Empirical Findings · September 2026 Run

OCR Benchmark Results & Error Taxonomy

Complete metrics, latency measurements, failure categorization, and visual evidence across all 10 evaluated language fixtures.

OCR Benchmark Results & Error Taxonomy?

Full results from the official September 20, 2026 test execution on Node.js v22 with Tesseract.js 5.1.1 (LSTM OEM 1, PSM 6). * Scripts without space word boundaries (Chinese, Japanese, Hindi, Arabic) report CER and exact key-field accuracy. WER is only evaluated for space-delimited Latin and Hangul scripts.

Last updated 2026-09-20

Complete Engine Benchmark Scorecard

Full results from the official September 20, 2026 test execution on Node.js v22 with Tesseract.js 5.1.1 (LSTM OEM 1, PSM 6).

10-Language Benchmark Metric Breakdown
Language / Fixture	Pack	CER	WER	Key Fields	Numeric Exact	Email Exact	URL Exact	Verdict
العربية OCR-AR-2026-010	`ara`	0.638	—*	0% (0/11)	0%	0%	0%	FAIL
Deutsch OCR-DE-2026-007	`deu`	0.302	0.738	100% (11/11)	100%	100%	100%	PASS
English OCR-EN-2026-001	`eng`	0.178	0.265	93.8% (15/16)	100%	100%	100%	WARN
Español OCR-ES-2026-005	`spa`	0.527	1.200	62.5% (10/16)	58.3%	0%	100%	FAIL
Français OCR-FR-2026-006	`fra`	0.402	0.809	83.3% (10/12)	87.5%	100%	100%	FAIL
हिन्दी OCR-HI-2026-009	`hin`	0.607	—*	22.2% (2/9)	20%	0%	0%	FAIL
日本語 OCR-JA-2026-003	`jpn`	0.667	—*	27.8% (5/18)	33.3%	100%	100%	FAIL
한국어 OCR-KO-2026-004	`kor`	0.627	0.860	63.6% (7/11)	85.7%	0%	0%	FAIL
Português OCR-PTBR-2026-008	`por`	0.431	0.900	70% (7/10)	83.3%	0%	100%	FAIL
简体中文 OCR-ZH-2026-002	`chi_sim`	0.668	—*	18.2% (2/11)	40%	0%	0%	FAIL

* Scripts without space word boundaries (Chinese, Japanese, Hindi, Arabic) report CER and exact key-field accuracy. WER is only evaluated for space-delimited Latin and Hangul scripts.

Where Browser OCR Struggles: Failure Taxonomy

High-contrast digital scans perform adequately in browser WebAssembly OCR. However, real-world creator and photo workflows present systematic failure modes:

Category 1

CJK Inter-character Spacing & Segmentation

Observed in: Simplified Chinese (OCR-ZH-2026-002, 18.2% key fields) and Japanese (OCR-JA-2026-003, 27.8% key fields).

The underlying Tesseract LSTM engine inserts ASCII spaces between contiguous Han and Kana glyphs (e.g. outputting 社区服务中心 instead of 社区服务中心). This degrades exact substring searches and form autofill. While space-stripping helps diagnostics, character substitutions (such as confusing '二' with '三' or '醤油' with '普油') remain prevalent in natural lighting.

Category 2

Non-Latin Scripts & Complex Shapers

Observed in: Arabic (OCR-AR-2026-010, 0% key fields) and Hindi (OCR-HI-2026-009, 22.2% key fields).

Arabic cursive joining rules and Devanagari vertical conjuncts fail on scene-photographed text. On our Arabic clinic fixture, zero of 11 scored fields matched; cursive headlines collapsed into disconnected glyph fragments. On Hindi notices, Latin phone numbers survived but the Devanagari headline was entirely lost.

Category 3

Photographic Noise & Graphic Elements

Observed in: Japanese ramen menu (OCR-JA-2026-003) and English desk invoice (OCR-EN-2026-001).

Without dedicated layout segmentation or picture masking, the OCR engine treats high-contrast texture in photographs as typography. High-resolution photos of food were transcribed into ghost kanji strings (such as 隊暫要衣 and ワニ). Similarly, decorative desk items (mugs, foliage) generated garbage text tokens around the borders.

Category 4

Thermal Receipt Creases & Faded Print

Observed in: Spanish supermarket receipt (OCR-ES-2026-005, 62.5% key fields).

Thermal receipt paper creases cause horizontal character slicing. Single-digit drops were observed in price columns: 2,85 € had the digit 5 dropped and the Euro sign misread as a cent mark (2,8¢). The tax summary line was completely lost, though overall document totals survived.

Category 5

Stylized Display Headers & Sizing Variance

Observed in: English office invoice (OCR-EN-2026-001, 93.8% key fields, WARN).

Tesseract is tuned for uniform body text. When documents feature heavy, oversized display titles (such as a 48pt bold 'INVOICE' header), the character strokes exceed the LSTM line segmenter's expected aspect ratio and are ignored, even while 8pt table items directly below are read with 100% precision.

