4.4 Azure AI Translator Service
Key Takeaways
- Azure AI Translator provides real-time and batch text translation across 100+ languages using neural machine translation (and increasingly LLMs).
- Document translation preserves original formatting (tables, headers, images, layout) for PDF, DOCX, PPTX, XLSX, HTML, and more.
- Transliteration converts text from one script to another WITHOUT changing the language (e.g., Kanji to Latin); it is not translation.
- Custom Translator trains a domain-specific model from your bilingual parallel data for specialized legal, medical, or technical terminology.
- Translator handles text; Azure AI Speech translation handles spoken audio — choosing between them is a common exam decision.
Quick Answer: Azure AI Translator provides text translation for 100+ languages, document translation (preserving formatting), transliteration (script conversion), and Custom Translator for domain-specific models. It is different from speech translation — Translator handles text, Speech handles audio.
What Is Azure AI Translator?
Azure AI Translator is a cloud-based machine-translation service that enables real-time and batch text translation across more than 100 languages and dialects. It uses neural machine translation (NMT) — and, increasingly, large language models — to produce fluent, context-aware translations rather than the word-by-word output of older statistical systems.
A defining trait for the exam: Translator is a stateless, text-in / text-out service. You do not train it for general use, you do not give it audio, and it does not "understand" sentiment or intent — it only converts text from a source language into one or more target languages.
Core Capabilities
1. Text Translation
Translate text strings from one language to another:
| Feature | Description |
|---|---|
| Languages | 100+ supported languages and dialects |
| Auto-detect source | Automatically identify the source language so you do not have to specify it |
| Multi-target | Translate into several target languages in a single request |
| HTML support | Translate HTML content while preserving the tags |
| Profanity handling | Options to mark, delete, or ignore profanity |
Because auto-detect is built in, you usually do not need a separate language-detection call before translating — Translator can identify the source language itself.
2. Document Translation
Translate entire documents while preserving their original formatting:
| Feature | Description |
|---|---|
| Supported formats | PDF, DOCX, PPTX, XLSX, HTML, TXT, Markdown, and more |
| Formatting preserved | Tables, images, headers, fonts, and overall layout maintained |
| Batch translation | Translate many documents in one operation from blob storage |
| Glossary support | Apply custom glossaries to keep terminology consistent |
This is the feature to choose whenever a scenario says "translate the document and keep the formatting/tables/layout." Plain text translation would strip the structure; a generative model could translate but is not guaranteed to preserve layout.
3. Transliteration
Convert text from one script to another without changing the language:
| Source | Transliterated |
|---|---|
| Japanese (Kanji): TOKYO in characters | Latin script: "Tokyo" |
| Arabic script greeting | Latin script: "marhaba" |
| Hindi (Devanagari) greeting | Latin script: "namaste" |
The text remains the same language — only the writing system changes. This helps users read content in an unfamiliar script. Transliteration is frequently confused with translation on the exam: translation changes the language, transliteration changes the script.
4. Custom Translator
Build a domain-specific translation model for higher accuracy in specialized fields:
| Step | Description |
|---|---|
| Upload parallel data | Provide bilingual document pairs (source + human translation) |
| Train a custom model | The service learns your domain's terminology and style |
| Test and evaluate | Compare the custom model against the generic model (BLEU score) |
| Deploy | Use your custom model via its own category ID |
When to use Custom Translator: legal documents with precise legal terminology, medical content with specialized vocabulary, technical documentation with product-specific terms, or any domain where the generic model misses nuance. The trigger phrase is "specialized/industry terminology."
Translator vs. Speech Translation
| Aspect | Azure AI Translator | Azure AI Speech Translation |
|---|---|---|
| Input | Written text | Spoken audio |
| Output | Translated text | Translated text or speech |
| Use case | Documents, web pages, app strings | Live conversations and meetings |
| Real-time | Yes (and batch) | Yes |
| API/SDK | Translator REST API | Speech SDK |
Dictionary Lookup and Alternatives
Beyond plain translation, the Translator API can return a dictionary lookup — alternative translations for a word along with examples of usage in context. This is useful when a single word has several valid translations and an application wants to surface options to the user. It is a smaller feature, but it reinforces a key idea: Translator is a suite of text operations (translate, transliterate, detect, dictionary lookup), not a single endpoint.
Choosing Translator vs. a Generative Model
A modern wrinkle is that a generative model (GPT in Azure OpenAI) can also translate text. So when is dedicated Azure AI Translator the better choice?
| Factor | Azure AI Translator | Generative model (GPT) |
|---|---|---|
| Formatting preservation | Strong (document translation keeps tables/layout) | Not guaranteed |
| Cost at scale | Lower per character for bulk translation | Higher per token |
| Determinism | Consistent output for the same input | Can vary run to run |
| Custom terminology | Custom Translator + glossaries | Prompt/grounding only |
| Broad language coverage | 100+ languages out of the box | Varies by model and language |
For straightforward, high-volume, format-preserving translation, Translator is the right tool. A generative model is more appropriate when translation must be blended with reasoning, summarization, or stylistic rewriting in the same step. On the AI-900, a scenario that simply says "translate these documents into many languages and keep the layout" points to Azure AI Translator.
Common Exam Scenarios
| Scenario | Answer |
|---|---|
| Translate a website's text strings in real time | Text translation |
| Translate 500 PDF manuals and keep tables/headers | Document translation |
| Show a Japanese place name in the Latin alphabet | Transliteration |
| Improve accuracy for legal contract translation | Custom Translator |
| Translate a live spoken conversation | Azure AI Speech translation (not Translator) |
On the Exam: If the question involves translating written text or documents, the answer is Azure AI Translator. If it involves translating spoken audio in real time, the answer is Azure AI Speech translation. "Preserve formatting" -> document translation; "change the script, not the language" -> transliteration; "industry-specific terminology" -> Custom Translator.
A company needs to translate their product documentation from English into 15 different languages while preserving the original formatting (tables, headers, images). Which Azure service should they use?
What is transliteration?
When should you use Custom Translator instead of the standard Azure AI Translator?
A team needs to translate text but does not know which language the incoming text is written in. What should they do with Azure AI Translator?