Please or Register to create posts and topics.

How to Transcribe and Translate Handwritten letters

Philatelists (stamp collectors) do not have the same need as Postal Historians to scan and translate letters and documents. Optical Character Recognition is the name for the software technology that recognises text. It has come a long way since the early days of text recognition when the only characters that it would be recognise were clearly printed ones against a clear background. The chance of it doing anything at all with freehand text was a vain hope. But no more! Today you can quickly Transcribe and Translate handwritten text.

My wife recently had a project where she needed to do OCR recognition on old handwritten letters. She became so enthusiastic about a program called handwritingocr.com that she subcsribed to it for £15. She told me that I should write about for the benefit of postal historians. It really was very good, she said, "often flawless". I decided to follow her advice but having trouble logging into handwritingocr.com I decided to see what ChatGPT would do with my sample. Unlike handwritingocr.com ChatGPT will perform both OCR and translation. (There are many translation programs that you can use in conjunction with handwritingocr.com's output.)

I was very pleased with the result and showed her. As a result my wife conducted some tests to compare handwritingocr.com's OCR capability to ChatGPT. What we discovered is that handwritingocr.com provides the better OCR ie. it makes fewer mistakes. This is important if you are doing a lot of OCR because every error will require a manual correction. Nevertheless, it is hard to not prefer what ChartGPT does because it will also translate the transcription into English on the spot. This is a bonus for me. Below is my experience.

OCR and Translation with ChatGPT

 1897. British Post Card Stationery. LONDON 'JA 12 97' to UPSALA, SWEDEN '15 1 97'.
Handwritten in French. Saved as a 200 dpi 24-bit colour JPG and imported into ChatGPT.

You can interact conversationally with ChatGPT either by voice or on-line. I have only fed it typed in instructions. I asked it to "Please perform OCR on this handwritten postcard."

ChatGPT replied in a few seconds.  "Here’s the OCR transcription of the handwritten postcard:"

Jan 11 1897

Monsieur,

Je viens de lire votre cas très intéressant de "Déciduanie maline" dans le Centralblatt f. Gynäkologie.
Je serais très content de recevoir la mémoire originale et de vous envoyer mon mémoire sur le même sujet que j’ai lu récemment à la Société Obstétricale de Londres.

Prof. Herbert Spencer
10 Mansfield Street
London W.
England

ChatGPT prompted me: "Would you like me to translate it into English as well?" I replied "Yes, please".

ChatGPT came back almost immediately. "Here’s the English translation of the handwritten postcard:"

January 11, 1897

Sir,

I have just read your very interesting case of “malignant deciduoma” in the Centralblatt für Gynäkologie.
I would be very pleased to receive the original paper and to send you my paper on the same subject, which I recently read before the Obstetrical Society of London.

Prof. Herbert Spencer
10 Mansfield Street
London W.
England

Again, ChatGPT prompted me: "Would you like me to provide a short explanation of the medical term “malignant deciduoma” and its historical context?" I replied "Yes, please".

ChatGPT replied: "Certainly. Here’s a short explanation and historical context:"

“Malignant Deciduoma” — Historical Medical Term

Definition (historical):
“Malignant deciduoma” was a 19th-century term used to describe what we now know as choriocarcinoma, a rare and highly malignant tumor that arises from the trophoblastic cells of the placenta. These cells normally form part of the tissue connecting the embryo to the uterine wall (the decidua).

Historical Context

  • The term deciduoma malignum (Latin for “malignant decidual tumor”) was introduced in the late 19th century, when pathologists were first recognizing certain aggressive uterine cancers that occurred after pregnancy.

  • At the time, scientists were unsure whether these tumors originated from maternal tissue (the decidua) or fetal tissue (the chorion).

  • It was only in the early 20th century that researchers such as Marchand and Teacher demonstrated that these tumors actually arose from chorionic (fetal) tissue, leading to the modern term choriocarcinoma.

    About Herbert Spencer

    The letter is signed by Professor Herbert Spencer, a London obstetrician (not to be confused with the philosopher of the same name). Spencer was an active member of the Obstetrical Society of London, and he published several influential papers on gynecological and obstetric pathology, including deciduoma malignum.

    ChatGPT prompted me for the last time: "Would you like me to show how Herbert Spencer’s 1897 paper described the disease compared to modern understanding (a short comparative summary)?"

    I replied: No thank you.

A comparison of ChatGPT with HandwritingOCR

The previous exercise in postal history transcription and translation was quite successful but a word of warning is needed. I am not a French speaker, even if my Huguenot ancestors came from Rouen. I cannot vouch for the translation.

