Frequently asked
questions.

Yes. The Windows desktop app is a free download with quick text translation and F6 screen translation. Get it from the download page.

Full Premium: complete document translation with layout preservation across PDF, Word, PowerPoint, and Excel, plus the Office add-ins, glossaries, and translation memory. No commitment when the month ends.

Per organization, based on how many people translate, the deployment model (cloud, sovereign cloud, or on-premises), and whether you need white-labelling. Write to support@lisan.com and we will scope it together.

Yes, on-premises and air-gapped deployments are an Enterprise option, including a local translation model. See security and on-prem.

Enterprise agreements are typically invoiced in USD, AED, or SAR. Talk to us and we will work with your procurement process.

It was engineered Arabic-first by an Arabic AI company. That means native RTL handling in the engine and in rebuilt documents, correct Arabic punctuation ؟ ، ؛, typography nuance, mirrored layouts when direction flips, and QA checks specific to Arabic orthography, all running on Lisan's own models.

Hamza seat verification, taa marbuta checks (ه versus ة), stray kashida and tatweel detection, and diacritics consistency, on top of the standard checks for untranslated segments, number mismatches, double spaces, punctuation, and length ratio.

The layout is mirrored. PowerPoint shape positions flip when the writing direction changes, and tables, margins, and alignment follow the new direction in Word and PDF too, so the translated copy reads naturally instead of backwards. See layout preservation.

Arabic, Urdu, and Persian, out of more than 20 languages in the product, with automatic source-language detection. Arabic and English are the flagship pair. See languages for the full list.

The top-quality tier of Lisan's own translation models. TranslateX lets you pick the model per job, from L1 for the fastest turnaround up to Lisan L3 for publication-grade output, plus a local model option for on-premises deployments.

Windows 10 build 19041 or later on x64. The app is published by LISAN AI - FZE and is a free download from the official installer page.

Right-click any file in Windows Explorer, choose Translate To..., and pick a language. A fully translated copy is saved next to the original with the layout preserved. The original file is never changed. See document translation.

It runs in the background so hotkeys and right-click translation are always available, and you can enable autostart so it launches with Windows. Both behaviors are optional and under your control.

F6 starts a screen translation capture and middle-click translates a single word in place. Hotkeys are global, so they work in any application, and configurable in settings. See screen translation.

PDF, Word (DOCX), PowerPoint (PPTX), Excel (XLSX) and CSV, plain text (TXT), and images such as screenshots and scans. Each comes back in its original format as a translated copy.

Up to 10 files per batch and up to 100 MB per file. Each file in a batch can have its own language pair, and you can track progress per file while the batch runs.

No. The original is never modified. TranslateX creates a new translated copy next to it, in the same format and with the layout preserved. See layout preservation.

Yes, as an option. When enabled, images inside the document are read with OCR, the original text is masked, and the translated text is drawn in its place, so charts, stamps, and infographics come out in the target language too.

No. Source-language detection is automatic, and you can still set the pair per file when you want control. See the full language list.

A glossary holds source and target term pairs per language pair, each with a category and a note. Terms are enforced automatically on every translation job in the workspace, so the ministry name, the product name, and the legal phrase come out the same in every document.

An exact match means the new sentence is identical to one your reviewers already approved: the stored translation is reused and the segment is locked. A fuzzy match is close but not identical: the stored translation is suggested with a similarity score, and the reviewer decides.

Yes. TranslateX imports and exports TMX, the standard translation-memory format, so years of past work from agencies or other tools start paying off immediately. XLIFF support keeps segment-level workflows compatible with established CAT tools.

Through review. When a reviewer approves a segment in the editor, it enters the workspace memory. Only approved segments are stored, which means the memory gets better with use instead of accumulating unreviewed machine output.

Workspace roles control that: admins and managers manage language assets, while translators and reviewers work with them in the editor. See security and on-prem for the full role model.

