Translation AI is Google Cloud’s set of services for translating text and documents between languages automatically.

The problem of language

Businesses today serve customers, partners and employees all over the world. Product pages, support chats, contracts, manuals, reviews and training videos all exist in one language but are needed in many. Translating everything by hand is slow and costly, and it cannot keep up with content that changes every day. Yet leaving content untranslated means losing customers who simply prefer, or need, to read in their own language.

What the service does

Translation AI, delivered mainly through the Cloud Translation API, lets your applications send text or documents to Google’s translation models and receive the translated result in moments. It supports more than a hundred languages, can detect which language a piece of text is written in, and can be customized so that your brand names, technical terms and style are handled correctly. It is available to programs through an interface, and to people through tools such as Vertex AI Studio for experiments.

Twelve words you will meet in every translation project.

Keep these close. Each is explained again with examples in the chapters ahead, and seeing how they fit together is the quickest way to learn them.

A short history that explains why today’s translations sound natural.

Why context matters

The English word bank can mean a place for money or the side of a river. A word by word system must guess. A neural or language model reads the whole sentence and sees words like river or loan, so it chooses correctly. The longer and better the context, the better the translation.

One name, several doors into the same house.

Google Cloud offers translation through several surfaces, and each gives a different level of control. The Cloud Translation API is the main programming interface. It comes in two editions, Basic and Advanced. Behind them sit several models. On top you can use Vertex AI Studio, a visual workspace for trying models, uploading example sentence pairs, and fine-tuning. Together they cover quick experiments, production applications and large document pipelines.

flowchart TD
APP["Your Application or Website"] --> API["Cloud Translation API"]
API --> BAS["Basic Edition (v2)"]
API --> ADV["Advanced Edition (v3)"]
BAS --> NMT["NMT Model"]
ADV --> NMT
ADV --> TLLM["Translation LLM"]
ADV --> ADP["Adaptive Translation"]
ADV --> CUS["Custom NMT Models"]
ADV --> GLO["Glossaries"]
ADV --> DOC["Document and Batch Translation"]
STUDIO["Vertex AI Studio"] --> TLLM
STUDIO --> CUS
        

The first decision you make is which edition to use.

Basic is optimized for simplicity and scale. It gives access to the standard pre-trained NMT model, accepts simple API keys as well as service accounts, and needs almost no configuration. Advanced is optimized for customization and long-form content. It adds a choice of models, glossaries, batch requests, document translation, custom models, regional endpoints and tighter integration with identity and access management. Basic remains available, but new capabilities arrive in Advanced, so new projects usually start there.

Different jobs call for different engines under the hood.

flowchart TD
Q1{"Is general quality enough, with speed and low cost the priority?"} -->|Yes| N["Standard NMT"]
Q1 -->|No| Q2{"Need better context, tone or longer passages?"}
Q2 -->|Yes| Q3{"Have a few good example pairs for your style?"}
Q3 -->|Yes| AD["Adaptive Translation"]
Q3 -->|No| TL["Translation LLM"]
Q2 -->|No| Q4{"Have thousands of in-domain sentence pairs?"}
Q4 -->|Yes| CU["Custom NMT or fine-tuned model"]
Q4 -->|No| N
        

A practical way to decide

Start with the standard model and measure. If terminology is the main problem, add a glossary first, because it is the cheapest fix. If style and tone matter, try adaptive translation with a handful of good examples. Only if you have a large body of high quality translated data and a clear need should you invest in training or fine-tuning a custom model.

A glossary is the simplest and most powerful way to keep your key terms consistent.

Every business has words that must be handled with care. A product name must stay in its original form. A technical term must always translate to the same word. A legal phrase has an approved version. Models do not know these preferences automatically. A glossary is a custom dictionary that you upload. When it is attached to a translation request, the service uses your entries for those terms and translates everything else normally.

Translate the whole file and keep it looking like the original.

Translating text strings is easy to picture, but companies often have files: manuals, contracts, slide decks, spreadsheets. Copying text out, translating it and pasting it back loses formatting and wastes hours. Document Translation sends the whole file to the service, which translates the content and keeps the layout, so paragraphs, headings and structure appear as in the original. Supported formats include PDF, Word, PowerPoint and Excel files. For native PDFs the service can preserve font size, color, style and links, and it supports right to left languages.

flowchart LR
IN["Source document: PDF, DOCX, PPTX, XLSX"] --> SVC["Document Translation"]
GL["Glossary (optional)"] --> SVC
MD["Chosen model (optional)"] --> SVC
SVC --> OUT["Translated document with layout kept"]
        

Things to watch

Scanned PDFs that contain pictures of text, rather than real text, may need text recognition first. Text that gets longer in the target language, such as German, can overflow narrow boxes in slides or forms. Pages are counted for billing, with limits on how much text each page can hold on average. Always inspect a sample of translated files before processing a whole library.

