Translate a Scanned PDF in Phrase TMS

A client sends a scanned PDF of a notarised contract, or a WhatsApp photo of a birth certificate, and the job has to be in Phrase TMS by tonight. You create the job and the analysis has next to nothing to count, because every page is a picture. If you’re searching for how to translate a scanned PDF in Phrase TMS, the answer starts with a setting you pick before the job exists: which of Phrase’s three PDF filters it uses.

Below: what Phrase TMS (formerly Memsource) does with scans as of October 2026, how to set up its PDF v2 filter with OCR, when to convert the scan to Word first, and what XTM Cloud does with the same file. Platform choice is covered in the Phrase TMS vs Smartcat comparison, and the same problem in other tools in the guides for Trados Studio and memoQ.

What Phrase TMS actually does with a scanned PDF or image¶

Phrase TMS handles PDF files through three import filters, and the filter a job uses decides whether a scan gets read at all. Phrase’s PDF help page, last edited in June 2026, lists them as Default, PDF v2 and TransPDF.

PDF filter in Phrase TMS	What Phrase’s help says	Translated output	Reads a scan?
Default	For analysis only, limited to 100 MB; the jobs table shows a warning	Unformatted text file	No OCR described
PDF v2	“Native .PDF support within Phrase TMS”, with an OCR Mode and a WYSIWYG Mode	DOCX	Yes, OCR Mode Auto (default), Force or Never
TransPDF	A third-party service (Iceni) that needs its own account	PDF, and downloading the finished PDF is paid	OCR runs on Iceni’s side

The Default filter is the trap for anyone who wants a usable target file. As Phrase’s PDF documentation puts it:

If an imported .PDF file is translated, output will only be in the form of an unformatted text file.

A job imported under Default can be counted for a quote, but the target loses the layout entirely: no tables, no headings, just running text.

PDF v2 is the filter with OCR, and it’s new. It appeared in Phrase’s help center during 2026; the October 2025 version of the same page listed only Default and TransPDF. Any forum answer written before 2026 that names TransPDF as Phrase’s only scan route is out of date on that point.

Photos and image files¶

Phrase TMS can’t take a photo as a translatable job. Phrase’s supported formats list has no JPG, PNG or TIFF; the only graphics formats it names are .psd and .svg. A JPG of a passport page goes in only as a reference file, and Phrase’s release notes for September 7, 2026 say image reference files now open inline in the editor, next to the segments. Agencies that get most documents as phone photos need a conversion step at intake; the WhatsApp document photo workflow shows one.

The OCR switch inside Word and PowerPoint imports¶

Phrase’s import settings for Word and PowerPoint files include “Extract text from embedded images (OCR)”, and Phrase’s Word documentation describes what it does:

When enabled, text from embedded images is extracted via OCR and inserted as a comment next to each image.

A comment is not layout. When a client’s “Word file” is really page photos pasted into Word, this option gives the translator the recognised text beside each picture, for reference, while the picture stays untranslated in the target. Phrase doesn’t document whether those comments become translatable segments, so don’t build a delivery on it.

Check what you actually have first. Open the PDF and try to select a sentence; if nothing highlights, the page is an image and needs OCR. The article on why Trados and memoQ fail to import scanned PDFs walks through that test and through mixed PDFs.

Why a scan breaks in a CAT tool, Phrase included¶

A scanned PDF is a picture of a page wrapped in a PDF container. A CAT tool or TMS works on text: it splits sentences into segments, matches them against translation memory and writes the translation back into the same structure. With no text in the file, there’s nothing to segment.

OCR (optical character recognition) reads letters off the picture, which is half the job. The other half is rebuilding the layout: which lines form a paragraph, whether a grid of lines is a table, which column is read first. Get the reading order wrong on a two-column certificate and the segments interleave the columns, as explained in why CAT tools break PDF tables and multi-column layouts.

Official documents carry things that aren’t text at all. A round seal, a signature, a handwritten date, a photo and a QR code are images; OCR either skips them or reads a stray word off the seal ring. Handwriting reads unreliably in any OCR, which is why handwritten documents need a human pass against the original.

Tag noise is the third problem. When OCR output switches font, size or character spacing every few words, a CAT tool marks each change with an inline tag, and segments fill up with tags the translator has to carry over. Phrase’s own Word import settings include a “Minimize number of tags” option, which its documentation describes as meant for scanned documents.

Rebuilding a usable Word copy from a scan is a separate job from OCR. In the scan pipeline behind ChatsControl (Route 3 below), a vision model looks at each page and writes the Word copy. Text is typed as printed, tables become tables, fields sit beside their labels, and stamps, seals, signatures and photos are cut from the page and placed where they stand. QR codes are read and drawn again clean. A second check compares the rendered copy with the page line by line and sends what it finds back to be fixed, and each scan gets a quality grade of good, fair or poor. Handwriting and very degraded scans still read unreliably, and the grade says so.

