AWS Machine Learning Services

Comprehend, Kendra, Lex, Polly, Rekognition, SageMaker, Textract, Transcribe, Translate — add AI to your apps without being a data scientist

ML services appear in the “Design Cost-Optimised” and “Design Performant” domains. The exam does NOT test deep ML knowledge — it tests which managed AI service to pick for a given scenario. Learn the one-line purpose of each service and its key differentiator. Every section has a dedicated exam tip box.

You don’t need a PhD to add AI to your application on AWS. AWS has pre-built, managed AI services that handle everything from reading documents to understanding speech to building full ML models — you just call an API.

Think of these services in three groups: Language AI (understand, generate, and translate text and speech), Vision AI (analyse images and video), and Custom ML (SageMaker, for when you need to build your own models). The exam almost always tests which pre-built service fits a scenario — not how to build ML from scratch.

The SAA-C03 mindset for ML: The exam gives you a business problem and asks which AWS service solves it. You don’t need to know how the ML works internally — you just need to match the problem to the right service. The key question is always: Is this about text, audio, images, or custom ML?

Amazon Comprehend — Understand What Text Means

Natural Language Processing (NLP) — extract meaning, sentiment, entities, and sensitive data from text

Comprehend is AWS’s NLP service. It reads text and tells you what it means — is this customer review positive or negative? Who and what is mentioned in this document? Does this text contain a credit card number or a person’s name that should be redacted? All without you building any ML model yourself.

The analogy: Comprehend is like a very fast, tireless reader who can process millions of documents and answer: “What’s the overall tone? Who is mentioned? What topics are discussed? Is there any sensitive information?” It does in milliseconds what would take a human team weeks.

What Comprehend can detect

Comprehend Medical

A specialised version that understands clinical and medical text — extracting medical conditions, medications, dosages, test results, and protected health information (PHI) from clinical notes and medical records. Uses HIPAA-eligible service standards.

SAA-C03 Exam Tips —

Amazon Comprehend Comprehend = NLP on text. Key use cases: sentiment analysis (product reviews, customer feedback), PII detection/redaction (GDPR compliance, data privacy), entity extraction (who/what/where in documents), topic modelling. Comprehend Medical = clinical text, PHI detection, medical terminology. Custom models: train your own classifier with labelled examples. Comprehend vs Textract: Comprehend = understands meaning of text. Textract = extracts text FROM documents (images/PDFs). Trigger phrase: “analyse customer reviews”, “detect PII in text”, “classify support tickets automatically”.

Amazon Kendra — Intelligent Search That Understands Questions

ML-powered enterprise search that returns specific answers, not just a list of document links

Traditional search returns a list of documents matching keywords. Kendra understands what you’re actually asking and returns the specific answer — from within those documents. Ask “What is the company’s parental leave policy?” and instead of 10 links to HR documents, Kendra returns: “Employees are entitled to 16 weeks of paid parental leave.”

The analogy: A keyword search engine is like a librarian who hands you a stack of books that might contain your answer. Kendra is like a librarian who has read every book, understood every page, and can answer your question directly — and point you to the exact paragraph it came from.

Kendra key features

• 40+ connectors — S3, SharePoint, Salesforce, ServiceNow, Google Drive, OneDrive, Confluence, Jira, Slack, databases, and more. Kendra ingests and indexes content from all of them
• Semantic search — understands natural language questions, synonyms, and context. Doesn’t require users to know the exact keyword in the document
• Specific answer extraction — returns the actual answer text highlighted within the source document, not just a document ranking
• Incremental learning — learns from user feedback (thumbs up/down on results) to improve relevance over time
• Access control — respects existing document-level permissions so users only see documents they’re authorised to access
• FAQ support — ingest structured FAQ lists for direct question-answer matching

Kendra vs standard search (OpenSearch)

SAA-C03 Exam Tips —

Amazon Kendra Kendra = intelligent ML-powered enterprise search. Returns specific answers, not just document links. 40+ built-in data source connectors. Trigger phrase: “employees search internal knowledge base”, “IT help desk search”, “find answers within documents”, “natural language enterprise search”. Kendra vs Comprehend: Kendra = search and retrieve answers. Comprehend = analyse and extract meaning from text. Kendra vs OpenSearch: Kendra = semantic ML-powered search, enterprise connectors, higher cost. OpenSearch = keyword search, log analytics, high volume, custom control.

