Azure AI Services Cheat Sheet

Azure AI services hierarchy

Vision services

Service	What it does	Key features
Azure AI Vision	Analyze images and video	Image Analysis, OCR, spatial analysis, image captioning
Face API (part of AI Vision)	Detect and recognize faces	Face detection, verification, identification, emotion (limited)
Custom Vision	Train custom image models	Image classification, object detection with your own data

When to use which:

• Need general image analysis? → Azure AI Vision
• Need to train on your own images? → Custom Vision
• Need to detect/verify faces? → Face API

Language services

Service	What it does	Key features
Azure AI Language	Understand and analyze text	Sentiment analysis, key phrase extraction, entity recognition, CLU, summarization, QnA
CLU (Conversational Language Understanding)	Build intent + entity models	Replaces LUIS; custom NLU for chatbots and apps
Azure AI Translator	Translate text between languages	100+ languages, custom models, document translation
QnA (Question Answering)	Build knowledge bases from FAQs	Part of Azure AI Language; replaces QnA Maker

Speech services

Service	What it does	Key features
Speech-to-Text	Convert audio to text	Real-time and batch transcription, custom models
Text-to-Speech	Convert text to audio	Neural voices, custom voice, SSML support
Speech Translation	Translate spoken language in real time	Combines recognition + translation
Speaker Recognition	Identify/verify who is speaking	Voice profiles, verification, identification

Document Intelligence

Service	What it does	Key features
Azure AI Document Intelligence	Extract structured data from documents	Pre-built models (invoices, receipts, ID), custom models, layout analysis

⚠️ Formerly called "Form Recognizer" — exam may reference either name.

Azure OpenAI Service

Feature	Details
GPT models	Text generation, chat completion, summarization, code generation
DALL-E	Image generation from text prompts
Embeddings	Vector representations for semantic search and RAG
Whisper	Audio transcription model
Access	Requires application/approval; deployed via Azure OpenAI resource

Azure Machine Learning

Component	Purpose
Workspace	Central hub for all ML assets (data, models, compute, experiments)
Compute	Training clusters, compute instances, inference endpoints
Designer	Drag-and-drop ML pipeline builder (no code)
AutoML	Automatically trains and tunes models from your data
Pipelines	Orchestrate multi-step ML workflows
Model Registry	Version and manage trained models
Endpoints	Deploy models as REST APIs (real-time or batch)

Responsible AI — 6 principles

Principle	Definition	Example
Fairness	AI treats all people equitably	Loan model doesn't discriminate by gender
Reliability & Safety	AI performs reliably under expected conditions	Self-driving car handles edge cases safely
Privacy & Security	AI respects privacy and is secure	Training data is anonymized and encrypted
Inclusiveness	AI empowers everyone	App works for users with disabilities
Transparency	AI systems are understandable	Users know when they're interacting with AI
Accountability	People are accountable for AI systems	Human review process for high-stakes decisions

Key differences table

Old name	New name	Notes
Cognitive Services	Azure AI services	Umbrella rebrand (2023)
Form Recognizer	Azure AI Document Intelligence	Same capabilities, new name
LUIS	CLU (Azure AI Language)	Conversational Language Understanding
QnA Maker	Question Answering (Azure AI Language)	Now part of Language service
Metrics Advisor	Azure AI Anomaly Detector	Consolidated

Comparison	When to use A	When to use B
Custom Vision vs Azure AI Vision	You need to train on your own labeled images	You need general-purpose image analysis (out-of-the-box)
CLU vs Prebuilt Language features	You need custom intents/entities for your domain	Standard tasks (sentiment, key phrases, NER) suffice
Azure ML vs Azure AI services	You need to build/train/deploy custom models	You need a pre-built API for a common AI task
Azure OpenAI vs Azure AI Language	You need generative text, chat, or code	You need analytical NLP (sentiment, extraction)

ML concepts quick reference

Term	Meaning
Feature	An input variable used for prediction (e.g., age, income)
Label	The value you're trying to predict (e.g., price, category)
Training	Feeding data to an algorithm so it learns patterns
Validation	Testing the model on held-out data to measure accuracy
Inference	Using a trained model to make predictions on new data
Overfitting	Model memorizes training data; performs poorly on new data
Underfitting	Model is too simple; misses patterns in data