Route 1: PDF v2 with OCR inside Phrase TMS, step by step¶

Route 1 keeps everything inside Phrase: you pick PDF v2 when you create the job, Phrase runs OCR, you translate in the editor and download a DOCX. Phrase’s Create a Job article places the import settings under File import settings when you add a job through project > Jobs > New.

Test the PDF. Try to select a sentence on body pages, tables and stamped pages. No selectable text means the file needs OCR; garbled selectable text means the scanner added a bad text layer.
Start a new job. In the project, go to Jobs > New and upload the scan.
Open File import settings. In the PDF settings, choose PDF v2 out of the three filters.
Set OCR Mode. Phrase’s help gives three values: Auto (the default) “performs OCR only when needed”, Force “applies OCR to all pages” and Never “disables OCR entirely”. For a fully scanned document, or one with a garbled scanner text layer, Force makes Phrase read every page itself.
Leave WYSIWYG Mode on. Phrase says it keeps the visual layout and is on by default.
Create the job and compare. Open the first segments next to the scan and check names, dates, document numbers and the order of table cells.
Attach page photos as reference files, if the client sent any.
Translate, run QA and download the target, which is a DOCX. Plan the step to PDF separately (see delivery below).

What Phrase doesn’t document about PDF v2¶

As of October 2026, the PDF help page doesn’t say which OCR engine PDF v2 uses, which languages its OCR reads or how it rebuilds tables. It also doesn’t say whether the 100 MB limit stated for Default applies to PDF v2, or whether the filter depends on your plan. If PDF v2 isn’t offered in your account, ask Phrase support.

Treat your first PDF v2 scan as a test, not a delivery. Run one hard page (a table, a stamp, two columns, Cyrillic next to Latin) and compare the segments with the image. If a Soviet-era certificate comes back with merged columns or the stamp’s ring text scattered into segments, move that document to Route 2 or 3.

TransPDF: the older, paid route¶

TransPDF is a service from Iceni, a separate company. According to Phrase’s PDF help page, you need your own TransPDF account, you connect it under Settings > Integrations > TransPDF with its username and password, and downloading the finished PDF is paid. The free tier described there covers 10 PDFs or 25 MB, with watermarked previews.

Iceni’s published TransPDF prices, as of October 2026: OCR of a scan costs 1 credit per page, the final PDF another 1 credit per page, and a credit costs 50 cents, falling to 15 cents in bulk (the page doesn’t name the currency). New accounts get 25 free credits. A 10-page scan taken through OCR and back to PDF uses 20 credits, which is 10.00 at the 50-cent rate.

TransPDF gives you a PDF back instead of a DOCX, at the price of a second account, per-page credits and the client’s file going to a third party.

XTM Cloud: no OCR, so the scan is read somewhere else first¶

XTM Cloud (now branded xtm.ai, help docs at version 26.4) has no OCR step. XTM’s knowledge base article on PDF source files, updated in February 2026, says a scanned PDF fails project analysis with “file could not be analyzed”, and explains why:

your PDF file might actually be are scans or images – XTM Cloud can not extract the text out of them. Unfortunately, we currently do not support OCR files.

XTM has no Route 1: a scan has to become text before it goes into an XTM project, and the analysis error tells you it didn’t. The same article sets the delivery format for text PDFs:

Source files in PDF format are always returned as target files in DOCX format.

XTM’s supported formats list describes PDF support as “PDF (converts PDF file to Word)” and has no JPG, PNG or TIFF among translatable formats; its graphics formats are PSD, AI, SVG, FXG and EPS. A phone photo needs converting before it can become an XTM job, same as in Phrase.

The XTM workflow for a scan: convert it to DOCX outside XTM (Route 2 or 3 below), check it against the page and upload it as the source file. A checked DOCX also gives XTM a real word count. To quote before the OCR is done, see how to quote a scanned document when word count finds nothing.

Route 2: OCR to DOCX first, then import as a normal Word file¶

Route 2 works the same in Phrase TMS and XTM Cloud: you make the Word file yourself, so the TMS only sees an ordinary DOCX. The OCR tool is your choice (Word’s PDF conversion, Adobe Acrobat, ABBYY FineReader or Google Docs), and the comparison of ways to turn a scanned PDF into editable Word covers each one’s limits.