Complete Fixture Evidence Archive

Examine the original rendered photograph, ground truth excerpt, recognized OCR output, and exact error notes for every fixture in the benchmark suite.

OCR-AR-2026-010 test fixture: Clinic information sheet and appointment card — Fixture **OCR-AR-2026-010** · Controlled scene

Key Field Accuracy 0%

CER (Error Rate) 0.638

Expected Ground Truth (sample)

مركز النور الطبي
المرجع: OCR-AR-2026-010
التاريخ: 20-09-2026
العنوان: شارع الملك فهد، مبنى 18، الرياض
الهاتف: +966 11 555 2480

Recognized Output (Tesseract.js)

✓ What survived:

Phone digit sequence partly visible (+966 11 555 2480).

✕ Failure observations:

0 of 11 key fields matched (0% key-field accuracy).
Arabic cursive ligatures and shaping lost on two-document desk scene.
Latin email domain truncated and colon corrupted: 'hitps:/alnourmed:s2'.

OCR-DE-2026-007 test fixture: Technical product sheet on clipboard — Fixture **OCR-DE-2026-007** · Controlled scene

Key Field Accuracy 100%

CER (Error Rate) 0.302

Expected Ground Truth (sample)

Geräteübersicht
Dokument-Nr.: OCR-DE-2026-007
Produkt: MobileScan Pro 12
Hersteller: DatenWerk GmbH
Adresse: Lindenstraße 44, 10117 Berlin

Recognized Output (Tesseract.js)

✓ What survived:

100% key-field accuracy (11 of 11 fields exact match).
Umlauts and German typography exact: 'Geräteübersicht', 'Auflösung', 'Lindenstraße'.
Exact decimal comma preserved in weight: '1,35 kg'.
Serial code 'MSP12-DE-884209', email 'support@datenwerk.de', and URL 'https://datenwerk.de/manual' fully matched.

✕ Failure observations:

High-contrast printed paper succeeded; camera photos with background clutter or low DPI require human review.

OCR-EN-2026-001 test fixture: Office invoice photographed on a desk — Fixture **OCR-EN-2026-001** · Controlled scene

Key Field Accuracy 93.8%

CER (Error Rate) 0.178

Expected Ground Truth (sample)

INVOICE
Invoice No: OCR-EN-2026-001
Date: 2026-09-20
Bill To: Harbor Metrics LLC
Address: 1450 Market Street, Suite 810, San Francisco, CA 94103

Recognized Output (Tesseract.js)

✓ What survived:

15 of 16 key fields matched (93.8% key-field accuracy).
Financial figures exact: Subtotal $134.00, Tax $11.06, Grand Total $145.06.
Itemized pricing exact: $49.50, $99.00, $35.00.
Email 'billing@harbormetrics.com' and payment URL matched.

✕ Failure observations:

Large stylized title 'INVOICE' missed due to oversized display font weight.
Desk background mug and plant pixels caused noise in border tokens.

OCR-ES-2026-005 test fixture: Crumpled supermarket receipt — Fixture **OCR-ES-2026-005** · Controlled scene

Key Field Accuracy 62.5%

CER (Error Rate) 0.527

Expected Ground Truth (sample)

SUPERMERCADO LA PLAZA
Ticket: OCR-ES-2026-005
Fecha: 20/09/2026 18:42
Dirección: Calle Mayor 18, 28013 Madrid
Tel: +34 91 555 24 80

Recognized Output (Tesseract.js)

✓ What survived:

10 of 16 key fields matched (62.5% key-field accuracy).
Total payable amount '13,42 €' and ticket number matched.
Store website 'https://supermercados-sol.es' matched.

✕ Failure observations:

Thermal paper wrinkles caused digit loss: '2,85 €' became '2,8¢' (5 dropped, € to ¢).
Tax breakdown 'IVA: 1,22 €' line completely missed due to faint thermal print.
Accents flattened: 'Dirección' became 'Direccion'.

OCR-FR-2026-006 test fixture: Cultural event flyer on cork board — Fixture **OCR-FR-2026-006** · Controlled scene

Key Field Accuracy 83.3%

CER (Error Rate) 0.402

Expected Ground Truth (sample)

Festival de Lecture
Référence : OCR-FR-2026-006
Date : 20 septembre 2026
Lieu : Maison de la Culture, 12 rue Victor Hugo, 69002 Lyon
Programme

Recognized Output (Tesseract.js)

✓ What survived:

10 of 12 key fields matched (83.3% key-field accuracy).
Reference code 'OCR-FR-2026-006' and Lyon venue address matched.
Accents largely retained on body lines ('Référence', 'numérique').

✕ Failure observations:

Phone number spacing lost and final digit dropped: '04 72 55 18 90' became '047255189'.
URL missed a slash: 'https:/festivallecture.fr'.