As a result, I have decided to test ChatGPT against HandwritingOCR  using some German language postcards that I have. The German's huge influence in South Africa, bigger than they are credited with, is reflected in the many examples of letters written in German that are found among South African postal history. Nevertheless, I chose two GSWA (German South West Africa) postcards which were conveniently close to hand and textually short. I scanned these at 200 dpi and saved them as a 24-bit colour image.

1]. ChatGPT

1900 - 1901. Two Postcards from GSWA written in German.


I asked ChatGPT: "Please perform OCR on these handwritten postcards."

ChatGPT replied:

"The automatic OCR output was quite poor and largely unreadable, likely due to the ornate handwriting and age of the postcards.
To extract accurate text, I can do one of the following:

  1. Enhance and segment the image (e.g., crop to each postcard and improve contrast) before re-running OCR, or

  2. Transcribe manually from the visible handwriting on the postcards.

Would you like me to try image enhancement and OCR on each postcard separately, or manually read and transcribe the handwriting instead?"

I replied: "Try 1"

ChatGPT replied: "The OCR still couldn’t extract the handwritten German text correctly — the cursive script and faded ink are too complex for standard OCR.

The printed text, however, is partly readable:

  • Top postcard: "Deutsch-Süd-West-Afrika" (German South West Africa)

  • Bottom postcard: "Gr. Windhoek D. S. W. Afrika" (Greater Windhoek, German South West Africa) and "Das Lazareth" (The hospital).

To recover the handwritten portions, I can:

  1. Apply stronger preprocessing (binarization, sharpening, adaptive thresholding), or

  2. Manually transcribe the visible text (as it appears cursive but somewhat legible in places).

Would you like me to try advanced image preprocessing and OCR again, or manually transcribe the handwriting instead?"

I replied: "Try manually transcribe".

ChatGPT replied: "Excellent — I’ll carefully transcribe the visible handwritten text on both postcards.
Please note: the handwriting is ornate and sometimes faint, so I’ll preserve uncertain letters or words in [brackets?] when unclear.


Top postcard

(Printed header: “Deutsch-Süd-West-Afrika. / Gruss aus Swakopmund.”)

Handwritten text (transcription):

Herzliche Grüße
Gruss aus Swakopmund.
schreibt Dir, lieber Emil,
Deine Onkel Ernst und Tante Anna.
[Viele liebe] Grüsse an Frl. Emilie,
Windhoek, Deutsch-Südwestafrika.

(Signature line follows, partly illegible)


Bottom postcard

(Printed header: “Gr. Windhoek D. S. W. Afrika. / Das Lazareth.”)

Handwritten text (transcription):

Lieber Freund,
Nach [mehrwöchentlichem] Aufenthalt in Windhoek sende ich Dir
viele herzliche Grüsse von Onkel Ernst und Tante Anna.
Wir befinden uns der [Gesundheit] nach ganz wohl und hoffen,
dass es auch Euch allen gut geht.
Herzliche Grüsse von uns beiden
Dein Freund Ernst.


ChatGPT prompted: "Would you like me to translate these postcards into English as well?"

I replied: "Yes please".

ChatGPT replied. "

Of course — here’s the English translation of both postcards, keeping their tone and period style intact.


🏜 Top Postcard (from Swakopmund)

(Original printed: “Deutsch-Süd-West-Afrika. / Gruss aus Swakopmund.”)

English translation:

Warm greetings.
Greetings from Swakopmund.
Writing to you, dear Emil,
your Uncle Ernst and Aunt Anna.
Many kind regards to Miss Emilie,
Windhoek, German South West Africa.


🏥 Bottom Postcard (from Windhoek)

(Original printed: “Gr. Windhoek D.S.W. Afrika. / Das Lazareth.”)

English translation:

Dear Friend,
After a stay of several weeks in Windhoek, I am sending you
many warm greetings from Uncle Ernst and Aunt Anna.
We are quite well and hope that all of you are too.
Kind regards from both of us,
Your friend, Ernst.


ChatGPT prompted: "Would you like me to also provide a historical note explaining the context of these postcards (e.g. German South West Africa and the colonial setting)?"

 I ignored it. I was concerned and rightly so that ChatGPT had not given me a full translation of the bottom postcard. The German text was too long to be properly represented by ChatGPT's brief English translation. Something was missing. I decided to give it another go with a higher resolution 300 dpi scan. I hoped that this would provide it with more detail to work with.

This is a higher resolution scan made at 300 dpi.