22 languages in the product list: Arabic, English, French, German, Spanish, Portuguese, Russian, Chinese, Japanese, Korean, Hindi, Urdu, Persian, Turkish, Indonesian, Bengali, Tamil, Malayalam, Nepali, Sinhala, Tagalog, and Amharic.

No. Source-language detection is automatic, so mixed folders can be batched with just a target language. You can still set the pair explicitly per file when you want to.

TranslateX was built Arabic-first, with native RTL and LTR handling, Arabic punctuation, and typography treated as core engineering rather than an afterthought. See Arabic & RTL excellence.

Yes. Each file in a batch carries its own language pair, so one run can translate a report to English, a press release to French, and a contract to Arabic. See document translation.

Fonts and text styles, tables, spacing and margins, headers and footers, footnotes, slide structure, and SmartArt. The translated copy comes back in the same format as the original, designed the same way.

The layout is mirrored along with the writing direction: alignment, reading order, and, in PowerPoint, shape positions on each slide all flip so the result reads naturally in the target language. See Arabic & RTL excellence.

Yes. Slide structure is preserved and shape positions are mirrored when the direction flips between RTL and LTR, so an Arabic deck becomes a natural English deck rather than a mirrored mess, and the reverse.

With the image-translation option on, TranslateX reads image text with OCR, masks the original, and draws the translation in its place, so charts and infographics match the rest of the document. See document translation.

That is the design goal: publish-ready output in the original design. Review the translation itself, run your terminology checks, and send it, rather than budgeting days for reformatting.

Microsoft Word and Microsoft PowerPoint. In Word you translate the document you are editing in place; in PowerPoint you can translate a slide or the entire deck with Translate all slides.

No. The point of the add-ins is removing the copy-paste-reformat loop entirely. Text is translated in place and the document structure stays intact. See layout preservation.

Yes. The same glossary enforcement, translation memory, and organizational style apply through the add-ins as through the web platform and desktop app. See style governance.

Yes. Document translation is also available inside the WriteX writing platform, and browser extensions plus an API and SDK exist at the Lisan platform level. The Windows desktop app covers files and screens.

Yes, and it applies TranslateX's native RTL and LTR handling while doing it, so an Arabic deck comes out as a naturally reading English deck rather than a direction-confused one.

Projects, segments, words translated, and active users, plus the language-pair mix across the workspace. Views are available per workspace and per project.

Yes. The top-users view ranks workspace members by translation volume and shows which language pairs each person works in, alongside how many members are actually active.

QA and activity reports export as CSV or text, so you can share them, file them, or load them into the tools your organization already uses.

No. They are simplified recreations with illustrative numbers, drawn to show the kind of data the reports section records: volume, pairs, and users. The real reports run on your workspace's actual activity.

Yes. Reports run wherever TranslateX runs: cloud, sovereign cloud, or fully on your own infrastructure. Audit logs and SIEM export cover the security side. See security and on-prem.

Press F6 or click the TranslateX toolbar icon, then draw a box around the region you want. The text is read with OCR and translated instantly. Hotkeys are configurable in the desktop app.

Copy it to the clipboard, open it in an editor window to adjust, or export it as a polished, formatted document. Each capture can go to a different destination.

Yes. Text extraction mode runs the OCR step only, which is handy for pulling text out of scans, screenshots, and images when you need the original language, not a translation.

Yes. Screen translation reads pixels, not the clipboard, so it works on scanned PDFs, remote desktops, dashboards, and applications that disable text selection or copying.

Middle-click a single word on screen and TranslateX translates it in place, without drawing a capture box. It is the quickest way to work through an interface or document in another language. Part of the Windows desktop app.

Yes. The on-premises deployment includes a local translation model and supports documented air-gapped installs, so translation runs entirely inside your network with no external calls. Deployment is via Docker.

LDAP and Active Directory integration, so access follows your existing directory, plus multi-factor authentication using TOTP authenticator apps.

Five: admin, manager, translator, reviewer, and viewer. Admins control deployment and security, managers own projects and language assets, translators and reviewers handle production and quality, and viewers get read-only access.

