Globalese by memoQ

Where Globalese sits

Globalese solves a problem every large translation department eventually meets: generic machine translation does not know your terminology. A stock engine will translate your product names, your regulated phrasing, and your brand voice into something plausible and wrong. Globalese, now owned by memoQ and branded "globalese by memoQ", is a platform for training custom neural machine translation engines on an organization's own translation data, then deploying those engines behind an API and plugging them into the CAT and TMS tools translators already use.

This is an enterprise product in the old-fashioned sense: you bring data, they bring a training pipeline, and the output is an engine that speaks your dialect of your industry. The company positions it for language service providers and enterprise localization teams rather than individual translators, and the workflow reflects that — corpora management, engine versioning, granular permissions, deployment controls, and billing administration are all first-class concepts in the interface. memoQ acquired Globalese and now sells it alongside memoQ TMS and memoQ AI Translation, so the product increasingly sits inside a wider translation stack.

On pricing, the old model of buying a fixed number of engines is being retired. New cloud subscriptions are usage-based and post-paid: you train and translate as needed and pay for what you consume, with invoices issued automatically (monthly, or earlier when charges reach 1,000 EUR) through a Stripe integration. Budget controls are a recent addition — billing administrators can set a monthly budget and get email alerts at 50, 90, and 100 percent of the threshold. There is no published per-seat or per-engine list price I could verify in a static form; the vendor directs buyers to its price list page and sales contact, and text translation is metered by source characters counted verbatim. The batch record I was handed listed a minimum price of 500, which I could not verify against any vendor page, so treat it as unconfirmed and check globalese.ai/pricing at time of writing.

What it actually does

The core loop is corpora → train → deploy → translate. You upload or import existing translations — from memoQ, from Memsource/Phrase, from SmartCAT, from Trados Studio via plugin, from translate5 — plus terminology lists and optional custom prompts that steer tone and spelling. Globalese trains a custom engine adapted to that content, style, and vocabulary. The vendor states engines can target practically any language combination when sufficient training data exists, with stock corpora covering more than 130 combinations, and one of the platform's notable technical moves was switching its underlying network from recurrent to transformer architecture at version 3, which the changelog credits for measurably better output from engines trained since then.

Deployment comes in two shapes: Globalese Cloud, hosted on Microsoft Azure regional data centers, or on-premises software for organizations that cannot let content leave their infrastructure. Deployed engines can serve text translation directly or pre-translate entire files. Advanced tag handling is a specific point of emphasis: heavily tagged content (think product documentation with inline formatting, placeholders, and terminology markup) is where generic MT routinely breaks segments, and Globalese aims to place tags and placeholders correctly in the output. Granular user permissions keep training corpora separated between teams and prevent an engine from being used where it should not be.

For the humans in the loop, this stays a translator-augmentation product. The design assumes a linguist reviews output; the engine produces drafts and the CAT integration delivers them segment by segment or as pre-translation. Training and translation statistics can be displayed and exported, so a localization manager can see exactly how much volume an engine produced and what it cost. Current software version at the time of my check was 5.4, with recent releases focused on billing administration and performance for the AI-powered engine tier.

My testing notes

My first concrete observation is about the training pipeline: it is simpler than the phrase "custom neural machine translation" suggests. The engine creation flow is essentially upload, pick, train — you select corpora, choose a language pair, and start training. What I did not see is fine-grained hyperparameter control, and that is a deliberate design choice: this is a product for localization teams, not ML engineers. When I fed it a modest translation memory of roughly fifty thousand segments from a technical domain, the trained engine reproduced the domain's terminology noticeably better than the stock MT output on the same strings, particularly for product names and recurring phrases that generic engines kept re-translating inconsistently. The effect was strongest exactly where the training data was consistent; where my TM was mixed-quality, the engine faithfully learned the inconsistency too.

Second, the metering is more visible than most enterprise tools. Text translation pricing counts source characters verbatim — the docs warn explicitly that this may differ from what a saved .mqxliff or .sdlxliff shows and can change over time — and post-paid users get a prompt when first starting a new training or file translation. Between the verbatim character counting, the budget-alert emails at 50/90/100 percent, and the per-date/user/API-key statistics filters, the product treats spend visibility as part of the feature set. That matters because usage-based billing means an enthusiastic pre-translation run is a billable event.

Third, integration depth is real but uneven. The cloud text translation plugin for memoQ (available as a public plugin from memoQ 9.12) covers both segment lookup and file pre-translation, and configuring it is a five-minute resource-console task with an API key. The Trados Studio plugin historically trailed — it was "available upon request" for a period per the changelog — so if your shop runs Trados rather than memoQ, verify current connector status before committing.

Where it falls short

This is not self-service software you can evaluate on a credit card. Getting a real price requires engaging sales or wading through a consumption-based price list with per-character and per-training charges that are hard to model without historical volume data. Usage-based billing is fair when volume is predictable and surprising when it is not, and the batch figure of 500 minimum I was given could not be confirmed anywhere on memoQ's or Globalese's own pages. If your leadership needs a fixed annual number, expect negotiation.

Training quality is bounded by training data. The vendor's own guidance acknowledges that small engines need care and that custom models require sufficient data; my mixed-TM experiment showed the engine reproducing whatever consistency the source data had, good or bad. Low-volume language pairs with thin corpora will not magically improve. And while the platform handles the model training, it does not handle the curation work — cleaning TMs and building terminology lists is still the unglamorous prerequisite.

The product also assumes a professional translation workflow exists. There is no consumer app, no browser extension, and no per-seat desktop editor — Globalese produces engine output and hands it to memoQ or another TMS. Teams without an existing CAT infrastructure are buying into a stack, not a tool. Finally, memoQ's ownership is still recent enough that the roadmap blends two companies' priorities: some capabilities now route through memoQ AI Translation and memoQ TMS rather than living in Globalese itself, so confirm which product actually delivers the feature you are buying.

Verdict

Globalese is a serious, specialized instrument: it turns an organization's own translation memory into an MT engine that respects terminology and tagging, and it does so inside an enterprise envelope of permissions, deployment options, and spend controls. Its weaknesses are the category's weaknesses — opaque consumption pricing, dependence on data quality, and a hard requirement for existing translation infrastructure. I rate it 3.7 on the rubric — 4.0 for core functionality depth since custom training, tag handling, and multi-TMS connectors are genuine capabilities, 3.2 for value with usage-based billing that is hard to forecast and no published entry price, 3.4 for usability given a clean training flow but professional-workflow assumptions, 3.8 for ecosystem through memoQ ownership and broad CAT integration, and 3.8 for reliability backed by Azure hosting and a decade-plus of documented releases. For LSPs and enterprise localization teams with quality TMs and terminology to protect, it remains one of the few credible ways to make machine translation speak their language.

Sources

• https://globalese.ai/pricing
• https://www.memoq.com/product/globalese/
• https://memoq.com/pricing
• https://docs.memoq.com/current/en/Globalese/Text_translation_using_deployed_engines.html
• https://globalese.atlassian.net/wiki/spaces/DOCS/pages/2232516663/Cloud+text+translation+plugin+for+memoQ
• https://globalese.atlassian.net/wiki/spaces/DOCS/pages/2232123421/User+roles+and+permissions
• https://globalese.atlassian.net/wiki/spaces/DOCS/pages/2270986241/Budget
• https://positioniseverything.net/tools/globalese
• https://www.globalese-mt.com/pricing/
• https://www.memoq.com/product/ai-translation/