Run OCR with the real source language. A Ukrainian certificate with a Russian-language Soviet stamp needs both languages selected, or recognition runs against the wrong dictionary.
Export to DOCX and keep the original scan untouched.
Join broken lines. OCR tends to put a hard return after every printed line, and a paragraph break ends a segment, so one sentence becomes three fragments in the editor.
Take text out of floating text boxes and put it in body paragraphs or table cells, or the reading order in the TMS won’t match the page.
Rebuild tables as real Word tables where OCR produced tab-separated lines.
Strip formatting noise. Remove font, size and character-spacing changes the page doesn’t need, because each change becomes a tag.
Leave stamps, seals and signatures as pictures where they stand. The translator adds a note such as [Seal: …] in the target if the authority wants the wording; see translating stamps and seals.
Compare the DOCX with the scan line by line: names, dates, amounts, document numbers.
Import into Phrase as a normal Word job. Consider “Minimize number of tags” in the Word import settings; leave “Extract text from embedded images (OCR)” off.
Attach the original pages as reference files and check the tag view after import.

Pre-import checklist for a scanned document¶

Check	Why it matters in Phrase or XTM	Done when
Text layer present	A picture gives no segments and no word count	You can select text on every page
OCR language set	The wrong dictionary garbles names and numbers	Every language on the page was selected
No line breaks inside sentences	Each paragraph break ends a segment	Sentences run to their full stop
No floating text boxes	Text boxes scramble the reading order	Text sits in paragraphs or table cells
Tables are real tables	Tab-separated lines merge columns into one segment	Rows and cells match the page
Formatting noise removed	Font and spacing changes turn into tags	One font per paragraph unless the page differs
Stamps, seals, signatures, QR codes	They are images; OCR misreads them	Kept as pictures where they stand
Numbers, dates, names checked	OCR errors become translation errors	Compared with the scan character by character
Original attached	The translator needs the page to resolve doubts	Scan or photos added as reference

If the scan itself is skewed, dark or blurred, every row gets harder; what to fix in a poor-quality scan before translation covers that first step.

Route 3: a Word copy from a scan-to-Word service, then Phrase TMS¶

Route 3 replaces the OCR and cleanup of Route 2 with a service that rebuilds the page as a Word file. One option is ChatsControl: you upload the scan or photo (PDF, PNG, JPG, TIFF, WebP), it reads the page and returns an editable Word copy that follows the page layout. Text is typed as printed, tables come back as real tables, stamps, seals, signatures and photos are cut from the page and placed where they stand, and QR codes are redrawn clean. You then import that .docx into Phrase TMS or XTM Cloud like any client Word file, or have it translated directly in the same service.

Upload the scan or photo and wait for the job to finish.
On the job page, download the Word copy of the original with the “Original” button, next to the translated Word file and the raw scan.
Check the Word copy against the scan: names, numbers, table cells and the quality grade.
Create the Phrase job (or XTM project) from that .docx and attach the original pages as reference files.
Check the tag view after import, as with any converted file.

The result is an ordinary Word document; Phrase TMS or XTM Cloud imports it the way it imports any client .docx. The scan examples page shows real scans next to the Word files the service returned, so you can judge tables and stamped pages before uploading a client file.

Prices as of October 2026, from ChatsControl: a free tier of 6,000 characters a month, then pay as you go at $1.50 per 1,000 characters ($2.70 per 1,800-character page). Scans are billed by recognised characters, at least 1,000 characters per page, and uploads are deleted after 30 days by default. The job always produces a translation as well, and that’s what the upload is billed for; there’s no separate OCR-only mode.

ChatsControl isn’t a CAT tool or TMS: no translation memory, no segment-level project management. It runs in the cloud only, with no on-premise or offline mode. Handwriting and very degraded scans read unreliably, and the quality grade says so. Stamps, seals and signatures stay as pictures, so if the receiving authority wants a seal’s wording rendered, the translator adds a bracketed note such as [Seal: …].

Three routes compared for Phrase TMS and XTM Cloud¶

	Route 1: PDF v2 in Phrase	Route 1b: TransPDF in Phrase	Route 2: your OCR to DOCX	Route 3: scan-to-Word service
Works in XTM Cloud	No	No	Yes	Yes
Where OCR happens	Inside Phrase	Iceni’s service	Your OCR tool	The service
Cost (October 2026)	No separate OCR price in Phrase’s PDF docs	1 credit per page OCR plus 1 per page for the final PDF, 0.50 to 0.15 per credit (Iceni)	Your OCR licence plus cleanup time	$1.50 per 1,000 characters after 6,000 free a month, at least 1,000 characters per scanned page
Accepts photos (JPG, PNG)	No	No	If your OCR tool does	Yes (PNG, JPG, TIFF, WebP)
Target file	DOCX	PDF (paid download)	DOCX	DOCX
Suits	Clean, simple scans already in Phrase	Teams that must return a PDF	Teams with an OCR licence and time to clean up	Stamped forms, tables, phone photos, XTM projects

What each route keeps from the page, by what each tool documents:

