Synthetic Tamil OCR Benchmark Dataset
- The synthetic Tamil OCR benchmark is derived from OPUS text, rendering 7,000 line images with six curated Google Fonts for controlled zero-shot evaluation.
- It focuses on printed Tamil text with clean black-on-white images, omitting noise and layout variations for clear performance insights.
- Benchmark protocols compare OCR engines using metrics like CER and WER, highlighting the challenge of word reconstruction despite high character accuracy.
Searching arXiv for papers relevant to synthetic Tamil OCR benchmarking datasets. Synthetic Tamil OCR benchmarking datasets are benchmark resources for evaluating optical character recognition on Tamil text using synthetically generated image–text pairs, typically under controlled rendering conditions and with line- or word-level transcriptions. Within the cited literature, the most direct instance is the synthetic Tamil benchmark introduced for zero-shot OCR comparison on printed Tamil text, where line images are rendered from OPUS OpenSubtitles Tamil text using Pillow and six Tamil Google Fonts, yielding a 7,000-sample benchmark designed for fair comparison with a matched synthetic Sinhala dataset (Jayatilleke et al., 24 Jul 2025). A broader multilingual context is provided by an Indic synthetic OCR benchmark containing Tamil among 23 Indic languages, with 3,111 synthetic Tamil line images distributed across three synthetic quality classes and additional real, noise-added-real, and scene-text Tamil subsets (Saini et al., 2022). Related work also supplies methodological components rather than Tamil-specific benchmark releases: Unicode-aware synthetic OCR generation pipelines for low-resource languages (Malik et al., 22 Jan 2026), multilingual Indic OCR benchmarks and CTC-based recognition protocols including Tamil (Mathew et al., 2022), legacy-font Tamil OCR rendering and evaluation strategies (Vasantharajan et al., 2021), and narrower character-level Tamil resources such as an MNIST-like synthetic Tamil vowel dataset (Annamalai, 2020). By contrast, some low-resource OCR papers are only indirectly relevant because they provide methodological guidance on script-specific OCR error analysis without any Tamil data (Ignat et al., 2022).
1. Emergence and scope
The most explicit synthetic Tamil OCR benchmarking dataset in the cited corpus appears in the study of zero-shot OCR accuracy on Sinhala and Tamil, where a “novel synthetic Tamil OCR benchmarking dataset” is introduced to enable a controlled comparison across the two low-resource languages (Jayatilleke et al., 24 Jul 2025). Its role is not ancillary: the authors report that they could find an existing synthetic Sinhala benchmark but not a comparable Tamil benchmark, and therefore created a synthetic Tamil dataset so that Tamil and Sinhala could be evaluated under a matched synthetic setting. The dataset targets printed Tamil text rather than scene text, handwriting, page layout understanding, or richly degraded scans.
A second important source is the multilingual “OCR Synthetic Benchmark Dataset for Indic Languages,” which presents “the largest publicly available synthetic OCR benchmark dataset for Indic languages,” with “a total of 90k images and their ground truth for 23 Indic languages,” explicitly including Tamil (Saini et al., 2022). This resource is not Tamil-only, but it contains a documented Tamil subset with language-wise image counts, font lists, and quality classes. In this broader sense, a synthetic Tamil OCR benchmarking dataset may denote either a Tamil-specific benchmark or the Tamil partition of a multilingual Indic benchmark.
Other cited works frame the synthetic Tamil benchmarking problem from adjacent directions. “Towards Deployable OCR models for Indic languages” includes Tamil in a 13-language OCR study and discusses synthetic-data rendering methodology, though its synthetic experiments are not run on Tamil (Mathew et al., 2022). “synthocr-gen” offers a language-agnostic synthetic OCR generator with Unicode normalization, script validation, segmentation at multiple granularities, and augmentation, but does not itself release a Tamil benchmark (Malik et al., 22 Jan 2026). “Adapting the Tesseract Open-Source OCR Engine for Tamil and Sinhala Legacy Fonts” shows how rendered Tamil legacy-font pages can support OCR training and per-font evaluation, again without packaging a standardized Tamil image benchmark (Vasantharajan et al., 2021). This suggests that the literature contains both explicit synthetic Tamil benchmarks and a wider methodological substrate for constructing them.