Audit logs record sign-ins, translation jobs, glossary changes, and admin actions, and SIEM export forwards those events into the monitoring stack your security team already uses. Data is encrypted in transit and at rest.

Yes. White-label branding lets an institution present TranslateX under its own name and identity, which is common for on-premises deployments offered as an internal service.

More than 24 government entities work with Lisan, and TranslateX's on-premises, air-gapped, and directory-integration capabilities were built for exactly those environments. See TranslateX for government.

It applies your organization's writing style when translating files: phrasing, register, and conventions follow how your organization writes, rather than a generic machine voice. It is a workspace setting, so it applies to every file translation once enabled.

A glossary pins specific terms to one approved translation. The style guide governs everything the glossary does not: how sentences are phrased and how formal the writing is. They work together, and translation memory adds sentence-level reuse on top. See glossary and memory.

Settings that shape how the translation reads, from formal and institutional to plainer and warmer. Terminology stays governed by the glossary either way; tone controls change the register of the writing around it.

From Lisan's writing-assistant heritage. Lisan builds AI that understands institutional writing in Arabic and English, and TranslateX applies that same understanding to translation, which is why it can produce output in your voice instead of a generic one.

Yes, and that is the hard part done properly. Arabic institutional register is its own discipline, and TranslateX was engineered Arabic-first. See Arabic and RTL excellence for what that means in practice.

Yes, and this is rarer than it should be. XLSX and CSV files translate as documents: headers, labels, and notes move to the target language while the sheet structure and figures stay untouched, and automated number checks verify nothing drifted.

Up to 10 files per batch, up to 100 MB each, mixing PDF, Word, PowerPoint, Excel, and images, with a language pair per file. Folders and projects keep department batches organized, with progress tracking and webhooks on completion.

Very little. They translate where they already work: directly inside Word and PowerPoint with the add-ins, by right-clicking a file in Windows Explorer, or with F6 for anything on screen. The organization's glossary and style apply automatically either way. See Office add-ins.

One workspace glossary enforces group terminology on every job, translation memory reuses approved wording, and the review editor with QA checks handles documents that need a human pass. Roles decide who translates, who reviews, and who administers. See glossary and memory.

A workspace summary of documents, words, and active users, translation activity over time, top users, and QA reports that export as CSV. See reports and analytics.

Yes. TranslateX deploys via Docker on your own infrastructure, with documented air-gapped installs and an On Premises (local) model option that runs the translation engine itself on your hardware. Sovereign-cloud deployment sits between full on-prem and our cloud. See security and on-prem.

Glossaries pin the approved rendering of entity names, honorifics, and legal instruments per language pair, and TranslateX enforces them on every job. Style governance applies your organization's writing style on top, and translation memory reuses wording your reviewers already approved. See glossary and memory.

Not unless you choose to. The review editor gives your own reviewers side-by-side segments, statuses, a track-changes diff, and automated QA checks on numbers and punctuation. The final document is rebuilt in the original layout only after review.

Modern Standard Arabic, the register official documents are written in, with native RTL and LTR handling, Arabic punctuation, and typography nuance. Arabic and English is the flagship pair among more than 20 supported languages. See Arabic and RTL.

Yes. More than 24 government entities work with Lisan, the company behind TranslateX. Deployments are scoped directly with each entity: write to support@lisan.com or use the contact page.

Trust the workflow, not the machine alone. TranslateX produces the draft, enforces your approved legal terminology, and runs automated QA on numbers, dates, and completeness; your counsel then reviews segment by segment with track changes before the final document is rebuilt. The machine accelerates, the human decides.

No, and no machine does. Where the law requires a certified or sworn translation, a qualified human translator must produce and sign it. TranslateX accelerates the draft and the review so that translator starts from a consistent, QA-checked document instead of a blank page.

The glossary holds one approved rendering per term, per language pair, with notes recording why counsel chose it, and it is enforced automatically on every job. Translation memory reuses boilerplate clauses your reviewers already approved. See glossary and memory.