Element on the page	PDF v2 in Phrase	Your OCR to DOCX	Scan-to-Word service (ChatsControl)
Printed text	Read by OCR (engine and languages not documented)	Read by your OCR tool	Typed as printed, then checked line by line
Tables	Not documented for TMS	Depends on the tool; you rebuild by hand	Rebuilt as real Word tables
Stamps, seals, signatures	Not documented	Images if the tool keeps them	Cut from the page and placed where they stand
QR codes	Not documented	An image, if kept	Read and redrawn clean
Handwriting	Not documented	Unreliable	Unreliable; the quality grade flags it
Layout	WYSIWYG Mode keeps the visual layout	Depends on cleanup	Follows the page layout

PDF v2 is the cheapest to try, since Phrase’s PDF documentation lists no separate charge for it. For an agency handling stamped certificates every week, cleanup time is where the money goes, as the real cost of retyping scans breaks down.

What goes wrong with scans in Phrase TMS and XTM Cloud¶

Automation creates jobs with a different filter¶

In Phrase’s API reference for job import settings, the PDF filter defaults to TRANS_PDF, and the PDF settings carry separate ocrMode and wysiwyg fields. If jobs come in through a connector, a script or an automation, check which filter they send; a scan imported that way doesn’t get PDF v2 just because a PM would have picked it in the UI.

Existing projects keep their defaults¶

Phrase’s Project Settings help says the default import settings of an existing project can’t be edited in the UI. Set the filter on each job when you create it, or change it through the API.

A job already imported with the Default filter¶

Phrase’s PDF page says the jobs table shows a warning for PDFs imported with Default, and the translated output is an unformatted text file. Don’t translate that job and hope: create it again with PDF v2 and an OCR Mode set.

Tables in scans: the caveat Phrase documents for Portal¶

Phrase has published table caveats for scanned PDFs, but for Phrase Portal, a separate product with its own built-in OCR that returns DOCX, not for PDF v2 in TMS. Phrase’s Portal page says:

Tables and charts may lose alignment, borders, and structure, especially in complex layouts. In scanned PDFs, content may occasionally be missed or incorrectly formatted.

PDF v2 has no equivalent statement either way. Take Portal’s warning as a reason to test a table page through PDF v2 before you trust it with a certificate annex, not as a description of PDF v2.

Delivery: a DOCX when the client expects a PDF¶

A scan through PDF v2 comes back as DOCX, and XTM Cloud returns DOCX for every PDF source. If the client wants a PDF that looks like the original, someone has to check the translated DOCX layout and export it, which is DTP time to price in; see DTP charges for scanned document translation. TransPDF returns a PDF at Iceni’s per-page price. Before sending, the free QA validator checks the translated file against the source for numbers and meaning.

FAQ¶

How do I translate a scanned PDF in Phrase TMS?¶

When you create the job (project > Jobs > New), open File import settings, choose the PDF v2 filter and set OCR Mode to Auto or Force. Phrase reads the scan and the translated target downloads as a DOCX file. For hard scans, convert the PDF to a checked Word file first and import that instead.

Does Phrase TMS (Memsource) have OCR?¶

Yes, as of October 2026 the PDF v2 import filter has an OCR Mode with Auto, Force and Never settings. The other two PDF filters are Default, which doesn’t read scans and returns unformatted text, and TransPDF, a paid third-party service from Iceni.

Can I import a JPG or PNG into Phrase TMS?¶

No. Phrase’s supported formats list has no JPG, PNG or TIFF; the only graphics formats are .psd and .svg. Images can be attached to a job as reference files, so convert a photo to a Word file before you create the job.

Does XTM Cloud support OCR for scanned PDFs?¶

No. XTM’s knowledge base says a scanned PDF fails project analysis and that XTM currently does not support OCR files. Run OCR outside XTM, check the result and upload the DOCX.

Will the segments and tags be messy after OCR?¶

Yes, they can be. OCR output that keeps switching fonts, sizes or character spacing creates extra tags, and line breaks inside sentences split segments. Clean the DOCX before import, try the ‘Minimize number of tags’ option for Word files and check the tag view after import, as with any converted file.

Can I deliver a PDF back to the client?¶

Not straight from PDF v2 or XTM, since both return a DOCX for a PDF source. TransPDF in Phrase returns a PDF, but Iceni charges per page for it; otherwise check the translated DOCX and export it to PDF yourself.

Is it safe to upload client scans to an online OCR service?¶

Check where the file goes and how long it stays there. PDF v2 runs inside Phrase, TransPDF sends the file to Iceni under your own account, and ChatsControl deletes uploads after 30 days by default as of October 2026. If the client contract forbids cloud processing, run OCR on your own machine.

Sources¶

Try ChatsControl

AI platform for professional translators

Try for free →