Amazon Lex — Build Conversational Chatbots and Voice Interfaces

The same AI that powers Alexa — automatic speech recognition (ASR) + natural language understanding (NLU)

Lex is the service that lets you build conversational interfaces — chatbots in your app or website, and voice-based interfaces like Alexa skills. It understands what a user is trying to do (their intent) and what information they’ve provided (the slots/entities), then triggers the right Lambda function or business logic.

The analogy: Lex is like a smart receptionist. You tell it: “I want to book a flight.” Lex knows you want to book (intent = BookFlight), asks follow-up questions to fill in missing information (slots = departure city, arrival city, date), and once it has everything, triggers the actual booking system (Lambda). It works whether the user types or speaks.

How Lex works — key concepts

Lex integrations

• Amazon Connect — Lex powers voice bots in Amazon Connect contact centres. Automates common customer service calls (check order status, update address, report issues) before escalating to a human agent
• AWS Lambda — the fulfillment layer. Every intent that needs to do something real connects to Lambda
• Messaging platforms — deploy Lex bots directly to Slack, Facebook Messenger, Twilio SMS, and Kik with built-in channel integrations

SAA-C03 Exam Tips —

Amazon Lex Lex = conversational chatbot + voice interface builder. The same technology as Alexa. ASR (speech-to-text) + NLU (understands intent and extracts slots). Integrates with Amazon Connect for automated contact centre voice bots. Lambda handles the fulfillment logic. Trigger phrase: “build a chatbot”, “voice-enabled customer service”, “automate call centre responses”, “conversational interface”. Lex vs Polly: Lex = understands speech input (ASR + NLU). Polly = generates speech output (TTS). Often used together: Lex understands what the user says → Lambda processes → Polly speaks the response.

Amazon Polly — Turn Text into Lifelike Speech

Text-to-speech (TTS) — convert written text into natural-sounding audio in 60+ voices, 30+ languages

Polly converts text into realistic spoken audio. It’s not a robotic voice reading letters — it uses deep learning to produce natural human-like speech with correct pronunciation, rhythm, and intonation. You can use it to add voice to apps, create audio versions of content, or power the speech output for a Lex chatbot.

The analogy: Polly is like a professional voice actor in every language who never gets tired, never needs a break, and costs fractions of a cent per word. You hand them a script (text), they record it (audio), you play it in your app.

Polly key features

• SSML (Speech Synthesis Markup Language) — fine-tune pronunciation, add pauses, change speaking rate, add emphasis, control breath sounds. Give Polly directions like a director giving a voice actor: “pause here”, “emphasise this word”, “speak this acronym letter by letter”
• Custom lexicons — define how specific words are pronounced. “AWS” → “Amazon Web Services”. “SQL” → “sequel”. Domain-specific terminology like medical or legal terms
• Speech marks — returns metadata alongside the audio: timestamps for each word and sentence, SSML marks for syncing visuals with audio (e.g. karaoke-style text highlighting or lip-sync animation)
• Output formats — MP3, OGG Vorbis, PCM (for telephony)

SAA-C03 Exam Tips —

Get Amit Naik’s stories in your inbox

Join Medium for free to get updates from this writer.

Remember me for faster sign in

Amazon Polly Polly = text-to-speech (TTS). Opposite of Transcribe (which is speech-to-text). Key differentiator: Neural TTS (NTTS) = highest quality, human-like voices. SSML = control pronunciation and pacing. Custom lexicons = how to pronounce domain-specific terms. Trigger phrase: “convert articles to audio”, “add voice to an app”, “generate speech from text”, “audio version of a blog post”. Polly + Lex pattern = Lex understands the user’s speech → processes it → Polly speaks the response. Polly stores generated audio in S3 for caching (avoid re-generating the same text repeatedly — cost saving).

Amazon Transcribe — Convert Speech into Accurate Text

Automatic Speech Recognition (ASR) — transcribe audio and video files or live streams into text

Transcribe converts audio and video into text — what was said, by whom, and when. It’s the opposite of Polly. Where Polly turns text into speech, Transcribe turns speech into text. It handles different accents, multiple speakers, background noise, and specialised vocabulary.