2. Tamil-specific benchmark instantiations
The Tamil-specific synthetic benchmark introduced in the zero-shot study is constructed from OPUS, specifically OpenSubtitles Tamil-English v2024 (Jayatilleke et al., 24 Jul 2025). The workflow is given as: text retrieval from OPUS, line segmentation, length-based filtration, random selection of records, and text-to-image rendering. Tamil text is divided line by line, and the preprocessing is restricted to Tamil characters only. This yields 2,437,960 records, after which only texts longer than 40 characters are retained, producing 222,658 records. From these, 7,000 samples are randomly selected so that the Tamil benchmark approximately matches the 6,969-image Sinhala synthetic benchmark used in the same study (Jayatilleke et al., 24 Jul 2025).
The benchmark is line-level. The paper states that the corpus was split line by line, that text strings were rendered into individual images, and that OCR output was compared against the corresponding reference string. The ground truth is therefore image-text pairing at the line level, with no mention of character bounding boxes, word boxes, polygon annotations, or page-structure labels. Rendering is performed with Pillow. The implementation computes image dimensions dynamically from the text bounding box, renders black text on a white background, and centers the text horizontally and vertically for visual consistency and quality (Jayatilleke et al., 24 Jul 2025). The paper further notes that empty reference strings and empty generated features were removed during post-processing.
The font inventory for this Tamil-specific benchmark is limited but explicit. The authors examined 17 Tamil fonts from Google Fonts and selected 6 unique font families after visual inspection because some appeared too similar. The six retained fonts are Hind Madurai, Noto Serif Tamil, Kavivanar, Noto Sans Tamil, Pavanam, and Anek Tamil (Jayatilleke et al., 24 Jul 2025). The font-assignment function was designed to ensure a proportional distribution of input text records across the selected font files. The result is therefore a clean, printed, centered, line-image benchmark with controlled typographic variation but no reported degradation, noise, or background complexity.
The larger multilingual Indic benchmark supplies a more extensive Tamil subset. Its Tamil synthetic image counts are Class 1 = 358, Class 2 = 1425, and Class 3 = 1328, for a total of 3111 synthetic Tamil line images (Saini et al., 2022). The same table reports Real Class 1 = 240, Real Class 2 = 796, Real Class 3 = 181, Noise-Added-Real = 1217, and Scene-Text = 40 for Tamil (Saini et al., 2022). This resource therefore situates synthetic Tamil OCR benchmarking within a mixed synthetic/real evaluation context, although it does not define a Tamil-only train/validation/test split or a dedicated benchmark partition.
3. Data generation and representation
The Tamil-specific zero-shot benchmark employs a simple rendering pipeline centered on OPUS text selection and Pillow-based image synthesis (Jayatilleke et al., 24 Jul 2025). Its salient characteristics are line-level images, Tamil-character-only filtering, and clean black-on-white rendering. The paper does not report font sizes, anti-aliasing policy, file format, DPI, Unicode normalization, bold/italic variants, blur, rotation, noise, or any geometric or photometric corruption. It explicitly frames the current dataset as clean and suggests future work should include more fonts, backgrounds, and varied noise conditions, which implies that the present dataset lacks such augmentations (Jayatilleke et al., 24 Jul 2025).
The multilingual Indic benchmark describes a richer but still relatively high-level synthetic pipeline. Its required ingredients are text corpus, fonts, background image, noise filters, and text color codes (Saini et al., 2022). The workflow is: collect a text corpus for each language, filter out other languages using language detection, optionally filter language-specific characters and auxiliaries with regular expressions, break long text into lines, choose rendering attributes randomly from pools of fonts, backgrounds, colors, and noise settings, render line-level images, attach exact ground-truth text, and then analyze character coverage per font so that missing characters and numerals can be generated and added back (Saini et al., 2022). The paper states that the synthetic generation tool “renders images at the line level,” and that missing characters and numbers were identified “through analysis at font level” and then generated synthetically to achieve coverage (Saini et al., 2022).
