Searchable PDF OCR made simpler, in your browser

If you can see words on a page but cannot select or copy them, the PDF may contain a picture of the document rather than encoded text. Optical character recognition analyzes that picture and proposes characters. This tool reads printed English and Simplified Chinese and offers two outputs. Choose searchable PDF to add invisible recognized characters to image-only pages while retaining the original page appearance when optional scan adjustments are disabled. Choose TXT to review the text in an editable box and download your corrections as UTF-8 text.

TXT output is plain text, while searchable PDF adds a recognition layer and keeps the original pages by default. Neither output is a Word document or a reconstructed editable page layout. Images, fonts, tables, and precise positioning are not carried into TXT. That makes it useful for notes, searching within a text editor, or preparing a passage for further editing. When the original PDF already has a reliable text layer, use the PDF-to-text tool first; direct extraction usually avoids unnecessary recognition errors.

Add a PDF, choose the recognition language, and start processing. The tool renders the pages and recognizes them in sequence. Page markers keep the output associated with the source's physical page positions. Wait for the complete result, then compare it with the original and edit the text box as needed. The TXT download updates when you make corrections, so the saved file reflects the text currently shown.

Begin with a short representative document if you are unsure about scan quality. A clean printed page is a better test than a cover containing decorative lettering or a photograph. Check a passage with punctuation, numbers, and any language-specific characters you care about. Recognition can appear plausible while changing a single digit or omitting a word, so judging only whether the output contains readable sentences is not enough.

Choose English for English printed material, Simplified Chinese for predominantly simplified Chinese text, or the combined option for pages that mix the two. A suitable language model helps the recognizer interpret likely characters and words. The combined option can be useful for Chinese technical documents containing English product names, abbreviations, or references. Searchable output using a Chinese model embeds a complete font for continuous Chinese search, so its file size can grow noticeably. It is not a promise that every script or specialized symbol will be recognized correctly.

This page does not offer dedicated models for every language. Traditional Chinese, Japanese, handwritten notes, mathematical notation, and unusual typography may produce poor results even when some characters overlap with the selected model. Do not describe such output as a faithful transcription without checking it. If you need a language outside the available choices, use a recognition workflow designed for that language rather than expecting the nearest-looking option to be equivalent.

OCR works from pixels, and several different characters can look similar at low resolution. A zero may resemble the letter O, a one may resemble a lowercase l, and punctuation may disappear into scan noise. Compression artifacts, skew, shadows, pale printing, and damaged paper can make those distinctions harder. A recognition model can also favor a familiar word over an unusual surname or technical term.

Proofread identifiers, dates, decimal points, negative signs, and units against the scan. For a table of amounts, check the relationship between each label and value instead of reading the numbers alone. If the source is blurry, rescanning at a clearer setting can help more than repeatedly running the same image through OCR. This tool cannot reconstruct detail that is absent from the scan, and it does not know which errors would matter most in your document.

Use a scan with upright pages, even lighting, readable contrast, and minimal shadows. Keep the paper flat when possible, and avoid cutting off letters at the edges. A photograph taken at a strong angle may need perspective correction before it becomes a suitable OCR source. If the PDF contains two facing book pages in one image, separating the spreads can make manual review easier, although it does not itself correct curved text near the spine.

Do not aggressively compress the source before recognition if doing so makes small characters less distinct. Keep a higher-quality original for OCR and create a smaller sharing copy separately if needed. This workflow renders pages at a fixed working scale with a pixel safety limit; optional small-angle, brightness and contrast adjustments are available. There are no automatic controls for removing folds, restoring faded ink, or choosing different OCR regions within a page.

The intended use is printed English and Simplified Chinese. Handwriting can produce missing or incorrect characters, and the tool does not promise reliable handwriting transcription. A page with several columns may also return text in an order that needs manual repair. Captions, sidebars, footnotes, and text over illustrations can interrupt the expected flow even when individual words are recognized reasonably well.

Tables are especially important to check. Plain text has no spreadsheet cell model, so the visual relationship between rows and columns may be lost. This tool does not export structured CSV or Excel data. If you need a reliable table, use a workflow designed for table recognition and verify the resulting cells. For a small passage, manually editing the recognized text may be quicker than trying to preserve the entire original layout.

Select Searchable PDF in the output-format control, choose the appropriate language, and process the document. The tool checks each page for existing text. Pages with text stay unchanged, avoiding a second overlapping recognition layer. Image-only pages are recognized, and an invisible text-only OCR page is mapped back onto the source page. With scan adjustments disabled, the original images and page order remain in the exported copy. Page rotation and visible crop boundaries are accounted for when placing the layer.

Download the result and search for a distinctive word in your reader. Select a short sentence and check that the selected area aligns with the visible sentence, particularly on rotated pages. Confirm that copied words match the image. A page with no recognized characters receives no new layer; the summary reports how many pages received OCR and how many existing text pages were kept. A document whose text is already selectable but incorrectly encoded is not automatically repaired by this mode.

No. TXT and searchable PDF are separate processing modes. Editing the TXT result updates that text download, but it does not rewrite the positions or characters in a searchable PDF. The PDF text layer uses the recognition engine's result. If its numbers or names are wrong, improve the source and recognize it again, or use a PDF OCR editor with text-layer correction. Do not treat a corrected transcription as evidence that a separately generated PDF was corrected too.