The analogy: Transcribe is like a court reporter who never gets tired and can process 1,000 hours of recordings simultaneously. They listen, type, note who spoke when, and produce a clean transcript — including timestamps and punctuation.

Transcribe key features

• Batch transcription — submit audio/video files stored in S3 for transcription. Output saved to S3 as a JSON transcript. Supports MP3, MP4, WAV, FLAC, OGG, and more
• Streaming transcription — real-time transcription of live audio streams over WebSocket or HTTP/2. For live captions, real-time meeting transcripts, live customer service call monitoring
• Speaker diarisation — identifies and separates different speakers: “Speaker 1: …”, “Speaker 2: …”. Great for meeting recordings and interviews
• Custom vocabulary — improve accuracy for domain-specific terms (medical, legal, technical jargon) by providing a list of words and their pronunciations
• PII redaction — automatically redact sensitive information (SSN, credit card numbers, phone numbers) from transcripts. Compliance feature for call recordings
• Amazon Transcribe Medical — specialised model trained on medical and clinical speech — understands drugs, procedures, anatomy, and clinical abbreviations with much higher accuracy than standard Transcribe

SAA-C03 Exam Tips —

Amazon Transcribe Transcribe = speech-to-text (ASR). Opposite of Polly (text-to-speech). Batch = audio/video files in S3. Streaming = real-time live audio. Speaker diarisation = who said what. PII redaction = redact sensitive info from call transcripts. Transcribe Medical = clinical/medical speech. Custom vocabulary = improve accuracy for domain-specific terms. Trigger phrase: “transcribe call recordings”, “generate subtitles from video”, “convert meeting audio to text”, “live captions”. Common pattern: Call recording → S3 → Transcribe → Comprehend (sentiment on transcript) → insights dashboard.

Amazon Translate — Neural Machine Translation at Scale

Fast, high-quality text translation across 75+ languages — real-time or batch, with custom terminology

Translate converts text from one language to another using neural machine translation — the same technology used by the best consumer translation services. It goes beyond word-for-word substitution to understand context and produce fluent, natural translations.

The analogy: Old machine translation was like replacing every word in a sentence with its dictionary equivalent — often producing nonsense. Neural machine translation (Translate) reads the whole sentence, understands what it means, then re-expresses that meaning naturally in the target language, the way a human translator would.

Translate key features

• 75+ language pairs — translate between any combination of supported languages. Detect source language automatically using Comprehend
• Real-time translation — synchronous API for translating text on-the-fly in your application. E.g. translate chat messages between users speaking different languages
• Batch translation (Async) — submit large volumes of documents in S3 for translation. Results written back to S3. Use for translating product catalogues, knowledge bases, or content libraries
• Custom terminology — define how specific terms (brand names, product names, technical jargon) should always be translated or kept as-is. “iPhone” should never be translated — keep it as “iPhone” in every language
• Active Custom Translation — provide example sentence pairs to fine-tune the translation style for your domain

SAA-C03 Exam Tips — Amazon TranslateTranslate = neural machine translation across 75+ languages. Real-time (API) or batch (S3 input → S3 output). Custom terminology = keep brand names/product names consistent across translations. Trigger phrase: “translate product listings to multiple languages”, “multilingual customer support”, “localise app content”. Common pattern: Comprehend (detect language) → Translate (convert to target language) → Comprehend (analyse translated sentiment). Translate vs Comprehend: Translate = convert text between languages. Comprehend = understand meaning within text.

Amazon Rekognition — See and Understand Images and Video

Computer vision — detect objects, faces, text, scenes, and activities in images and video without ML expertise

Rekognition is AWS’s computer vision service. You give it an image or video and it tells you what’s in it — faces, objects, scenes, text, activities, and whether the content is appropriate. No ML model to build or train. Just call the API with an image from S3 or as bytes, and get back structured results.

The analogy: Rekognition is like having a team of specialists who look at every image: a face recognition expert, an object identification specialist, a content moderator, and an OCR reader — all working simultaneously and returning their findings in one API response in milliseconds.

