ஓலைச் சுவடி · Olai Suvadi · research
ஓலையிலிருந்து உரைக்கு
From palm leaf to readable text
ஓலையிலிருந்து உரைக்கு எந்த வழியில் செல்வது என்பதை கட்டம் கட்டமாகக் காட்டுகிறது — எந்தக் கட்டம் முடிந்தது, எதை மாற்றி மீண்டும் யோசிக்க வேண்டியது என்பதும் உடன்.
Six stages, and one of them is a correction. Character segmentation looked like the blocker — we measured a detector finding 72 text regions on a leaf whose text has no word boundaries at all. Then it turned out to be the wrong question: modern handwritten text recognition reads a whole line at once and never cuts it into letters. Tap through the stages; the fourth is the one that changed.
வாசிப்பு அறைThe reading room
Working today, with no model involved. Every line below was read off the leaf by a scholar and published by CICT — and until now there was no way to search any of it. Each transcription is shown over the strip of palm leaf it came from, so you can always check the text against the thing itself.
ஏற்றுகிறது…
இயந்திரம் படிக்கTeaching a machine to read it
The 1,329 lines above are 1,329 of roughly 100,000 leaves. The rest need a model, and this is where that stands — stage by stage, including one stage we got wrong.
எழுத்துப் பிரித்தல்Character segmentation
தேவையில்லை· Not needed any moreCut each line into individual letters — the old way, and no longer necessary.
WE MEASURED THIS AND THEN FOUND IT WAS THE WRONG QUESTION. Running a standard detector (CRAFT) over four real manuscripts and a printed Tamil control: the control gave 15 regions for 16 printed words — it found the words. A Tholkappiyam leaf of roughly five lines gave 72, a Tevaram leaf 133, numbers that correspond to nothing, because manuscript Tamil has no spaces and the glyphs touch. That is a real result, and it is why the old pipeline fails. But modern handwritten text recognition does not cut lines into characters at all: Kraken reads a whole line with a recurrent network trained with CTC, TrOCR with a transformer. Give either one line images and their transcriptions and the segmentation problem simply does not arise. The 97% figures in the older literature are character classification on hand-cut crops, measured after this step — which is why they never transferred to a page.
1,330
சரிபார்க்கப்பட்ட குறள்
verified kurals in CICT's ground-truth corpus — 100% complete
865
சுவடிக் கட்டுகள்
manuscript bundles digitised at CICT, ~100,000 leaves
2
பயிற்சி வடிவங்கள்
HTR export formats published: PageXML (Kraken) and JSONL (TrOCR)
This page said something wrong until 30 October 2026. It said no line-level Tamil ground truth existed publicly and that the next step was an agreement with an archive rather than a training run. Both were wrong. The Central Institute of Classical Tamil publishes a Tirukkural ground-truth corpus — 1,330 of 1,330 kurals, complete, labelled as Tamil HTR training specimens — and exports it as PageXML for Kraken and JSONL for TrOCR, which are precisely the inputs those toolchains take. The corpus was built so a model could be trained on it. So this is engineering with a known starting point, not research waiting on data, and the correction is left visible here rather than quietly edited out.