rapidocr-spanish

Two text-recognition checkpoints for Spanish documents, fine-tuned for RapidOCR from PP-OCRv5_latin_mobile_rec with a reduced Spanish vocabulary.

Everyday Spanish paperwork is not just clean paragraphs: identity cards, payslips, certificates, receipts, family records and engineering drawings each bring their own fonts, layouts and symbol sets. These checkpoints target that mix.

Files

file	size	use
spanish_rec_v18.onnx	7.7 MB	default checkpoint
spanish_rec_v19.onnx	7.7 MB	alternative for receipts and records
ppocr_keys_es.txt	116 lines	vocabulary for both; required

Load each model with ppocr_keys_es.txt. Pairing a model with a different dictionary corrupts the CTC decoding: indices no longer match characters.

Vocabulary

116 characters: lowercase and uppercase Spanish letters including ñ and the accented vowels, digits, common punctuation, € £ º ª, and the symbols the multilingual Latin vocabulary had but naive pruning drops: < > _ Ø | · □ Φ ≠. The < matters in practice — machine-readable zones on ID cards are full of it, and without it those lines cannot be transcribed at all.

Metrics

Held-out gold of 3,079 crops (44,947 reference characters) from 14 Spanish document types, compared against the stock Latin checkpoint on the same crops:

model	CER micro	CER macro
PP-OCRv5_latin_mobile_rec (stock)	0.1745	0.1412
spanish_rec_v18	0.1476	0.1238

micro is the standard character-error rate (edit distance over reference characters); macro is the mean of per-crop error rates and flatters a model on short fields. The checkpoints win on 10 of the 14 document types; blueprints −19%, CVs −18%, health-insurance slips −25%.

Per-type routing between the two checkpoints adds a further ~0.8% relative with no regression, because an unknown document type falls back to the default model.

Training data

• Real crops mined from 48 worker files covering 17 document types (identity cards, CVs, certificates, receipts, health-insurance slips, driving licences, family records, transfer deeds, experience records), rendered at 300 DPI.
• Synthetic Spanish lines composed over real scanned-paper backgrounds.
• Labels produced by a three-model vision consensus and refined by repeated passes; a fraction was reviewed by hand. This is the main limitation, see below.
• Fine-tuned from the Latin checkpoint with the head pruned to the 116-character vocabulary, low learning rate (5e-5) and 8 epochs. Reusing a from-scratch schedule on a converged model wrecks it, which is worth knowing if you retrain.

No client documents, names or identifiers are part of this release.

Usage

from rapidocr import RapidOCR, LangRec

ocr = RapidOCR(params={
    "Rec.lang_type": LangRec.ES,
    "Rec.model_path": "spanish_rec_v18.onnx",
    "Rec.rec_keys_path": "ppocr_keys_es.txt",
})
result = ocr("document.png")
print(result.txts)

Consumers

These checkpoints are consumed by rapidocr-worker, a Rust daemon on ONNX Runtime that serves the full PP-OCR pipeline over a UNIX socket with bounded memory (about 74 MB per worker after load, no Python GIL contention).

It needs no code change to use them: the checkpoint goes in --rec-model and the vocabulary in --keys.

rapidocr-worker \
  --rec-model spanish_rec_v18.onnx \
  --keys ppocr_keys_es.txt

--keys is mandatory here because the checkpoint carries no embedded character metadata, and pairing it with a different vocabulary corrupts the CTC decoding silently.

Limitations

• Label noise. Part of the gold set was labelled by vision models without human review, so the measured error includes some reference error. Treat the numbers as a comparison between checkpoints, not as an absolute quality claim.
• Small text. On drawings rendered at 300 DPI the error is much higher (CER ≈ 0.30 micro) because the detector also finds dimension marks and tiny labels. This is the weakest area.
• Tables and formulas are not covered by these checkpoints; use PPStructureV3 for table HTML and LaTeX.
• No paragraph model. The recognizer reads one line at a time; merged multi-line regions are recognized poorly.
• Mobile-size architecture (7.7 MB). A larger model fine-tuned on the same data measured worse, so capacity was not the bottleneck here.

License

Apache-2.0, inherited from PaddleOCR's released weights.