This multilingual benchmark also formalizes three synthetic quality classes. Class 1 is “high quality,” with white background, no noise filter, no color text, and clear background; it is described as “the most basic OCR complexity, with clear text on a white background and font variation” (Saini et al., 2022). Class 2 is “average quality,” with color or white background, minimal noise filters, color text, and less noisy backgrounds, intended to simulate high-quality scanned documents (Saini et al., 2022). Class 3 is “bad quality,” with white or color backgrounds, heavy noise filters, color text, and complex backgrounds, intended to simulate poor-quality scanned documents with font variation, noise, and blurring (Saini et al., 2022). Tamil-specific counts are reported across these classes, but the paper does not define exact corruption parameters such as blur sigma, JPEG levels, or geometric-transform ranges.
At the methodological frontier, “synthocr-gen” defines a general Unicode-to-image pipeline that is directly relevant to synthetic Tamil OCR construction even though the released dataset is Kashmiri rather than Tamil (Malik et al., 22 Jan 2026). It formalizes dataset generation as from a Unicode corpus , with segmentation, validation, rendering, augmentation, and output packaging as distinct stages (Malik et al., 22 Jan 2026). It supports segmentation modes , uses grapheme clusters for character mode, applies NFC normalization, enforces script-purity validation through Unicode-range filtering, performs weighted multi-font rendering, and exports datasets in OCR-framework-specific formats (Malik et al., 22 Jan 2026). For Tamil, the paper does not provide an implementation, but it explicitly states that the method is applicable to “any Unicode-supported writing system” and only requires an appropriate digital Unicode corpus and fonts (Malik et al., 22 Jan 2026). This suggests a technically stronger future Tamil benchmark could move beyond line-only clean rendering toward grapheme-aware, normalization-aware, reproducible generation.
4. Fonts, scripts, and Unicode considerations
Font selection is a defining variable in Tamil synthetic OCR benchmarking. The Tamil-specific zero-shot benchmark uses six Google Fonts—Hind Madurai, Noto Serif Tamil, Kavivanar, Noto Sans Tamil, Pavanam, and Anek Tamil—selected from an initial pool of 17 after visual inspection to reduce redundancy (Jayatilleke et al., 24 Jul 2025). The multilingual Indic benchmark reports a substantially broader Tamil font inventory, listing 17 Unicode fonts and 3 non-Unicode fonts for Tamil, while also noting that the visible list appears longer than 20 entries because of table-formatting inconsistencies (Saini et al., 2022). The listed Tamil fonts include liberationserif, helvetica old, symbol old, tscu-saiindira-old, courier old, akshar, tmotabbi-ship-old, muktamalar-bold, kavivanar-regular, baloothambi2-semibold, arial-unicode-ms, pavanam-regular, notoseriftamil-italic-variablefont, arimamadurai-black, latha, catamaran-variablefont-wght, nirmala, lohit-tamil, meerainimai-regular, hindmadurai-semibold, latha old, tmotabb-ship-old, and karlatamilinclined regular (Saini et al., 2022). The inclusion of both Unicode and non-Unicode or legacy fonts is particularly important for Tamil OCR because legacy encodings remain relevant in practical document collections.
Legacy-font OCR is treated more directly in the Tesseract adaptation study, which identifies 10 Tamil fonts “mostly used in Sri Lankan portable documents”: Aabohi, AnbeSivam, Baamini, Eelanadu, Kamaas, Keeravani, Kilavi, Klaimakal, Tamilweb, and Nagananthini (Vasantharajan et al., 2021). The authors create TIFF/Box training pairs from Tamil training text rendered in these fonts and manually correct bounding-box and spacing errors using jTessBoxEditor (Vasantharajan et al., 2021). This work is not a formal benchmark release, but it demonstrates that font-stratified evaluation is a critical dimension of Tamil OCR benchmarking.
