Is there an add-on to help with transcribing a document?
Something equivalent to Transcript or GenScriber (which I haven’t tried since they’re built into my current software)
I looked through the list of add-ons, but I didn’t find anything…
Thank you
For Windows, I use Genscriber.
You can find it at
I don’t know if there’s a plugin.
You might want to check if the WebSearch addon, which allows integration with an AI, can do that (??)
Thanks,
if there isn’t one, it’ll be a chance to test Genscriber 
Well, I wasn’t referring to using an AI to transcribe handwritten text (like Généatique’s announcement).
I meant a “simple” utility that allows you to zoom, adjust contrast, etc., on a document, along with a window for entering the transcription, so you don’t have to keep switching back and forth endlessly.
But thanks for the info on AI; it could still come in handy!
I use Transkribus:
https://www.transkribus.org/information-extraction
https://www.transkribus.org/transcribing
Then, if specific formatting is needed to enter data into Gramps, I use an AI (“instructions/prompts”) to generate a spreadsheet or separate the data.
Hi Jérôme 
Do you happen to have a quick tutorial lying around?
Edit: Ah, I just saw the thread about a Transkribus addon you replied to. I’ll follow that.
By following a few links, you can discover related projects (or alternative interfaces to Transkribus). Some are interesting for genealogists and the indexing of personal data in metropolitan France.
https://teklia.com/fr/blog/projet-socface-reconstituer-lhistoire-de-la-france-par-le-bas/
https://teklia.com/our-solutions/arkindex/
https://teklia.com/research/open-source/
etc.
Actually, I use Transkribus because some models perform very well with the German language. Certain historical periods require long and tedious deciphering.
Generative AI sometimes takes too many liberties or, unfortunately, seems to be fed by the semantics of business schools… I spent more time correcting hallucinations. Admittedly, it can be fun to discover a fictionalized text, except when doing genealogical research based on factual leads.
In Gramps, my needs were limited to transcription assistance (and translation in the case of German). A small gramplet or tool, an open API, and an API key (or ACCESS_TOKEN) were enough to make me happy.
Then, by digging a bit deeper into the subject, I came across this video:
Cool! A whole ecosystem in Python and SQLite databases, with exports in JSON or text formats…
An environment likely well-suited for French, Norwegian, perhaps Spanish, Swedish, and English (deductions are not the exclusive privilege of AI…). Letter recognition should also work well with Italian or Portuguese. I still lack a strong reference for German (Germany, Austria, Switzerland, Belgium, Luxembourg, etc.).
A genealogical circle, an association, a municipal archive, etc., can very well deploy such solutions. Indeed, who has never cursed at a former civil registrar or census taker during the attempt to decipher a handwritten document? A few hours to train the tool, and several days or weeks available to analyze these deciphered data more deeply!
As for predictions (variable temperature in some AIs), this is clearly for more complex corpora than civil registration or census records.
https://atr.pages.teklia.com/dan/dan/
arkindex can provide an alternative to local deployment, Transkribus, escriptorium, or a dedicated gramplet.
Sorry for my ignorance in this fascinating field.
Does that mean that every time there’s a change in the script (for example, a change of curate), do you need to restart the learning process?
Just like the parishioners of the time! 
More seriously, by default, a “modern” handwritten text in French doesn’t pose major difficulties (LLMs, APIs, transformers, AI, etc.). These models have so much experience with these ‘partner’ archives in France and Quebec that you’d have to stumble upon the only dyslexic officiant in the kingdom or republic, who authored the batch of documents to be deciphered!
With a particular or “flyspeck” handwriting, it does indeed start to require a “dictionary” or estimates and probabilities. A fifty-letter set (lowercase + uppercase) to decipher, a mini-index for our Rosetta Stone, seems to be the minimum. I imagine that this is anyway the first common step in most methods before deciphering unknown text? Even if unconscious, this step takes a little time.
In the video, it mentions two hours of learning (I think) for a small project (less than 20 documents → LLMs). That’s why using an already proven AI model for French handwriting is suggested before training a new custom model. Transkribus and other service platforms are more presentable, offering more advanced functions at the click of a button, but even there, you have to adapt or correct the deciphering.