What Rekognition can detect

SAA-C03 Exam Tips —

Amazon Rekognition Rekognition = computer vision for images and video. No ML model needed — just API calls. Key capabilities: object detection, facial analysis (emotions, attributes), face comparison/search, text-in-image OCR, content moderation, activity detection in video. Content moderation is a very common exam scenario: user-generated content → Rekognition → flag inappropriate content → human review queue (via SQS + Lambda). Face comparison = identify if two faces are the same person. Rekognition vs Textract: Rekognition = text in natural images (signs, photos). Textract = structured text extraction from documents, forms, tables (invoices, tax forms). Trigger phrase: “moderate user-uploaded photos”, “find a person in security footage”, “detect faces in images”, “identify objects in photos”.

Amazon Textract — Extract Structured Data from Documents

Beyond OCR — extract text, tables, forms, and signatures from scanned documents and PDFs with structure preserved

Textract is more than OCR (Optical Character Recognition). While OCR just reads characters, Textract understands document structure — it knows that a value in a table belongs to a specific column header, that a checkbox is checked, and that a field labelled “Date of Birth” contains a specific value. It returns the data with its structural context preserved.

The analogy: Basic OCR is like a scanner that types up every character it sees, left to right, top to bottom — losing all structure. Textract is like a human assistant who reads a tax form and creates a structured spreadsheet: “Field: Taxpayer Name → Value: John Smith. Field: Annual Income → Value: $85,000. Table: Deductions → Row 1: Mortgage Interest $12,000.” The structure is preserved, not just the text.

What Textract extracts

SAA-C03 Exam Tips

Amazon Textract Textract = structured data extraction from documents (forms, tables, PDFs, scans). Goes beyond OCR — preserves structure (key-value pairs, table relationships, checkbox state). Textract vs Rekognition: Textract = structured document processing (invoices, tax forms, contracts). Rekognition = text in natural scene images (signs, photos). Textract vs Comprehend: Textract = WHAT the document contains (structured extraction). Comprehend = WHAT the text MEANS (sentiment, entities, PII). Common pipeline: Document uploaded to S3 → Textract (extract structured data) → Lambda (process/validate) → DynamoDB or RDS (store). Trigger phrase: “process invoices automatically”, “extract data from scanned forms”, “automate document intake”, “read PDF contracts”.

Amazon SageMaker — Build, Train, and Deploy Custom ML Models

End-to-end managed ML platform — when the pre-built AI services don’t fit your unique problem

The pre-built AI services (Comprehend, Rekognition, Transcribe, etc.) are excellent for common problems. But what if you have a unique business problem — predicting customer churn, detecting manufacturing defects from your specific production line photos, forecasting demand for your particular product mix? That’s SageMaker. It gives you the complete infrastructure to build, train, and deploy your own ML models without managing servers.

The analogy: The pre-built AI services are like ready-made meals from a restaurant — fast, convenient, and good for most people. SageMaker is a fully equipped professional kitchen — you have every tool, appliance, and ingredient you need to cook anything you can imagine. More work, but unlimited possibilities for your specific recipe.

Key SageMaker components

SAA-C03 Exam Tips —

Amazon SageMaker SageMaker = end-to-end ML platform for custom models. Use when pre-built services (Comprehend, Rekognition, etc.) don’t fit your specific problem. Key components: Studio (IDE), Training (managed compute), Endpoints (deployed model API), Ground Truth (data labelling), Pipelines (MLOps). SageMaker Savings Plans = up to 64% discount on SageMaker instance usage (1 or 3 year commitment). Spot Instances for training = up to 90% cost savings (managed checkpointing handles interruptions). SageMaker Model Monitor = detect data drift and model quality degradation in production. Trigger phrase: “build a custom ML model”, “train on proprietary data”, “deploy ML endpoint”, “MLOps pipeline”. Pre-built AI services first — SageMaker only when you need a custom model.

All nine services — quick reference

The most confused pairs — side by side

Scenario keyword map

The SAA-C03 decision framework for ML: Start by asking — is this a general problem (sentiment analysis, face detection, speech-to-text, language translation)? If yes → use the pre-built AI service. Is this a unique business problem requiring your own data and a custom model? → SageMaker. Then narrow down by modality: Text understanding → Comprehend/Kendra/Textract. Audio → Transcribe (in) / Polly (out). Translation → Translate. Conversation → Lex. Images/Video → Rekognition. Custom model → SageMaker.