Unicode handling is unevenly documented across papers. The zero-shot Tamil benchmark states that the content was limited to Tamil characters only, but does not describe Unicode normalization or grapheme-cluster handling (Jayatilleke et al., 24 Jul 2025). The multilingual Indic benchmark reports filtering of language-specific characters and empirical font validation to ensure correct rendering, but does not discuss NFC/NFD/NFKC/NFKD, grapheme segmentation, or script-shaping engines (Saini et al., 2022). By contrast, “synthocr-gen” makes Unicode normalization and script-purity enforcement explicit. It defines NFC normalization as and validates segments through script-range filtering, rejecting mixed-script contamination unless explicitly allowed (Malik et al., 22 Jan 2026). The same paper treats grapheme clusters as the unit of character segmentation, which is especially relevant for Indic scripts where visual characters can correspond to multiple code points (Malik et al., 22 Jan 2026).
A related methodological lesson comes from low-resource OCR error analysis outside Tamil. The OCR4MT paper does not include Tamil, but its table of Khmer, Lao, and Pashto character-level substitution and deletion errors shows that script-specific OCR evaluation should distinguish visually similar yet Unicode-distinct characters and track deletions explicitly (Ignat et al., 2022). The cited details are substitution and deletion patterns such as Khmer ដ → ត, Lao ໃ → ໄ ; ລ → ຣ, and Pashto ګ → گ ; ي → ی, together with deletion examples (Ignat et al., 2022). This suggests, methodologically rather than directly, that a robust Tamil synthetic benchmark should be Unicode-aware and character-diagnostic.
5. Benchmark protocols and measured behavior
The most fully specified Tamil benchmark protocol is the zero-shot evaluation of five OCR engines on the 7,000-sample synthetic Tamil dataset: Google Cloud Vision API, Google Document AI, Tesseract 5.5.0, Surya, and EasyOCR (Jayatilleke et al., 24 Jul 2025). Each synthetic Tamil image is passed to the OCR engine as-is, without fine-tuning, and the recognized text is compared against the reference using five metrics: Character Error Rate (CER), Word Error Rate (WER), BLEU, ANLS, and METEOR (Jayatilleke et al., 24 Jul 2025). CER and WER are both defined as and , with the unit of comparison being characters or words respectively (Jayatilleke et al., 24 Jul 2025). ANLS is described in thresholded normalized-Levenshtein form, and BLEU and METEOR are invoked in their standard sequence-overlap senses (Jayatilleke et al., 24 Jul 2025).
The reported Tamil results show that Document AI is the best-performing system on all five metrics, with CER = 0.0078, WER = 0.1198, BLEU = 0.5803, ANLS = 0.9923, and METEOR = 0.8762 (Jayatilleke et al., 24 Jul 2025). Cloud Vision API is nearly identical, with CER = 0.0079 and WER = 0.1204 (Jayatilleke et al., 24 Jul 2025). EasyOCR is the strongest open-source system in that study, while Tesseract and Surya perform substantially worse on Tamil in the zero-shot clean-synthetic setting (Jayatilleke et al., 24 Jul 2025). The authors emphasize that Tamil OCR can exhibit a marked gap between excellent character recognition and much weaker word-level reconstruction: the best Tamil CER is only 0.78%, yet the best Tamil WER is 11.98% (Jayatilleke et al., 24 Jul 2025). They interpret this as a difficulty in word formation, spacing, and whitespace detection rather than raw character classification.
The same paper also performs character-level error analysis for the best Tamil model, Document AI, using a threshold of more than 1,600 errors (Jayatilleke et al., 24 Jul 2025). It reports that the diacritic “O” is the most error-prone character in Tamil, and that letters such as ‘க’, ‘த’, ‘ன’, and another symbol corrupted in the extract are also among the most problematic (Jayatilleke et al., 24 Jul 2025). This is one of the few directly Tamil-specific character-diagnostic findings in the cited dataset literature.
