Day 3 - AWS AI Services (1): Managed APIs Handling Images, Documents, Text, and Speech

📌 Key Takeaways

• AWS AI services hide pre-trained models behind the scenes and expose only finished capabilities as APIs. You do not need to build or choose a model.
• Purpose mapping: Rekognition for images and video, Textract for extracting tables and forms from documents, Comprehend for analyzing the meaning of text, Transcribe for speech→text, Polly for text→speech, and Translate for translation.
• Tell the confusing pairs apart by input and purpose. Objects and faces in a photo mean Rekognition, and pulling fields from paperwork means Textract. Analyzing text that already exists means Comprehend.
• Services are used connected in pipelines. Call→text→sentiment (Transcribe→Comprehend), speech→translation→speech (Transcribe→Translate→Polly).
• Where a general model is weak, such as on specialized terms or classifications, supplement it with each service's custom features, and if that is still not enough, move on to SageMaker.

The 6 Services at a Glance

The Bedrock and SageMaker we have seen so far were services that deal with "the model itself." The services we look at today are of a different kind. They are managed AI services that finish a specific task with a single API call, without even needing to build or choose a model. AWS calls these "AI Services."

• They hide models that AWS has pre-trained behind the scenes and expose only finished capabilities such as "find text in an image" and "speech to text."
• The AIF-C01 exam relentlessly asks "which service for this task?" The goal is to pair the purposes of the 6 core services precisely.

This table is the core of today. The trick for memorizing it is to look first at what the input is. Photos and video mean Rekognition, paperwork means Textract, writing means Comprehend or Translate, sound means Transcribe, and turning writing into sound means Polly.

Rekognition: Eyes That See Images and Video

Amazon Rekognition is a computer vision service that analyzes images and video.

• Object and scene detection: What is in the photo (a car, a puppy, a beach).
• Face analysis and comparison: Detecting faces, comparing whether two faces are the same person, and estimating expression and age range.
• Text extraction: Reading text inside images (signs, license plates).
• Inappropriate content detection (content moderation): Filtering violent and sexually explicit images.
• Celebrity recognition, labeling, and more.

Going a bit deeper, there are three more features likely to appear on the exam.

• Custom Labels: Trains the service on your own images so it can find things a general model does not know, such as your company's logo, products, or parts. It is the answer to "I want to find our company's products in images without ML expertise."
• Protective equipment (PPE) detection: Checks whether workers are wearing hard hats, masks, and so on.
• Face Liveness: During face authentication, tells whether a real person is in front of the camera or someone is trying to cheat with a photo or video.

As a hypothetical example, a social media platform had people inspect, one by one, millions of images that users uploaded for inappropriate content. After introducing first-pass automatic filtering with Rekognition's content moderation API, people only needed to review the ambiguous cases, and the review staff shrank considerably. When a clue says "you must understand images/video," it is Rekognition.

• Detection results come with a confidence score. A setup like the case above, where high-scoring items are handled automatically and humans look only at the ambiguous range, is the typical form of Human-in-the-loop.
• Face analysis is a feature that touches people's rights. From a responsible AI perspective, you must check its purpose and limits before using it, which connects to Week 5.

Textract: Truly Reading Documents

Amazon Textract goes beyond simple OCR. If ordinary OCR extracts "only the characters," Textract understands even the structure of tables and forms. It recognizes key-value pairs such as "Total: $1,200" in an invoice and extracts tables accurately by rows and columns.

💡 Related theory: Rekognition also reads text inside images, but Textract specializes in document processing. Its purpose is to pull data out of "structured documents" such as receipts, tax forms, insurance claims, and ID cards and process them automatically. On the exam, split them like this: "extract fields from a scanned invoice/form and enter them automatically" is Textract, and "analyze objects or faces in a photo" is Rekognition.

Even for the same "reading text," the targets differ. One line on a sign in a street photo is Rekognition, and which value is in which box on one sheet of paperwork is Textract.

Comprehend: Grasping the Meaning of Text

Amazon Comprehend is a natural language processing (NLP) service that analyzes the meaning of text.

• Sentiment analysis: Positive, negative, neutral, and mixed (a blend of the two).
• Entity recognition: Extracting people, places, dates, and organizations.
• Key phrases, language detection, topic modeling.
• PII detection: Identifying personal information in text.

There is also Comprehend Medical, specialized for medical text.

• Targeted sentiment: Looks at sentiment per entity rather than for the whole document. In "The food was great but the delivery was late," the food is positive and the delivery is negative.
• Comprehend Custom: With your own data, you build a custom classification that sorts documents into your own taxonomy and a custom entity recognition that finds terms unique to your industry. The service handles the model design for you.

🔍 Going deeper: Comprehend analyzes meaning "when the text already exists." It works like automatically tallying the sentiment of thousands of customer reviews, or pulling people and dates out of documents to index them. The clue "analyze the sentiment/entities/language of text" is the key signal.

Comprehend does not write new text. It only classifies and extracts. If you need to produce a summary or a reply, that is a job for generative AI (Bedrock).

Transcribe & Polly: Moving Between Speech and Text

These two go in opposite directions.