When you have thousands of files, you send a job and come back for the results.

Real-time translation answers in a moment, but it is meant for small pieces such as a chat message or a web page paragraph. For large amounts, such as a whole product catalog or an archive of manuals, you use batch translation. You place the input files in a Cloud Storage bucket, send a batch request naming the languages and optional glossary, and the service runs it as a long-running operation. It writes the translated files to an output bucket that you specify. Inline batch requests are not supported, so the input always comes from Cloud Storage.

The small details that decide whether a translation request goes smoothly.

Detecting the language

If you do not know what language a piece of text is in, perhaps because it comes from customers around the world, the service can detect it. You can ask for detection on its own, or simply leave out the source language in a translation request and let the service work it out. Detection is more reliable on longer text than on a single word. Very short or mixed-language text can confuse it, so supply the source language whenever you know it.

Plain text and markup

Requests can carry plain text or HTML. When text contains markup, such as bold tags or links, it is important to tell the service so that it translates the words and leaves the structure alone. Translating HTML as plain text can damage tags, while translating plain text as HTML can mangle special characters. Choose the correct input type for each case.

Multiple strings and limits

You can send several strings in one request, which is more efficient than many tiny calls. There are limits on request size, number of strings and requests per minute. Large jobs should use batch translation, and busy applications should handle rate limit errors with a short wait and retry.

Follow one sentence from your application to the translated answer.

sequenceDiagram
participant A as Your Application
participant S as Translation Service
participant M as Chosen Model
participant G as Glossary
A->>S: Send text, target language, options
S->>S: Check identity, quota and request size
S->>S: Detect language if source not given
S->>G: Look up glossary terms if attached
S->>M: Translate with context
M-->>S: Translated text
S-->>A: Return translation and metadata
        

How good is good enough, and how do you know?

Translation quality is not a single number. A casual product review needs to be understandable. A legal contract needs to be exact. A marketing slogan needs to feel natural and persuasive in the target culture. The right level of quality, and the right amount of human involvement, depends on the risk of getting it wrong.

Ways to measure

Teams combine several approaches. Automatic metrics compare machine output with human reference translations and give a score, which is handy for tracking trends and comparing models, but they do not capture everything. Human reviewers rate fluency, accuracy and tone on a sample. Side-by-side comparison shows which of two models or settings people prefer. Business measures, such as customer satisfaction and support tickets, show whether translations actually work.

Translation becomes most useful when it is a step inside a larger system.

flowchart LR
CMS["New content published"] --> PS["Pub/Sub event"]
PS --> RUN["Cloud Run service"]
RUN --> TR["Translation API with glossary"]
TR --> GCS["Cloud Storage: translated content"]
GCS --> REV["Human review for key languages"]
REV --> PUB["Publish to each regional site"]
        

Translated text is often sensitive, so treat the service as part of your data protection plan.

Access control

Advanced integrates with identity and access management. Roles decide who can translate text, use documents, manage glossaries or train models. Applications authenticate through service accounts, and each application should have its own account with only the permissions it needs. Avoid sharing simple keys among many applications, because they are easy to leak and hard to trace.

Where your data goes

Advanced offers regional endpoints, so that requests and resources stay within a chosen location, such as the European Union. Multi-regional endpoints are also available for some features. For organizations with strict rules, VPC Service Controls can add a boundary that limits how data can leave a protected perimeter. Check data usage terms and the current documentation on how customer content is handled, and match them to your legal obligations.

Sensitive content

Before sending personal or confidential text for translation, decide whether it is necessary. Where possible, remove or mask personal details, such as names and account numbers, before translation, and restore them afterward. Keep glossaries and training data free of secrets, because they are stored resources.

A translation feature that fails silently is worse than one that was never built.

Treat the translation call like any other dependency of your application. Watch how many requests you send, how many fail, how long they take and how many characters you consume. Quotas limit characters per minute and requests per minute, and exceeding them returns errors. Cloud Monitoring and logging show these signals, and alerts warn you early.