A broader OCR benchmarking perspective comes from the Indic CTC study. Tamil appears in both an internal dataset and the public Mozhi dataset, with word- and line-level cropped-image tasks evaluated using Character Accuracy (CA) and Sequence Accuracy (SA) (Mathew et al., 2022). For Tamil, internal test performance of the CRNN recognizer is word CA 95.4 and SA 84.5, versus line CA 95.9 and SA 41.2 (Mathew et al., 2022). On Mozhi, Tamil achieves word CA 98.0 and SA 91.6, versus line CA 98.3 and SA 68.1 (Mathew et al., 2022). These numbers indicate that line-level exact-match evaluation is considerably harsher than character-level scoring even when character accuracy remains high. This supports the view that synthetic Tamil benchmarks should report more than a single low-level metric.
6. Limitations, adjacent resources, and research implications
The current synthetic Tamil benchmarks are deliberately narrow. The Tamil-specific zero-shot dataset is synthetic-only, clean, and restricted to black text on a white background (Jayatilleke et al., 24 Jul 2025). It does not model poor print quality, low resolution, shading, blur, distortion, textured backgrounds, clutter, disconnected strokes, isolated dots, line breaks, motion blur, focus blur, or camera-capture artifacts (Jayatilleke et al., 24 Jul 2025). It also covers printed text only, not handwriting, natural-scene text, historical page scans with rich layout, or mobile document photography. The paper explicitly acknowledges these limitations and proposes future expansion in dataset size, number of fonts, background variation, and noise conditions (Jayatilleke et al., 24 Jul 2025).
The multilingual Indic synthetic benchmark is richer in conditions, but it remains line-level and under-specifies rendering parameters, annotation schema, public download path, and licensing for the final dataset images (Saini et al., 2022). It is highly useful for Tamil because it reports synthetic and real counts, font support, and quality classes, but it does not define a Tamil-specific split or evaluation protocol. “synthocr-gen” solves many reproducibility problems—segmentation granularity, normalization, augmentation formalization, deterministic seeded generation, and export formats—yet it stops short of providing a Tamil benchmark or OCR baseline results on Tamil (Malik et al., 22 Jan 2026).
Adjacent Tamil OCR resources expose complementary benchmarking needs. The MNIST-like Tamil vowel dataset is synthetic, public, and easy to benchmark, but it covers only 13 isolated classes—12 vowels plus aytham—in 28×28 grayscale images, which makes it a narrow character-classification benchmark rather than a full OCR benchmark (Annamalai, 2020). The uTHCD dataset, by contrast, is real rather than synthetic and targets isolated handwritten character recognition with around 91,000 samples across 156 classes (Shaffi et al., 2021). This suggests that printed synthetic benchmarking and real handwritten benchmarking currently occupy different parts of the Tamil OCR ecosystem. The binary document super-resolution study further shows that Tamil OCR evaluation can be centered on low-resolution binary scans and paired LR/HR data, with OCR accuracy improving from 25.83% character accuracy on 75 dpi input to 63.68% after the best CTS + PReLU + gamma + enhancement pipeline (Pandey et al., 2018). Although not a synthetic font-rendered benchmark in the strict sense, it demonstrates a scan-derived synthetic-degradation regime relevant to Tamil OCR benchmarking.
A plausible implication is that a mature synthetic Tamil OCR benchmarking dataset would integrate the strengths of several currently separate strands: the fairness and simplicity of the 7,000-sample zero-shot benchmark (Jayatilleke et al., 24 Jul 2025), the font diversity and quality stratification of the multilingual Indic benchmark (Saini et al., 2022), the Unicode-aware, grapheme-aware, reproducible generator design of “synthocr-gen” (Malik et al., 22 Jan 2026), the sequence-level CTC evaluation framework and real-data transfer setting of Mozhi (Mathew et al., 2022), and the font-stratified legacy-print evaluation found in Tesseract adaptation work (Vasantharajan et al., 2021). What is directly established in the cited literature, however, is more limited: synthetic Tamil OCR benchmarking currently exists chiefly as a clean printed line-image benchmark of 7,000 OPUS-derived samples (Jayatilleke et al., 24 Jul 2025), as a Tamil subset of a larger 23-language Indic synthetic OCR collection with 3,111 synthetic Tamil line images plus auxiliary real subsets (Saini et al., 2022), and as a set of reusable methodological patterns for expanding beyond those constraints.
