The Writing Is on the Wall for Handwriting Recognition
newsletter.dancohen.org71 points by speckx 7 days ago
71 points by speckx 7 days ago
> If AI can diminish some of the monotony of research, perhaps we can spend more time thinking, writing, playing piano, and taking walks — with other people.
Whenever any progress is made, this is the logical conclusion. And yet, those who decide about how your time is being used, have an opposing view.
I feel that we’re reaching a limit to our context switching. Any further process improvements or optimizations will be bottlenecked on humans. And I don’t think AI will help here as jobs will account for that and we’ll have to do context switching on even broader and more complex scopes.
Great post and amazing progress in this field! However, I have to wonder if some of these letters were part of the training data for Gemini, since they are well-known and someone has probably already done the painstaking work of transcribing them...
Most likely, and probably inferring the structure on texts with "similar" writing forms. Tried with my handwriting (in italian) and the performance wasn't that stellar. More annoyingly, it is still a LLM and not a "pure" OCR, so some sentences were partially rephrased with different words than the one in the text. This is crucially problematic if they would be used to transcribe historical documents
> Tried with my handwriting (in italian) and the performance wasn't that stellar.
Same here, for diaries/journals written in mixed Swedish/English/Spanish and with absolutely terrible hand-writing.
I'd love for the day where the writing is on the wall for handwriting recognition, which is something I bet on when I started with my journals, but seems that day has yet to come. I'm eager to get there though so I can archive all of it!
I have a personal corpus of letters between my grandparents in WW2. My grandfather fighting in Europe and my grandmother in England. The ability of Claude and ChatGPT to transcribe them is extremely impressive. Though I haven’t worked on them in months and this uses older models. At that time neither system could properly organize pages though and chatGPT would sometimes skip a paragraph.
I've also been working on half a dozen crates of old family letters. ChatGPT does well with them and is especially good at summarizing the letters. Unfortunately, all the output still has to be verified because it hallucinates words and phrases and drops lines here and there. So at this point, I still transcribe them by hand, because the verification process is actually more tiresome than just typing them up in the first place. Maybe I should just have ChatGPT verify MY transcriptions instead.
It helps when you can see the confidence of each token, which downloadable weights usually gives you. Then whenever you (your software) detects a low confidence token, run over that section multiple times to generate alternatives, and either go with the highest confidence one, or manually review the suggestions. Easier than having to manually transcribe those parts at least.
Possibly, but given it can also read my handwriting- which is much, MUCH worse than Boole’s - with better accuracy than any human I’ve given it to- that’s probably not the explanation.
Shhhhh no one cares about data contamination anymore.
Then write something down yourself and upload a picture to gemini.google.com or chatgpt. Hell, combine it. Make yourself a quick math test, print it, solve with pen and ask these models to correct it.
They're very good at it.
Any self-hosted open source solution? I would like to digitize my paper notebooks but I do not want to use anything proprietary or that uses external services. What is the state of the art on the FOSS side?
Ideally something that I can train with my own handwriting. I had a look at Tesseract, wondering if there’s anything better out there.
Regular handwriting there are many.
Historical handwriting, Gemini 3 is the only one which gave a decent result on a 19th century minutes from a town court in Northern Norway (Danish gothic handwriting with bleed through). I'm not 100% sure it's correct, but that's because it's so dang hard to read it to verify it. At least I see it gets many names, dates and locations right.
I've been waiting a long time for this.
> Regular handwriting there are many.
Please share. I am out of the loop and my searches have not pointed me to the state of the art, which has seen major steps forward in the past 3 or 4 years but most of it seems to be closed or attached to larger AI products.
Is it even still called OCR?
Totally not what you asked, but making an OCR model is a learning exercise for AI research students. Using the Kaggle-hosted dataset https://www.kaggle.com/datasets/landlord/handwriting-recogni... and a tutorial, eg https://pyimagesearch.com/2020/08/17/ocr-with-keras-tensorfl... you can follow along and train your own OCR model!
I became convinced of this after the release of KuroNet: https://arxiv.org/pdf/1910.09433 (High-quality OCR of Japanese manuscripts, which look almost impossible to read.)
Maybe for English, for the other human languages I use, it is still kind of hit and miss, just like speaking recognition, even with English it suffices to have an accent that is off the standard TV one.
As always, this depends on the amount of training data available. Japanese is another success story: https://digitalorientalist.com/2020/02/18/cursive-japanese-a...
> "transmitted": In the second line of the body, the word "transmitted" is crossed out in the original text
Am I nuts or is this wrong, not “perfect”?
It doesn’t look crossed out at all to me in the image, just some bleeding?
Still very impressive, of course
Call me when it can do Russian Cursive.
Seems to do an OK job:
https://g.co/gemini/share/e173d18d1d80
This is a random image from Twitter with no transcript or English translation provided, so it's not going to be in the training data.
No, transcription has nothing to do with written text, it guessed few words here and there but not even general topic. That's doctors note about patient visit, beginning with "Прием: состояние удовл., t*, но кашель / patient visit: condition is OK, t(temperature normal?) but coughing". But unreadable doctors handwriting is a meme...
That's Gemini 2.5 Flash btw
The result from Gemini 3 Pro using the default media resolution (the medium one): "(Заголовок / Header): Арсеньев (Фамилия / Surname - likely "Arsenyev")
Состояние удовл-
t N, кожные
покровы чистые,
[л/у не увел.]
В зеве умерен. [умеренная]
гипер. [гиперемия]
В легких дыха-
ние жесткое, хрипов
нет. Тоны серд-
[ца] [ритм]ичные.
Живот мяг-
кий, б/б [безболезненный].
мочеисп. [мочеиспускание] своб. [свободное]
Ds: ОРЗ [или ОРВИ]" and with the translation: "Arsenyev
> Here’s Transkribus’s best guess at George’s letter to Maryann, above:
Transkribus got a new model architecture around the corner and the results look impressive. Not only for trivial cases like text, but also for table structures and layouting.
Best of all, you can train it on your own corpus of text to support obscure languages and handwriting systems.
Really looking forward to it.
Surely the true prize is to be able to ditch computers altogether and just write with pencil on paper.
If I went back in time to the 90s when I was doing my PhD I would absolutely blow my mind with how well handwriting OCR works now.
It feels unbelievable that in Europe literacy rate could be 10% of lower. Then I look at documents even as young as 150 years... fraktur, blackletter, elaborate handwritting. I guess I'm illiterate now.
Hopefully next generations will feel the same about legal contracts, law in general, and Java code bases. They're incomprehensible not because of fonts but because of unfathomable complexity.
You can learn fraktur or blackletter in a day and cyrillic in a few days, if you already know the latin alphabet.
> learn fraktur or blackletter in a day and cyrillic in a few days
Not a chance, sorry.
My question for OCR automation is always which digits within the numbers being read are allowed to be incorrect?
The writing is on the wall for handwriting. Zoomers use speech recognition or touchscreen keyboards, millennials use keyboards. Boomers use pens
I call out the Lindy effect. Handwriting survived printed characters, typewriters, and the last 50-70 years of computers and keyboards, it will survive this too.