was actually an HTML page which I could open in my browser, and take a screenshot of.

Leave a Reply

Your email address will not be published. Required fields are marked *

This site uses Akismet to reduce spam. Learn how your comment data is processed.