Number mismatches between source and target, untranslated segments, double spaces, punctuation including Arabic marks, and suspicious length ratios, plus Arabic-specific checks such as hamza seats and taa marbuta. Findings attach to segments in the editor and QA reports export as CSV.

That is your choice. TranslateX runs in the cloud with enterprise-grade security, or fully on-premises with a local translation model and documented air-gapped installs, so privileged documents never leave your network. See security and on-prem.

A typical release translates in minutes as a complete document, formatting included, so the practical timeline is your review: the comms team reads the translation in the editor, adjusts the sensitive quote, and both language versions go out inside the same hour.

Yes. TranslateX rebuilds PDFs, Word documents, and PowerPoint decks in their original design, keeping fonts, tables, images, and slide structure, and mirrors slide layouts when the direction flips between Arabic and English. See layout preservation.

Press F6 and capture the article, post, or broadcast caption on screen. TranslateX reads the region, detects the source language, and translates it instantly, with output to clipboard, an editor window, or a document. See screen translation.

Pin each name, title, and brand in the glossary, one approved rendering per language pair, and TranslateX enforces it on every job. Translation memory keeps recurring boilerplate identical across releases. See glossary and memory.

More than 20 languages, with Arabic and English as the flagship pair and automatic source-language detection for monitoring. RTL languages including Arabic, Urdu, and Persian are handled natively.

Yes, DeepL added Arabic support, and for text translation it can produce good results. The honest distinction is depth: as of writing its RTL document handling is more limited than its European-pair strengths, while TranslateX was engineered Arabic-first, with full layout mirroring for documents and Arabic-specific QA in the review editor. Test both on a real Arabic document of yours.

It depends on the pair, and we will not invent benchmark numbers. On European pairs DeepL is deservedly well regarded. On Arabic and English, TranslateX's Arabic-first engineering, glossaries, memory, and review workflow are built for exactly that job. The only test that matters is your own documents, and both products let you start free.

DeepL does offer glossaries, and credit where due. TranslateX pairs term bases with translation memory, an organization style guide, and a review editor that surfaces glossary hits per segment with QA checks, then merges the edits back into the original document layout. See glossary and memory and style governance.

As of writing, DeepL is a cloud service. TranslateX offers full on-premises deployment via Docker with a local translation model, including documented air-gapped installs, plus roles, LDAP, MFA, audit logs, and white-labelling. See security and on-prem.

Both products can translate documents. TranslateX rebuilds the translated copy in the original design across PDF, Word, PowerPoint, Excel, and images, and mirrors layouts when the direction flips between Arabic and English, which is the hard case. For RTL documents specifically, that mirroring is the difference to test. See layout preservation.

We do not publish head-to-head accuracy claims against Google, and you should be suspicious of anyone who does without showing their work. The honest difference is the workflow around the engine: layout preservation, enforced glossaries, translation memory, and a review editor with QA checks. The right test is to run both on one of your own documents.

Google Translate accepts document uploads, but as of writing the translated result comes back with simplified formatting, so tables, fonts, and slide design need rebuilding. TranslateX returns a translated copy in the original design, mirrored when the reading direction flips. See layout preservation.

No. The consumer product is a cloud service, so whatever you paste or upload travels to Google. TranslateX offers full on-premises deployment with a local translation model, including documented air-gapped installs. See security and on-prem.

The TranslateX desktop app is a free download, premium comes with a 1-month free trial, and enterprise deployment is a conversation. For truly casual translation, Google Translate is free forever and that is fine: use each tool for what it is best at. See pricing.

For a quick personal gist, no, it remains a great tool. The risk starts when staff paste confidential documents into a public website, or when published documents carry inconsistent terminology. That is the work to move to TranslateX, where glossaries, review, and deployment controls exist.

Word's built-in translate creates a new machine-translated document, and as of writing the result comes back with simplified formatting, so styles, tables, and layout need rework. TranslateX returns a translated copy in the original design, applies your glossary, and adds a review step. See layout preservation.

