
Cargando...
Neural machine translation that makes your applications speak every language — instantly and at scale.
Amazon Translate is a neural machine translation (NMT) service that delivers fast, high-quality, and affordable language translation. It supports real-time and batch translation across a wide range of language pairs, enabling developers to build multilingual applications without managing translation infrastructure. Translate integrates natively with the broader AWS AI/ML ecosystem, making it a key component in pipelines involving Comprehend, Transcribe, Textract, and Bedrock.
Automate high-quality text translation across dozens of languages for real-time and batch workloads in applications, content pipelines, and AI workflows.
Use When
Avoid When
Real-Time (Synchronous) Translation
Single API call, up to 5,000 bytes per request
Batch Translation (Asynchronous)
S3 input → S3 output; handles large-scale document jobs
Custom Terminology
CSV or TMX files to enforce specific term translations (brand names, product names)
Parallel Data (Active Custom Translation)
Train translation style with example sentence pairs for domain adaptation
Automatic Language Detection
Can auto-detect source language using Amazon Comprehend under the hood; no separate call needed
Formality Setting
Control formal vs. informal register for supported language pairs
Profanity Masking
Mask profane words in translated output with a grawlix (****)
Brevity Setting
Request shorter translations when length matters (e.g., UI labels, SMS)
Do Not Translate Tags
HTML/XML tags can mark content to be preserved as-is (e.g., brand names in markup)
Encryption at Rest
Uses AWS KMS; customer-managed keys (CMK) supported
Encryption in Transit
TLS enforced for all API calls
VPC Endpoints (PrivateLink)
Keep translation traffic off the public internet
CloudWatch Metrics
Monitor request counts, latency, and errors
AWS CloudTrail Integration
API call logging for compliance and audit
Speech-to-Translated-Text Pipeline
high freqTranscribe converts audio to text in the source language; Translate converts the transcript to the target language. Used for multilingual call center analytics, meeting transcription, and subtitle generation.
Multilingual Sentiment & Entity Analysis
high freqTranslate normalizes non-English text to English first, then Comprehend performs sentiment analysis, entity recognition, or key phrase extraction. Comprehend's primary NLP models are optimized for English — Translate bridges the gap.
Multilingual Document Intelligence
high freqTextract extracts text from scanned documents (PDFs, images); Translate converts extracted text to a target language. Enables cross-language document processing workflows for legal, financial, and compliance use cases.
Multilingual Chatbot
high freqLex handles conversational logic in a single language (e.g., English); Translate converts user input to English before Lex processing and translates Lex responses back to the user's language. Avoids building separate Lex bots per language.
Multilingual Generative AI Applications
high freqTranslate normalizes multilingual user prompts to English before sending to a foundation model in Bedrock, then translates the model's English response back to the user's language. Maximizes FM effectiveness while supporting global users.
Text-to-Translated-Speech (TTS Pipeline)
medium freqTranslate converts text to target language; Polly synthesizes speech in that language. Builds multilingual voice assistants, IVR systems, and accessibility features.
Multilingual Image/Video Metadata Pipeline
medium freqRekognition detects labels, text, or moderation signals in images/video; Translate localizes the resulting labels or detected text for international audiences.
Event-Driven Batch Translation
medium freqS3 PUT event triggers Lambda to start a Translate Batch Translation job; results written back to S3. Serverless, cost-efficient pipeline for processing uploaded documents at scale.
Real-Time Content Localization Cache
medium freqLambda translates content on first request and caches translated strings in DynamoDB; subsequent requests for the same content/language pair are served from cache, reducing translation costs and latency.
When a question asks how to make a single-language NLP service (like Comprehend) work with multilingual input, the answer is almost always: Translate the input to English FIRST, then pass it to Comprehend. This is the canonical AWS multilingual NLP architecture.
Real-time translation limit is 5,000 BYTES (not characters). For large documents or bulk processing, Batch Translation (S3 → Translate → S3) is the correct architectural choice. If a question mentions 'thousands of documents' or 'large files', Batch Translation is the answer.
Custom Terminology enforces SPECIFIC WORD MAPPINGS (e.g., brand names, product names, technical terms) — it does NOT change the translation style or domain. Parallel Data (Active Custom Translation) is used to adapt translation STYLE and DOMAIN. Exams test this distinction directly.
For AIF-C01: In a generative AI pipeline serving global users, Translate is the correct service to normalize multilingual prompts before sending to a foundation model via Bedrock. This is a common architecture question in the AI Practitioner exam.
Multilingual NLP = Translate FIRST (normalize to English), THEN Comprehend (analyze). Translate does not perform NLP — it only moves text between languages.
Real-time limit = 5,000 bytes. Large documents or bulk processing = Batch Translation (S3-based). This distinction drives architecture decisions on SAA-C03 and SAP-C02.
Custom Terminology = enforce specific word substitutions (brand names). Parallel Data = adapt translation style/domain with example sentences. These are NOT interchangeable.
Amazon Translate does NOT perform sentiment analysis, entity recognition, language detection as standalone features for downstream use — those are Amazon Comprehend's jobs. Translate's auto-detect feature uses Comprehend internally, but you cannot use Translate as a replacement for Comprehend in NLP pipelines.