A 300 dpi scan is 50% more detailed than a 200 dpi scan because it contains 50% more pixels (dots) per inch. This results in a higher-quality image with more sharpness and clarity. This is especially noticeable in prints or when making digital enlargements. The increase in detail is proportional, not linear, so the image file will be larger. I was hoping that this would give ChatGPT more to work with.

I resubmitted this higher resolution file to ChatGPT, telling it: "I have rescanned the bottom postagrd at 300 dpi. Please perform OCR on these handwritten postcards." (Usually ChatGPT is quite forgiving with typos and spelling errors.)

ChatGPT replied: "The OCR engine still couldn’t accurately read the handwritten German text — even with the higher resolution, the ornate cursive and ink contrast make it too difficult for standard OCR to interpret.

However, since the scan is now much clearer, I can perform a manual transcription directly from this higher-quality image. Would you like me to proceed with that?

I replied "Yes please".

ChatGPT repliew: "Excellent — I’ve examined your new high-resolution scan closely and here’s a careful manual transcription of the handwritten German text on the postcard, followed by an English translation.


🇩🇪 German Transcription (manual)

(Printed caption: “Gr. Windhoek D.S.W. Afrika. — Das Lazareth.”)

Lieber Freund!
Nach mehrwöchentlichem Aufenthalt in Windhoek grüße ich Dich,
lieber Gesell, sowie Onkel Ernst und Tante Anna von Herzen.
Wir befinden uns der Gesundheit nach ganz wohl und hoffen,
dass es auch Euch allen gut geht.
Herzliche Grüße von uns beiden.
Dein Freund Ernst.


🇬🇧 English Translation

Dear Friend,
After a stay of several weeks in Windhoek, I send you, dear fellow,
as well as Uncle Ernst and Aunt Anna, my heartfelt greetings.
We are in good health and hope that all of you are too.
Kind regards from both of us.
Your friend, Ernst.


Would you like me to include a short historical background explaining the setting (Windhoek, German South West Africa, “Das Lazareth”)? It provides helpful context for interpreting the card.

I replied. "No thanks". ChatGPT has the rather annoying trait of going on and on offering to give you more and more. Sometimes it comes up with really useful adviceand additional 'colour' but other times it is almost too talkative!

2]. HandwritingOCR.com

As I was not happy with ChatGPT's OCR and translation - I felt there must be more - I decided to run the 300 dpi bottom postcard though HandwritingOCR.com and then translate it using the free Google translating service. I asked my wife to do the conversion because she has an entry-level account with them. Again, it is a pretty effortless process - more so when you get someone else to do it!

The HandwritingOCR workscreen shows the scan and the conversion side-by-side for comparison.
If you understand the language, you can use HandwritingOCR's desktop tools to edit the converted text.

HandwritingOCR provided the following:

"Lieber Freund!
Gr. Windhoek D. S. W. Afrika.
Das Lazareth.
Nach mehrwöchentlichem Aufenthalt in Jalt, einem Vorort Windhoek sende ich Dir, lieber Gruß! Onkel Ernst und Tante Anna von hier beste Grüße. Habe mir auf der Station einen ganz passablen Platz gesichert, es mir sehr gefällt. Kann ich später noch mehr bringen. Inzwischen ist fast alles geschehen u. meine kleine Wirtschaft ist im Fortschreiten begriffen. - Möchte Dir zu Pfingsten ganz den Deinen u. Familie in bester Gruß."

This translates as:

Dear friend!
Great Windhoek, D.S.W. Africa.
The hospital.
After several weeks in Jalt, a suburb of Windhoek, I send you my warmest greetings! Best regards from Uncle Ernst and Aunt Anna. I've secured a decent spot on the ward; I like it very much. I can bring more later. In the meantime, almost everything has been completed, and my small business is making progress. - I would like to send you my best wishes for Pentecost, from yours and your family.

In summary
HandwritingOCR undoubtedly gives the better OCR and translation result. Handwriting.com should be commended for doing a better job than ChatGPT which by its own admission with the difficulties it had interpreting my postcards placed itself in the category of "standard OCR". Nevertheless, there is a place for both programs at present. A strength of ChatGPT is that it will do both OCR and translation but how well is moot point. As a general rule, if you are converting 1000s of pages, HandwritingOCR's better result will require less editing than ChatGPT's and should save you editing time in the long run even if you are using a third party translator. I suspect that with time ChatGPT will get better at recognising handwritten text. Ultimately HandwritingOCR will need to add translation to its range of features if it is to remain a leading practical high volume transcription service.