As of writing, PowerPoint has no built-in one-action full-deck translation comparable to TranslateX's Translate all slides. TranslateX translates the entire deck, preserves slide structure, and mirrors shape positions when the direction flips between Arabic and English. See Office add-ins.

Custom terminology exists in Microsoft's ecosystem, mainly through the Azure AI Translator developer platform, which is aimed at engineers building applications. TranslateX makes glossaries an everyday organizational control: term bases per language pair, managed in the workspace, enforced automatically on every job across every format including PDF and images.

Yes, that is the normal case. TranslateX ships add-ins for Word and PowerPoint, a Windows desktop app with right-click translation in Explorer, and LDAP / Active Directory sign-in for the enterprise platform, so it slots into a Microsoft-centric environment rather than fighting it. See the desktop app.

Microsoft Translator supports Arabic text well for comprehension. TranslateX was engineered Arabic-first for documents: full RTL layout mirroring, Arabic punctuation and typography handling, and Arabic-specific QA checks (hamza seats, taa marbuta, kashida) in the review editor. For publish-ready Arabic documents, test both on a real file. See Arabic and RTL.

Usually not, and we would not advise trying. Certified and sworn translations, and creative transcreation, still need professional human linguists. What TranslateX replaces is the routine volume: the reports, contracts, and decks where speed, consistency, and confidentiality matter more than a stamp. Most organizations resize the agency contract rather than cancel it.

For routine institutional documents, that is exactly the workflow much of the industry itself uses: a machine first draft with human post-editing. TranslateX gives your bilingual staff a proper editor for it: segment statuses, QA checks including Arabic-specific ones, track changes, and a final file that keeps the original layout. See document translation.

Agency pricing scales with every word, so a heavy quarter means a heavy invoice. TranslateX is software: the desktop app is a free download, premium has a 1-month free trial, and enterprise deployment is a conversation. For a rough sense of your volume, try the free word count and pricing estimator.

An NDA is a promise; on-prem is a boundary. Agencies handle files through external staff and systems, however professionally. TranslateX can run entirely on your own servers with a local translation model, including air-gapped installs, so sensitive documents never leave your network. See security and on-prem.

Yes, and it is a good pattern. Agencies can work from TranslateX's layout-preserved first drafts and your workspace glossary, and your translation memory moves in standard TMX format, with XLIFF supported for segment-level workflows. Outsourced jobs start further ahead, and everything the agency approves flows back into your memory.

Stripping tashkeel removes the combining marks U+064B through U+065F plus the superscript alef U+0670. Kashida removal deletes the tatweel character U+0640. Alef normalization maps أ, إ, and آ to bare ا; the taa marbuta option maps ة to ه. Digit conversion swaps ٠ through ٩ with 0 through 9 in either direction.

When building search keys, term bases, or translation-memory match text, where you want spelling variants to collapse into one form. Do not apply them to text you will publish: the changes are lossy and alter correct spellings.

A memory matches on the text itself, and a segment written with tashkeel is a different string from the same segment without it. Cleaning input to a canonical form means previously approved translations are found instead of retranslated.

No. Every operation is plain JavaScript running in your browser; there are no network calls and nothing you paste leaves the page.

Either source = target on each line, or two tab-separated columns, which is what you get when you copy two columns from Excel or Google Sheets. Blank lines are skipped.

Excel assumes a legacy encoding for plain CSV files and mangles Arabic text into mojibake. The invisible UTF-8 BOM at the start of the file tells Excel to read it as UTF-8, so Arabic targets open correctly.

A term base should give exactly one approved target per source term. Two rows with the same source means the terminology decision has not been made yet; the flag tells you to settle it before importing.

No. Parsing, preview, and CSV generation all run in your browser; the download is created locally with no network calls.

They are industry rules of thumb: roughly 2,500 words per day for professional human translation and 6,000 words per day for post-editing machine output. Real throughput depends on the language pair, subject matter, and the person; adjust the fields to match your team.

Only roughly. The model divides throughput by team size, which works when documents split cleanly across people. Splitting a single document adds coordination and consistency overhead the estimate ignores, so treat results for large teams as optimistic.