In Alsace, there are parish registers in German (often Protestant), sometimes a jumble of “Latin-French-Alsatian”. By default, a “French”-oriented model will have difficulties, just like anyone discovering the documents in the region. Moreover, for another project, I had to translate a batch of administrative documents in German (corpus). They all had more or less the same structure, but not always… Anyway, quickly one looks for help, a tool for these repetitive tasks, preferably semi-automatic. We obviously go behind for the final review.
In my usage, this is close to the help manual (or assistant) of 1990s/2000s software… It’s a common base that we feed with our local inputs. It’s difficult to provide a list of suggestions without having fed the common base a minimum.
In my opinion, a change of priest is above all a change…
I’m almost sure that these tools can detect this change and adapt the model (smooth transition).
Thank you.
But in practice, what tools are we talking about?
LLMs, if I’m not mistaken, ChatGPT, Gemini, … for which there is no learning (proprietary). I have already tried the two mentioned above, the first impression is Wow 
then upon rereading, you see that there are quite a few hallucinations 
Same with Mistral! You can clearly sense the intent to lean into imagination and a pseudo-creative process. After a quick and fairly strict initial framing, you generally avoid many “deviations.”
Currently, we still get the impression that some “big/major” AI projects are flirting with fraud!
For the analysis of handwriting (or printed/typed), there are models (LLM tools) that are less “universal/mainstream”.
This may explain certain Microsoft projects and directions! Alsatian (Baden German, Swiss German, or Yiddish) is the new Enigma across the Atlantic…
It’s often the same impression, that of a salesperson who must have an answer for everything to drive their business. Except that, as far as we know, the universality of the French language is a myth.
Transcriptions and translations of French sources are a headache for Family Search or most American companies. ChatGPT or Gemini still struggle with cultural subtleties. I’ve fully understood that the French language is a bit like Scottish to English speakers: Klingon (or Esperanto) with Master Yoda’s phrasing! 
To refocus on AI models. I was using mistral for reformatting and transkribus for transcription for convenience (interface, cost/time, etc.). If mistral indeed uses telkia models for its transcriptions, while adding its own layers of hallucinations, then why not test a Telkia LLM locally?
We’re closer to potential additional features for gramps-web, but on localhost for gramps… 
Raw transcription output via the API is not easy to optimize through a gramplet and the default options (prompt/query).
So, outside of the Gramps context, try with transkribus?
… or arkindex [Full Document Processing (Classification, Structure Analysis, Text Recognition, Entity Recognition)], eScriptorium [Text Recognition (printed and handwritten), Table Recognition], etc.
| Arkindex | Transkribus | eScriptorium | |
|---|---|---|---|
| Text Recognition |  |
|
|
| Text Line Detection | |
|
|
| Table Detection and Recognition | |
|
 |
| Form Recognition | |
|
|
| Page Classification | |
|
|
| Custom Structure Recognition | |
|
|
| Named Entity Recognition | |
|
|
| Key Value Extraction | |
|
|
| Open-source | |
|
|
| Self-hosting | |
|
|
| API Access | Full access | Limited to download and processing | Limited |
| Exporting Trained Models | |
|
|
| Custom Algorithm Integration | |
|
|
Thanks, I have a few videos to watch 
I tried two documents (notarial deeds from 1874, reasonably well-written).
First one: a few words found.
Second one: nothing found!
In short, disappointing (and worse than Gemini,…)
The few basic models on Ocelus’s ‘web’ demo are very limited compared to the capabilities of their custom models. Additionally, I find that they should provide some guidance for user-side preprocessing before sending the document for testing. For example, I often get better results after applying a filter like “negative via Gimp”. There are indeed “educational” blogs
https://harmoniseatr.hypotheses.org/3271
https://harmoniseatr.hypotheses.org/2866
etc.
yet this is rarely illustrated in Telkia’s articles.
https://teklia.com/fr/nos-solutions-fr/cas-dusage/reconnaissance-automatique-de-textes-multilingues/
https://teklia.com/blog/202212-atr/
… and on top of that, you have to correct the false positives! 