For multilingual chatbots using Amazon Lex: the recommended pattern is Translate user input → Lex (English) → Translate Lex response back. Do NOT build separate Lex bots per language — that's operationally expensive and not the AWS-recommended architecture.
Translate is NOT a content moderation service. It does not detect harmful content, prompt injection, or policy violations. For AI safety in multilingual pipelines, you still need Comprehend (toxicity/sentiment), Bedrock Guardrails, or AWS WAF — but WAF cannot perform semantic analysis. Don't confuse translation with moderation.
Pricing is per CHARACTER translated. When designing cost-optimized architectures, implement a translation cache (e.g., DynamoDB or ElastiCache) to avoid re-translating identical strings. This is a classic SAA-C03/SAP-C02 cost-optimization pattern.
The Formality, Profanity Masking, and Brevity settings are OPTIONAL parameters on the translation API — not separate services. Know that these exist for exam questions about customizing translation output quality.
Common Mistake
Amazon Translate can detect sentiment or intent in translated text — I can use it as an all-in-one NLP service.
Correct
Amazon Translate ONLY translates text between languages. Sentiment analysis, entity recognition, key phrase extraction, and language detection for NLP purposes are all Amazon Comprehend features. The two services are complementary, not interchangeable.
This is one of the most common service-boundary confusion traps on SAA-C03 and AIF-C01. The correct architecture is: Translate (normalize language) → Comprehend (analyze text). Remember: Translate moves words between languages; Comprehend understands what those words mean.
Common Mistake
I can use Translate's auto-detect feature to identify the source language and then use that result in my application logic — it's a language detection service.
Correct
While Translate can auto-detect the source language (using Comprehend internally), if you need reliable language detection as a standalone capability for routing, classification, or analytics, use Amazon Comprehend's DetectDominantLanguage API directly. Translate's auto-detect is a convenience feature for translation calls, not a primary language detection service.
Exam questions about 'detect the language of incoming customer messages' point to Comprehend, not Translate. Using Translate just to detect language (without actually translating) is wasteful and architecturally incorrect.
Common Mistake
Custom Terminology and Parallel Data do the same thing — they both customize how Translate handles specific content.
Correct
Custom Terminology is a lookup-table approach: it enforces specific word-for-word substitutions (e.g., always translate 'AWS' as 'AWS', never localize brand names). Parallel Data (Active Custom Translation) trains the model on example sentence pairs to adapt the STYLE, TONE, and DOMAIN of translations (e.g., making translations sound more formal or more like your industry's terminology).
Exam questions will describe a scenario and ask which customization feature to use. Key signal: if the requirement is 'never translate this specific term' → Custom Terminology. If the requirement is 'make translations sound like our industry documents' → Parallel Data.
Common Mistake
Batch Translation is more expensive than real-time translation because it's a managed job service.
Correct
Batch Translation uses the same per-character pricing as real-time translation. There is no premium for using the asynchronous batch job capability. The choice between real-time and batch is driven by use case (latency vs. volume) and document size limits, not cost.
Cost-optimization questions may tempt you to choose real-time translation as 'cheaper' — it is not. Choose Batch Translation for large documents (>5,000 bytes) and high-volume processing regardless of cost concerns.
Common Mistake
Amazon Translate can moderate translated content and flag harmful or inappropriate translations before they reach users.
Correct
Amazon Translate has a Profanity Masking feature, but this is limited to masking profane words with grawlix characters (****). It is NOT a content moderation service. For comprehensive content moderation, use Amazon Comprehend (toxicity detection), Amazon Rekognition (image/video moderation), or AWS Bedrock Guardrails. Translate processes language — it does not evaluate content safety.
This is directly relevant to AIF-C01's AI safety domain. In multilingual AI pipelines, you need BOTH Translate (for language normalization) AND a separate moderation/safety service. Translate alone cannot protect against harmful content, prompt injection, or policy violations.
Common Mistake
To build a multilingual chatbot with Amazon Lex, I need to create a separate Lex bot for each language I want to support.
Correct
The AWS-recommended pattern is to build ONE Lex bot in a primary language (typically English) and use Amazon Translate to translate user inputs INTO that language before sending to Lex, then translate Lex's responses BACK to the user's language. This dramatically reduces operational complexity.
Exam questions about 'cost-effective multilingual chatbot' or 'scale chatbot to 10 new languages quickly' point to the Translate+Lex pattern, not creating N separate bots. Creating separate bots per language is expensive, hard to maintain, and not the AWS-recommended architecture.
TRANSLATE = Text Relay And Navigation for Speaking Languages Across The Enterprise. It MOVES text between languages — it doesn't READ meaning (that's Comprehend's job).
Custom Terminology = DICTIONARY (specific word → specific word). Parallel Data = STYLE GUIDE (here are example sentences in our style — learn from them).
The 5,000-byte real-time limit: Think '5K bytes = 5 kilobytes = roughly one page of text'. More than one page? Use Batch (S3).
Translate pipeline order: TRANSCRIBE (ears) → TRANSLATE (tongue) → COMPREHEND (brain) → POLLY (voice). Each service does ONE job.
CertAI Tutor · AIF-C01, SAA-C03, SAP-C02, CLF-C02 · 2026-02-22
In the Same Category
Comparisons
Guides & Patterns