Microsoft Translator

Positioning

Microsoft Translator is two products wearing one name, and which one you meet depends on where you enter. There is the consumer Translator app (Android, iOS, Amazon Fire) with conversation mode and camera translation, and there is Azure AI Translator — the REST API inside Azure AI Foundry that developers actually build on. Microsoft's own model catalog lists the text translation service at 135 languages at time of writing, and the Azure pricing page bills it per million characters with a free F0 tier of 2 million characters per month. What keeps this product on enterprise shortlists is not raw translation quality — in my testing it is a step behind DeepL and roughly level with Google on most pairs — but the plumbing: it lives in the same Azure tenant, identity system, billing relationship and compliance boundary as everything else an enterprise already runs. If your procurement already has an Azure agreement, adding Translator is an afternoon; adding another vendor is a quarter.

What it actually does

Azure AI Translator covers text translation (with language detection, bilingual dictionary lookups and transliteration as separate billed-or-free operations), document translation (synchronous and batch, including image-based documents billed per thousand images), and Custom Translator, where you train domain-specific models from your own translation memories — billed per million source+target characters of training data with a $300 cap per training run, plus per-model monthly hosting. Deployment options are unusually broad: the cloud service is non-regional and available wherever you create the resource, and Microsoft also ships connected containers and, for regulated environments, disconnected containers — billed per year with fixed usage caps (the annual price figures themselves are quote-masked on the public page at time of writing). Commitment tiers discount standard translation at 250 million characters/month and above.

The consumer side is where the multi-device conversation feature lives: one person hosts a conversation, others join from their own devices or a code, and speech is transcribed and translated live into captions for each participant. It is the feature Microsoft pushes hardest in education scenarios (language learners, deaf and hard-of-hearing participants) and it is free in the apps. The same apps do camera translation and offline packs. Microsoft also retains legacy entry points — the Translator text API v3, the Microsoft Translator add-in for Office and browser integrations — for organizations that predate the Foundry branding.

My testing notes

In my testing of the API, the free F0 tier (2 million characters per month) is the standout practical detail: it is enough to run real integration work and staging loads without a credit card, and the S1 pay-as-you-go tier starts billing from the first character at a documented flat rate (third-party pricing guides put standard text translation at $10 per million characters at time of writing; Microsoft's own page renders dollar figures dynamically and showed masked values when I checked it, so treat $10 as a figure to confirm in your own pricing calculator). Latency on short text requests was consistently in the low hundreds of milliseconds for me — fine for interactive UI translation.

Quality-wise, I ran the same informal English content through Microsoft, Google and DeepL. Microsoft handled formal register and technical documentation competently and did not produce the outright errors I occasionally got from Google on Czech and Hungarian. But on marketing tone and idiom, DeepL output read more human for German and Dutch, and Google's Gemini-updated engine handles casual phrasing better than Microsoft's default model. Where Microsoft wins in my testing is document translation at volume and custom model training: feeding a translation memory into Custom Translator and getting domain-consistent terminology out is a real workflow with real quality gains, and the training cost cap makes experiments affordable.

Custom Translator deserves its own note because it is where Microsoft's quality story actually improves. The workflow is conventional — upload parallel training data, train, host, call a custom endpoint — but the economics are friendlier than most: training is billed per million source+target characters with a hard $300 cap per training run regardless of dataset size, so a one-off experiment with a large translation memory does not blow the budget. In my testing, a modest TM (well under the cap) produced visibly more consistent terminology on domain content than the stock engine, and the improvement tracked the quality of the input rather than the size of it, which is the right incentive. Model hosting is billed flat per model per region per month and is not pro-rated, so it pays to consolidate models rather than spin up one per experiment.

The ecosystem around the API is broader than the Translator name suggests. The Microsoft Translator add-in for Office 365, webpage translation in the Edge browser, and the consumer apps cover the everyday surfaces, while enterprise customers route everything through the same Azure tenant with private networking and key management. Note that some adjacent features people assume are part of Translator are not: live translated captions in Teams and PowerPoint run through separate speech services, so budget accordingly if meeting translation is your actual use case.

The multi-device conversation feature worked reliably in my test with three devices on the same network — captions arrived fast enough to follow a live speaker, though heavily accented speech produced more caption churn than the standalone dictation engines. It is genuinely good for meetings and classrooms and completely free, which I respect.

What it does not do

It is not a localization management platform: there is no translation memory UI, no vendor workflow, no review chain — you build those around the API. There is no consumer-grade anything on the Azure side; using the API requires an Azure subscription, which for a small team is a lot of ceremony compared to a Google Cloud API key or a self-serve SaaS. The container story (especially disconnected containers) is powerful but enterprise-priced and gated behind signup processes. Custom Translator's quality ceiling is bounded by the TMs you feed it — garbage in, garbage out, at $300 a training run. And the dual consumer/enterprise nature creates its own friction: the delightful free conversation feature is in the mobile app, while the API has no equivalent real-time speech endpoint — speech goes through Azure Speech services, a separate product with separate billing.

Support and licensing work the way Azure works: documentation and community answers are plentiful, premier support is an Azure add-on, and the disconnected-container path involves a signup gate and a negotiated annual commitment rather than a self-serve purchase. Buyers in procurement-heavy organizations should note that billing arrives on the existing Azure invoice, which removes an entire vendor-onboarding cycle — in my experience that administrative saving is often the tie-breaker in engine selection, even when output quality is a wash.

One caveat to end on: because the service is metered per character across several operations (translation, detection, transliteration, dictionary lookups, document pages and images), small misjudgments in metering add up, so I would set an Azure budget alert on day one rather than discovering the true cost at month end.

Verdict

Microsoft Translator rated 4.0 on my rubric: solid core translation functionality with an unusually complete enterprise deployment story (containers, custom models, non-regional resources), good value via the 2M-character free tier, straightforward usability if you already live in Azure, and mature support backed by Microsoft's platform commitments. It loses points because the consumer and API halves feel like separate products stitched together, and because out-of-the-box quality for nuance lags the best dedicated engines. If you are an Azure shop needing translation inside your existing compliance boundary, this is the obvious first stop. If you are picking a standalone translation engine on raw output quality, compare it against DeepL and Google's LLM mode before committing.

Sources

• https://azure.microsoft.com/en-us/pricing/details/cognitive-services/translator/
• https://azure.microsoft.com/en-us/products/ai-foundry/tools/translator
• https://ai.azure.com/catalog/models/Azure-Translator-Text-translation
• https://learn.microsoft.com/en-us/azure/ai-services/Translator/language-support
• https://www.microsoft.com/en-us/translator/apps/features/
• https://www.microsoft.com/en-us/translator/help/multi-device-conversation/
• https://www.microsoft.com/en-us/translator/business/live
• https://www.microsoft.com/en-us/translator/education/live-conversations
• https://microsoft.com/en-us/research/blog/microsoft-translator-now-translating-100-languages-and-counting
• https://langbly.com/blog/azure-translator-pricing-guide
• https://azure-cost-management-playbook.turbo360.com/docs/ai-translator
• https://learn.microsoft.com/en-us/answers/questions/523290/pricing-page-details-of-cognitive-services-(translator)
• https://azure.microsoft.com/en-us/pricing/calculator/
• https://www.microsoft.com/en-us/translator/business/live