• Amazon Transcribe (STT, Speech-to-Text): Speech → text. Transcribing meeting recordings, logging call center calls, generating captions. It also supports speaker diarization (who spoke).
• Amazon Polly (TTS, Text-to-Speech): Text → speech. Reads text aloud in a natural human voice. Audiobooks, voice guidance, accessibility.

• Transcribe supports both batch processing of files uploaded to S3 and real-time streaming. Organizing recorded files is batch, and live-broadcast captions are streaming.
• If it mis-transcribes technical terms or proper nouns, supplement it with a custom vocabulary or a custom language model. It also has features for masking personal information in the transcript and filtering inappropriate words.
• There are variants tuned for call center call analysis and medical dictation (Call Analytics, Transcribe Medical).
• Polly offers several kinds of voices, such as standard, neural, long-form, and generative, and adjusts pronunciation, pauses, and speed with SSML tags.

As a hypothetical example, a call center used Transcribe to convert calls to text, then fed that text into Comprehend to analyze customer sentiment. It automatically flagged calls with strong complaint signals and notified managers. It is a typical example of connecting several AI services in a pipeline (speech→text→sentiment).

Translate: Crossing the Language Barrier

Amazon Translate is a neural network-based machine translation service. It translates text from one language to another in real time. It is used for multilingual websites, translating customer support messages, real-time chat interpretation, and so on.

Translate is also often combined with other services. For example, if you transcribe foreign-language speech with Transcribe → translate it with Translate → output it as speech again with Polly, you get a simple interpretation pipeline.

• Words that must not be translated, or must be rendered in a fixed expression, such as brand and product names, are pinned with a custom terminology.
• A combination that translates reviews arriving in several languages into one language and then analyzes them with Comprehend is also common (Translate→Comprehend).

What these 6 services have in common is that they "expose pre-trained models as APIs." Users can use them right away without ML knowledge. However, in very specialized domains (for example, specialized medical terms or a company's own product names), a general model may be weak, so some services offer custom model / user dictionary features. If you still need "to train it yourself," you must move on to SageMaker.

Telling Look-Alike Features Apart

A finished AI service or a foundation model? These days an FM can also read images and translate. Even so, the exam has you pick the finished service by the following criteria.

Customization and Cost

• All six services are pay-as-you-go, and there is no server to keep running in advance. The closest thing to an exception is the real-time endpoint of a custom model, which is billed for the time it is kept up.
• A pipeline adds a charge at every step. If you run one minute of a call through Transcribe→Translate→Comprehend, the charges of all three services apply.

Common misconceptions.

• "You have to prepare training data to use an AI service too" → The basic features are called right away with no data. Only the custom features need data.
• "Rekognition reads text, so it can also process documents" → It reads the characters, but table and key-value structure is Textract's job.
• "Sentiment analysis needs generative AI" → Positive/negative classification is enough with Comprehend, and the shape of the result is also consistent.
• "Polly transcribes calls" → The direction is reversed. Transcription is Transcribe.

Scenario → Answer Decision Table

Tomorrow we look at the remaining AI services (Lex, Kendra, Personalize, Forecast) and Amazon Q, then tie all the services together by the criterion of "which one to pick for which task."

📖 Glossary

• Computer vision: The AI field that recognizes objects, faces, text, and so on in images and video.
• OCR: Technology that reads characters in an image out as text. Textract adds an understanding of table and form structure to it.
• Key-value pair: An item name paired with its value, like "Total: 1,200." It is the basic structure of form documents.
• NLP (natural language processing): Technology that lets computers analyze and understand human writing.
• Entity recognition: Finding names such as people, places, dates, and organizations in writing.
• PII: Information that can identify an individual, such as name, address, and account number.
• STT / TTS: Turning speech into text (Speech-to-Text) and turning text into speech (Text-to-Speech).
• Speaker diarization: Separating who spoke which part of a recording.
• Content moderation: The work of finding and filtering inappropriate images, video, and text.
• Confidence score: A value showing how sure the service is about its own judgment. If it is low, hand off to human review.

📝 Practice Questions

Click a choice to reveal the answer and explanation.

Question 1

You want to detect objects and faces in user-uploaded photos and automatically filter inappropriate content. Which is the most appropriate service?

• AAmazon Comprehend
• BAmazon Rekognition
• CAmazon Polly
• DAmazon Translate

Question 2

You want to extract key-value pairs like "Total" and "Date" and table data from a scanned invoice image into structured form. Which is the most appropriate service?

• AAmazon Textract
• BAmazon Transcribe
• CAmazon Rekognition
• DAmazon Polly

Question 3

You want to automatically classify thousands of customer reviews as positive or negative and extract entities like people and places from text. Which service is appropriate?

• AAmazon Translate
• BAmazon Comprehend
• CAmazon Textract
• DAmazon Lex

Question 4

You want to convert call center recording to text and then analyze the text's customer sentiment. Which service combination is appropriate?

• APolly → Translate
• BTranscribe → Comprehend
• CRekognition → Textract
• DTranslate → Polly

Question 5

You want to convert your website's guidance text to natural-sounding human voice and play it. Which is the most appropriate service?

• AAmazon Transcribe
• BAmazon Polly
• CAmazon Comprehend
• DAmazon Rekognition