Designing for failure

Use retries with increasing waits for temporary errors. Cache translations of content that does not change, such as product descriptions, so you do not pay for or wait on the same translation twice. If the service is unavailable, show the original text, perhaps with a note, rather than breaking the page.

You pay by the amount of text you send, so volume and repetition drive the bill.

Standard text translation is billed by the number of characters sent to the service, per target language. Translating one text into five languages counts five times. At the time of writing, the standard model has a monthly free allowance, applied as a credit across both editions, and a published rate per million characters above it. Document translation is billed per page. Custom models, adaptive translation and the Translation LLM follow their own price tables, and training a custom model is billed by the hour. Prices change, so use the official pricing page and calculator for current numbers.

Where automatic translation quietly opens doors.

These are common patterns rather than named customer stories, because specific designs and contracts change over time.

Speed and scale are bought with the need for care about quality and context.

Habits that make multilingual products shine, and habits that embarrass them.

Patterns that work

Strong teams translate content at the right moment: user-generated text on demand, product and help content ahead of time through batch jobs. They keep a glossary as a living asset with an owner. They cache translations and track which glossary and model version produced each one. They route content by risk, sending low-risk text straight out and high-risk text through human review. They separate text from code and design interfaces so strings can be replaced and longer text fits.

A checklist to keep your multilingual features accurate, safe and affordable.

Mistakes beginners make most often

The first is trusting machine output for high-stakes text without review. The second is skipping glossaries. The third is retranslating unchanged content. The fourth is testing only in English or in one language pair. The fifth is ignoring the layout effects of longer text. The sixth is exposing API keys in client-side code. The seventh is launching in many languages without a plan to handle feedback and corrections.

There is more than one way to get text into another language.

flowchart TD
Q1{"Is a wrong word dangerous or very costly?"} -->|Yes| HU["Human translation or machine draft with expert review"]
Q1 -->|No| Q2{"Is brand voice or tone critical?"}
Q2 -->|Yes| PE["Machine translation with adaptive examples plus editing"]
Q2 -->|No| Q3{"Large volume of ordinary content?"}
Q3 -->|Yes| MT["Cloud Translation with glossary and caching"]
Q3 -->|No| QK["Basic or standard NMT for quick needs"]
        

The best systems combine these. Machines handle the bulk, glossaries keep terms consistent, and people focus on the content where their judgment adds the most value.

It is the family of Google Cloud services, centered on the Cloud Translation API, that translates text and documents between more than a hundred languages using neural models and language models, with options to customize terminology and style.

Basic gives simple access to the standard NMT model. Advanced adds the Translation LLM, adaptive translation, custom models, glossaries, batch and document translation, regional endpoints and stronger access controls.

Create a glossary with those names and attach it to your translation requests. The service will then use your entries for those terms, and you can review output to confirm.

Yes. Document translation handles formats such as PDF, Word, PowerPoint and Excel and keeps layout as far as possible. Always check a sample of the results for layout issues.

Use batch translation. Put the files in Cloud Storage, send a batch request, and the service writes the translated files to an output bucket as a background operation.

It uses a language model guided by a few example sentence pairs of your own, so the output follows your preferred wording and style, without training a whole custom model.

Usually not at first. Start with the standard model plus a glossary, try adaptive translation next, and consider a custom model only if you have plenty of clean in-domain data and a clear gap to close.

Yes. You can request detection or leave out the source language. It is more reliable for longer text, so give the source language when you know it.

Text is billed per character, per target language, documents per page, and custom training by the hour, with different tables for different models. A monthly free allowance exists for the standard model. Check the official pricing page for current rates.

It is a useful fast first pass, but important legal, medical or safety text should be reviewed by qualified people before it is relied upon.

Advanced supports regional endpoints and, for stricter needs, perimeter controls with VPC Service Controls. Review the current documentation to confirm what applies to the features you use.

Translate only the languages you need, cache results, translate only content that changed, use the simplest model that meets your quality needs, and set budgets and alerts.

Translation AI in Google Cloud turns language into a service. The Cloud Translation API, in its Basic and Advanced editions, gives access to a standard neural model, a Translation LLM, adaptive translation and custom models, with glossaries, document translation and batch processing for business needs. Success comes from choosing the right model for each kind of content, protecting your terminology, caching what does not change, securing access and data location, and keeping humans in the loop where accuracy matters most.