No. The estimator is plain JavaScript running in your browser; nothing you type is sent anywhere.

Word count times your per-word rate, times a rush multiplier (1x standard, 1.25x for 48-hour, 1.5x for 24-hour). It is an illustrative model of how agencies quote, not a price list.

Rates vary widely by language pair, direction, domain, and country, from well under 0.10 to over 0.50 in major currencies. Use a figure from a recent quote for your pair, or leave the default to compare scenarios.

Yes. Counting is Unicode-aware and splits on whitespace, so Arabic, English, and mixed-direction text all count correctly, including text with diacritics.

No. Counting and the estimate run entirely in your browser; nothing you paste leaves the page.

Within one business day for demo requests and support. Enterprise scoping calls are usually booked within the same week.

Yes, that is our favorite kind of demo. Bring a PDF, a Word file, or a deck, and we will translate it live with layout preserved.

Yes. More than 24 government entities work with Lisan, and TranslateX supports sovereign-cloud and on-premises deployment for sensitive environments.

Lisan (LISAN AI - FZE) is headquartered in Dubai, United Arab Emirates, and works across the region.

Because the text hides in dozens of surfaces (runs, tables, text boxes, headers, SmartArt, speaker notes, images), the translated text is a different length than the original, and the layout may need to mirror when direction changes. Each translation must also go back exactly where it came from, with formatting intact.

It varies by pair. Arabic to English typically expands around 20 to 40 percent, English to German grows too, and Chinese contracts sharply. Any layout designed snugly for one language will overflow or look sparse in another unless the tool adjusts sizing and flow.

PDF is a print format: it stores positioned glyphs, not paragraphs or tables. A translator has to reconstruct reading order, paragraph boundaries, and table structure from coordinates, essentially reverse-engineering the document from its printout, and scanned PDFs add OCR on top of that.

Pixels carry no text, so the pipeline must OCR the image, translate the recognized words, then re-render the image with the original text masked and the translation drawn in its place. Tools that skip the re-rendering step return documents with untranslated pictures inside them.

A glossary works at the level of terms: one approved translation per name, title, or term of art, enforced everywhere. A translation memory works at the level of segments: whole approved sentences stored and reused when they recur. Glossaries decide the words, memory decides the sentences.

Names of the organization, departments, and key people; job titles and honorifics; product and program names (often marked do-not-translate); legal and technical terms of art; and transliteration choices. Keep it small and owned: five hundred trusted terms beat five thousand ignored ones.

A fuzzy match is a stored segment that is similar but not identical to the new one, offered with a similarity score. It only suggests because a small character difference can carry a large meaning difference: a changed year is trivial, a changed 'shall' to 'shall not' is not. A human confirms or adapts it.

Institutional documents repeat: boilerplate, disclaimers, standard clauses. Every approved segment adds to the memory, so each new document starts with more of itself already translated and locked. Coverage that begins around a tenth of a document can reach half of similar documents within a few years.

Yes, through TMX, the standard exchange format for translation memories. Any serious platform, TranslateX included, imports existing TMX so past work starts matching immediately, and exports your memory whenever you want it. XLIFF plays the same role for in-progress segment work.

Phrase-based MT chopped sentences into short fragments, translated each from statistics, and stitched the pieces back together, so long sentences came out mangled. Neural MT reads the whole sentence, models its meaning, and writes a new sentence in the target language, which is why modern output is fluent and grammatical.

Because fluency and accuracy are separate. A neural model can drop a negation, alter a number, or invent a clause while still producing polished prose, so errors are harder to spot than in the old broken-sounding output. For consequential documents, a human review pass plus automated QA checks catch what fluency hides.

Words underdetermine meaning; the surrounding document supplies register and domain. 'Execute the agreement' means signing in a contract and simply doing something in a chat message. Models that see the whole document pick up these signals, which is why document-level translation beats pasting isolated sentences.