OCR-HI-2026-009 test fixture: Public information poster — Fixture **OCR-HI-2026-009** · Controlled scene

Key Field Accuracy 22.2%

CER (Error Rate) 0.607

Expected Ground Truth (sample)

जन सूचना
दस्तावेज़ संख्या: OCR-HI-2026-009
तारीख: 20 सितंबर 2026
विषय: डिजिटल सेवा शिविर
स्थान: सामुदायिक भवन, सेक्टर 12, नोएडा

Recognized Output (Tesseract.js)

✓ What survived:

Helpline phone '+91 98765 43210' and support email 'help@digitalseva.in' exact.

✕ Failure observations:

Devanagari conjuncts and matras collapsed under generic scene lighting.
Notice header and dates failed exact substring match (22.2% key fields).

OCR-JA-2026-003 test fixture: Restaurant menu board — Fixture **OCR-JA-2026-003** · Controlled scene

Key Field Accuracy 27.8%

CER (Error Rate) 0.667

Expected Ground Truth (sample)

本日のおすすめ
店舗コード：OCR-JA-2026-003
営業時間 11:30-22:00
住所：東京都新宿区西新宿2-8-1
電話：03-5550-2481

Recognized Output (Tesseract.js)

✓ What survived:

Email 'info@tokyoramen.jp' and URL 'https://tokyoramen.jp' exact.
Restaurant hours and shop identifier matched.

✕ Failure observations:

Yen price format broken: '1,080円' became '1.080' (comma to period, 円 symbol lost).
Kanji substitution: '醤油' (soy sauce) recognized as '普油'.
Food photograph background noise read as ghost kanji ('隊暫要衣', 'ワニ').

OCR-KO-2026-004 test fixture: Cafe poster taped to glass — Fixture **OCR-KO-2026-004** · Controlled scene

Key Field Accuracy 63.6%

CER (Error Rate) 0.627

Expected Ground Truth (sample)

가을 시즌 안내
문서번호: OCR-KO-2026-004
행사 기간: 2026년 9월 20일 - 2026년 10월 5일
매장명: 모닝브루 카페
주소: 서울특별시 강남구 테헤란로 152

Recognized Output (Tesseract.js)

✓ What survived:

7 of 11 key fields matched (63.6% key-field accuracy).
Won currency prices (6,300원) and promotional dates recognized.

✕ Failure observations:

Public ID corrupted: 'OCR-KO-2026-004' read as '008-0-2026-004' (O/0 confusable).
Latin email and website links on the poster collapsed due to glass reflection glare.

OCR-PTBR-2026-008 test fixture: Shipping label on cardboard parcel — Fixture **OCR-PTBR-2026-008** · Controlled scene

Key Field Accuracy 70%

CER (Error Rate) 0.431

Expected Ground Truth (sample)

ETIQUETA DE ENVIO
Código: OCR-PTBR-2026-008
Destinatário: Camila Souza
Rua das Palmeiras, 245, Apto 32
Bairro Centro

Recognized Output (Tesseract.js)

✓ What survived:

7 of 10 key fields matched (70.0% key-field accuracy).
Package tracking code 'BR420998771' and recipient city 'Curitiba - PR' matched.
Store website 'https://lojanovaera.com.br' matched.

✕ Failure observations:

Postal code digit misread: 'CEP 80010-120' had digit '8' recognized as 's' ('ceps0010-120').
Package code 'OCR-PTBR-2026-008' was truncated to 'OCR-PTBR-2026-0 2'.

OCR-ZH-2026-002 test fixture: Community notice pinned to a wall — Fixture **OCR-ZH-2026-002** · Controlled scene

Key Field Accuracy 18.2%

CER (Error Rate) 0.668

Expected Ground Truth (sample)

社区服务中心通知
文件编号：OCR-ZH-2026-002
发布时间：2026年9月20日
主题：周末设备维护安排
维护时间：2026年9月22日 09:00-17:30

Recognized Output (Tesseract.js)

✓ What survived:

Phone '021-5558-1266' and public ID 'OCR-ZH-2026-002' exact.

✕ Failure observations:

Tesseract LSTM inserted spaces between contiguous Han characters ('社区服务中心').
Support email corrupted from 'support@example.cn' to 'support@example:cn'.
URL corrupted to 'htips:J/examplecnne'.

Cite this benchmark

Empirical research published by CreatorPrivacyKit Labs is freely citeable with attribution under CC BY 4.0.

APA / Text Format

CreatorPrivacyKit Labs. (2026). OCR Benchmark Results & Error Taxonomy (Version 1.0) [Data set]. CreatorPrivacyKit. https://creatorprivacykit.com/labs/ocr-benchmark/results/

View BibTeX entry

@misc{creatorprivacykit_2026_ocr_benchmark,
  title = {OCR Benchmark Results & Error Taxonomy},
  author = {{CreatorPrivacyKit Labs}},
  year = {2026},
  version = {1.0},
  howpublished = {\url{https://creatorprivacykit.com/labs/ocr-benchmark/results/}},
  note = {Tested 2026-09-20. In-browser Tesseract.js 5.1.1 multilingual evaluation}
}

Data sources: Dataset landing & download · results.json · results.csv