The page image remains the reference when recognition and appearance disagree. Search can miss a word that OCR misread even though the scan looks correct. Likewise, copying a plausible but incorrect amount can cause problems downstream. Keep the original, check important passages, and describe an unreviewed result accordingly. The tool does not convert paragraphs into editable Word objects, restore table cells, or supply a complete accessibility structure.

With scan adjustments disabled, the workflow adds text to the existing PDF rather than rebuilding visible pages. Existing page objects, images, geometry, and ordinary navigation objects remain in the document. Check the downloaded copy's links and bookmark destinations if you rely on them. Signature fields, certified documents, and XFA forms are not supported for searchable output. The tool does not promise to preserve the validity of a digital signature or the behavior of specialized document formats.

An invisible text layer does not itself create a correctly tagged accessible PDF. Reading order, headings, image descriptions, and table relationships need a separate accessibility review. OCR also does not permanently remove sensitive content from an image. If you later compress the result using an image-only compressor, the searchable layer can be lost. Choose subsequent transformations carefully and repeat your search and selection checks after the last processing step.

The browser needs the recognition engine and the selected language resources before it can recognize text. These resources are loaded from this site's host rather than requiring your document to be uploaded to an OCR service. A slower connection can make initialization noticeable. Recognition itself then uses your device's processor and memory, so a long or complex scan can take more time on a phone than on a desktop computer.

Keep the page open while processing. If you cancel, the running recognition worker is stopped and no partial result is offered for download. You can then try a shorter file or a more appropriate language setting. A model-loading or rendering failure produces an error rather than a success message with an empty file. In TXT mode, a page with no recognized text receives an explicit marker. In searchable mode, that page remains visually intact without a new text layer.

Use one unencrypted PDF up to 30 MB and thirty pages. Output cannot exceed 120 MB. Rendering also has a per-page pixel limit to avoid unexpectedly large canvas allocations. A very large physical page may therefore be rejected even when the file itself is small. Damaged files and unsupported page geometry are not processed as successful results. The site code and recognition resources load over the network, but selected document bytes remain in the browser's processing workflow.

Before downloading, check the beginning and end of each recognized page, unusual names, numbers, and any passage you intend to quote. Correct line order and punctuation where necessary. Save the selected output with a name that connects it to the source scan, and retain that scan for future verification. In PDF mode, also test search and text selection in the downloaded copy. OCR is a useful first draft of a transcription; your review is what makes the result appropriate for its next use.

Enable Adjust scanned pages before OCR, enter the physical pages to adjust, and select a preview page. You can set a correction angle from minus five to five degrees; positive values turn the displayed page clockwise. Preview before confirming. This small-angle adjustment is different from a quarter-turn rotation and does not correct perspective, curved book pages, folded paper, or uneven shadows.

The optional automatic estimate looks for horizontal ink alignment. Sparse text, decorative lines, tables, or noisy backgrounds can mislead it, so the estimate is a suggestion to inspect. Use a manual angle if the suggested result is unsuitable. Correction fits the page within its original visible bounds to avoid cutting off corners; it can shrink the content slightly and introduce white margins. Inspect the edges as well as the main paragraph.

Brightness and contrast each accept values from minus thirty to thirty, and grayscale is optional. These adjustments can make a gray scan easier to read, but an aggressive setting can make faint handwriting or a light stamp disappear. Nothing here automatically identifies which marks are important. Compare both preview images and keep adjustments modest when a page contains subtle detail or meaningful color.

Reset adjustments returns the angle, brightness and contrast to zero and turns off automatic estimation and grayscale. Review the new preview before confirming again. Selected image-only pages are rebuilt at the working resolution of 144 DPI in searchable PDF output, so their original image bytes are not retained. Links and annotations on those rebuilt pages are removed. Other pages stay unchanged, and pages with existing text are never scan-adjusted. Interactive forms are unsupported with adjustments enabled.

Yes. Enter a range such as 1,3-5 rather than assuming every page needs the same treatment. Check a representative page from each section, because lighting and skew can vary across a scan. The result summary reports how many scan pages were rebuilt. In TXT mode, the adjustment changes the image used for recognition, but the download remains plain text; it is not an enhanced scan file.

After producing a searchable PDF, inspect the adjusted pages, search for a distinctive word, and compare text selection with the visible characters. Unadjusted image pages keep their original appearance, while existing text pages stay unchanged. The adjustment process is not permanent redaction: the surrounding source document can still contain hidden objects or metadata. Use the separate redaction tool if your goal is to remove private content before sharing.

Select two through eight files in the same picker to start a sequential queue. The selected language and output format apply to every file. Each file may contain up to thirty pages and thirty megabytes; total selected input and successful output each have a one hundred twenty megabyte limit. A damaged or encrypted file is listed as failed, and the queue continues with the remaining documents.

Download the successful results individually or together as a ZIP. The archive includes manifest.json with source names, outcome details, and output names. Duplicate source names receive distinct output names so that one result cannot overwrite another. Batch TXT files should be proofread separately in your text editor; there is no shared text box that merges unrelated documents. A searchable PDF retains its own page count and existing text pages. Scan adjustments are available in single-file mode, where you can preview and confirm the changed appearance. Canceling clears the unfinished queue results, and selecting another set or changing settings invalidates the previous download.

Choose a tool for the next step and review each output.

• Extract an existing text layer
• Separate book spreads