Match the tier to the consequences of an error. Internal gist and high-volume batches suit the fastest tier; published or legally binding documents deserve the best-quality tier plus human review. In TranslateX the choice is per job, from Lisan L1 for speed to Lisan L3 for best quality, with a local model option for on-premises deployments.

The workflow where a machine translates the whole document first and a human reviewer then walks the output segment by segment, accepting, editing, or rejecting each one. The machine does the typing and the human does the judgment, which is faster than human translation and safer than raw machine output.

The mechanical failure modes: numbers that differ between source and target, untranslated segments, doubled spaces, missing or mismatched punctuation, and length ratios that suggest dropped content. They exist because these errors hide inside fluent prose where human eyes skip past them.

Light post-editing fixes only errors of meaning: wrong numbers, dropped negations, mistranslations, leaving awkward but accurate phrasing alone. Full post-editing also fixes terminology, style, and tone to publication standard. Light suits internal working documents; full is for anything that ships under your name.

When the reader knows the text is machine output, an error costs nearly nothing, and the alternative is no translation at all: gisting foreign news, triaging inbound documents, internal lookups. It is never acceptable for legal instruments, published material, or any document where a wrong number or dropped negation has consequences.

The whole visual structure flips, not just the text alignment: the page anchors top right, sidebars and images swap sides, table columns run right to left, and presentation shapes sit in mirrored positions. A translated document that keeps the old skeleton feels backwards to native readers even when every word is correct.

Both are legitimate; convention varies by country, institution, and document type. Gulf financial material often keeps Latin digits, while more formal traditions prefer Arabic-Indic (٠١٢٣). The real rule is consistency: pick one convention per document type, record it in the style guide, and never mix systems within a document.

Taa marbuta (ة) loses its two dots and becomes haa (ه), silently changing the word. Hamza sits on different seats (أ إ ؤ ئ ء) depending on rules subtle enough that native writers get them wrong, and models trained on native writing inherit the mistakes. Both are mechanical enough for Arabic-aware QA checks to flag automatically.

A mixed line contains multiple directional runs (RTL Arabic, LTR acronyms, LTR digits), and the Unicode bidirectional algorithm decides their visual order. It usually guesses right, but punctuation at run boundaries, parentheses, and adjacent Latin fragments are classic failure points, which only correct handling in the layout engine fixes.

Not historically: it is a legitimate typesetting device for justification and emphasis. In digital body text it mostly causes harm, breaking search, matching, and glossary lookups because elongated and plain spellings are different strings. Most institutional style guides ban it from body text and let the justification engine do the work.

For gisting, yes: understanding an inbound document, triaging a pile, skimming a competitor's announcement, provided the output is labeled as machine translation. For anything published, signed, or otherwise consequential, no. Fluent machine output can still contain wrong numbers, dropped negations, and mistranslated terms that read perfectly.

By consequence. Gist-only documents need no review, internal working documents need a light pass fixing errors of meaning, published material needs full post-editing to publication standard, and legally binding texts need bilingual legal review and often certified translation. Decide the level before review starts, not during.

Wherever the law requires it: courts, immigration, and official government filings in many jurisdictions accept only certified or sworn translations, where the translator's accreditation and liability are part of the product. Machine translation plus review does not substitute for that, and a good workflow routes those documents out to certified translators explicitly.

The text leaves your network and jurisdiction the moment it is submitted, lands on servers governed by another country's laws, and is retained according to terms of service you cannot verify or control. For sensitive documents, confidentiality has to come from architecture, not from a provider's policy.

On-premises means the software runs on your own servers inside your own network, with data never leaving your perimeter, though the environment may still have internet access. Air-gapped removes the internet path entirely: no outbound connections, with updates delivered by controlled offline transfer. Air-gapped is the standard for classified environments.

Three heavy pieces: the translation model itself, OCR for scanned PDFs and images, and the document pipeline that parses and rebuilds PDF, Word, PowerPoint, and Excel files, plus local storage for glossaries, translation memory, and audit logs. If any of these calls a cloud service, documents are leaving.

