Exam-style
A contact centre records every call. The recordings need to become text. The transcripts then need checking for personal data and scoring for sentiment. A summary of each call has to be readable by a Spanish-speaking team in Madrid. And a voice bot on the front door should handle the three most common requests, such as checking a delivery date, before anyone reaches a human. Which AWS managed AI services do those four jobs, in that order?
Reveal the answer
C. Amazon Transcribe, Amazon Comprehend, Amazon Translate, Amazon Lex
Take the jobs one at a time and each has a purpose-built service with the job in its description. Audio into text is Amazon Transcribe. Understanding text that already exists, which covers both the personal-data check and the sentiment score, is Amazon Comprehend. Moving text between languages is Amazon Translate. A conversational front door built from defined intents and slots, where “when is my delivery” is an intent and the order number is a slot, is Amazon Lex. Amazon Polly appears in the finished pipeline too, speaking the bot’s replies down the phone line, but it is not one of the four jobs asked about. The first distractor swaps the two speech services and the two analysis services, which is the swap worth rehearsing. Amazon Polly turns text into speech and Amazon Transcribe turns speech into text, so the direction of the conversion decides between them, and here the conversion runs from a recording towards a transcript. Amazon Rekognition analyses images and video, not text. The second distractor is closer but puts Amazon Quick, the enterprise assistant, on the front door. That is an open-ended assistant answering questions from an indexed corpus of company documents, a different architecture from a bot with three defined intents wired to the delivery-tracking system. A foundation model on Amazon Bedrock genuinely could do the middle three, and it would cost more per call, vary its output between runs, and still need something else to handle the audio. Amazon SageMaker AI is where you build and host a model of your own when no managed service covers the task, which means data, labelling, training and an endpoint to operate for four jobs that already have a managed service.
Q. Calls to text, transcripts checked for personal data and sentiment, summaries into Spanish, and a voice bot handling three common requests. Which services, in order?
A. Amazon Transcribe, Amazon Comprehend, Amazon Translate, Amazon Lex, with Amazon Polly speaking the bot’s replies.
Why? The direction of the conversion separates the pair that gets swapped most: audio to text is Transcribe, text to audio is Polly. Comprehend reads text and reports what is in it (entities, personal data, sentiment) rather than writing anything new. Lex handles a fixed set of intents and slots, which is a narrower job than an assistant answering open questions from a document store. A foundation model on Amazon Bedrock could cover three of the four, at a higher price per call and with less consistency, and training your own in Amazon SageMaker AI turns four API calls into four build projects.