Yes. TranslateX offers a local on-premises model option, deploys via Docker, and air-gapped installation is a documented, supported path, alongside roles, LDAP and Active Directory sign-in, MFA, audit logs, and SIEM export for the security team.

Desktop publishing: the rebuilding phase after translation, where someone restores the formatting the copy-paste workflow destroyed, re-flowing tables, fixing fonts, mirroring layouts, and fitting expanded text back into slides and pages. Agencies typically bill it as a separate line item because it often takes longer than the translation itself.

Two mechanical reasons. The direction flips between right-to-left and left-to-right, so alignment, table column order, and slide layouts all need mirroring. And Arabic is compact, so the English version typically needs roughly 20 to 40 percent more room, which overflows text boxes and cells that fit before.

They are illustrative, chosen to show the shape of the problem rather than a benchmark. The honest test is your own last document: split the elapsed time into translating, rebuilding, and reviewing, and see which bucket dominates.

Because the whole document moves through one pipeline, a glossary and translation memory can be enforced across every segment: the same ministry, title, or product name is rendered the same way everywhere, in this document and in the next one, instead of being re-guessed fragment by fragment.

Arabic attaches conjunctions, prepositions, articles, and pronouns directly onto words, so one Arabic token can correspond to four or five English words. Add a right-to-left script, formal-versus-dialect diglossia in the training data, and far less parallel text than European pairs, and the error rate on formal documents climbs fast.

The same Arabic name or institution being rendered several different ways in one translation, for example a ministry appearing under three different English names across a single report. It happens because generic engines guess each occurrence independently. Enforced glossaries fix it by pinning the official rendering once.

Roughly 20 to 40 percent is a reasonable planning range, since Arabic is more compact than English. That expansion is why translated documents overflow tables, slides, and text boxes unless the translation system rebuilds the layout as it translates.

TranslateX is built for written institutional documents, which are almost always Modern Standard Arabic, and that is the register its engine is optimized for. Arabic to English with native RTL and typography handling is the flagship pair.

Yes. The desktop app is free to install and use for text and screen translation. Document translation at full strength comes with Premium, and there is one month free to try it.

Yes. TranslateX supports Windows 10 (build 19041 and later) and Windows 11, on 64-bit systems.

The app installs as an MSIX package through Windows App Installer, so updates download and apply automatically. No manual reinstalls.

Not yet, they are on the roadmap. The enterprise web platform works in any modern browser on any operating system today.

Yes. MSIX packages deploy through Intune and standard enterprise device-management tooling, and enterprise plans support on-premises hosting of the platform itself.

PDF, Word (DOCX), PowerPoint (PPTX), Excel (XLSX and CSV), plain text, images and screenshots, and anything you capture on screen. You can batch up to 10 files at a time, up to 100 MB each.

Yes. TranslateX rebuilds the translated document in the original design: fonts, tables, spacing, and slide structure. When translating between right-to-left and left-to-right languages, layouts are mirrored so the result reads naturally.

More than 20 languages including Arabic, English, French, German, Spanish, Portuguese, Chinese, Japanese, Korean, Hindi, Turkish, Urdu, and Persian, with automatic source-language detection. Arabic and English are the flagship pair.

Yes. Glossaries enforce your approved terms, translation memory reuses your past translations, and the style guide option applies your organization's writing style. See style governance.

Yes. Full on-premises and local deployment options exist for sensitive environments, including a local translation model and documented air-gapped installs. See security and on-prem.

Press F6 (or click the toolbar icon), draw a box around anything on your screen, and TranslateX reads the text, translates it instantly, and can copy the result or export a document. It works on scans, dashboards, and apps that block copy-paste.

The visuals on this page are simplified recreations, drawn to stay honest about proportions and flow. The official videos on the homepage show the real product.

Download the free desktop app and right-click a document. Two minutes, no meeting needed. Get it here.

The review editor, shared glossaries and memory at team scale, reports, style governance, and on-premises deployment come with the enterprise platform. See pricing.