Episode 05 · Identify AI concepts
Four text techniques the skills list names on purpose, plus the two arrows of speech. Mira uses hotel reviews; Owen hunts the lookalike answers.
Transcript
Mira Cole
Episode five. The skills list is unusually specific here. Common text analysis techniques: keyword extraction, entity detection, sentiment analysis, and summarization. Then features of speech recognition and speech synthesis. We will treat those as six labeled drawers.
Owen Park
Keyword or key-phrase extraction pulls the important terms. 'Wait time,' 'checkout,' 'parking.' It does not tell you if the guest was happy. It tells you what the text is about.
Mira Cole
Entity detection finds named things: people, places, organizations, dates, quantities, product names. 'June fourteenth,' 'Contoso Suites,' 'three nights.' Keywords are topics. Entities are the proper nouns and typed values living in the sentence.
Owen Park
A picture: keywords are the index at the back of a book. Entities are the cast list. If the question says 'who or what is named,' it is entities. If it says 'what themes keep appearing,' it is key phrases.
Mira Cole
Sentiment analysis judges attitude: positive, negative, mixed, neutral, sometimes at sentence level. It does not extract the date of the stay. It scores the vibe.
Owen Park
Summarization compresses. Extractive summary lifts sentences. Abstractive summary rewrites. For the exam, 'give me the short version' is summarization, even though a generative model is usually the engine.
Mira Cole
You can chain them. Extract entities, score sentiment, summarize for a manager. The exam still wants the name of the technique that matches the stated goal.
Owen Park
Speech recognition is transcription. A call center records audio, you get a transcript. Speaker diarization, timestamps, punctuation, languages, and custom vocabularies are the features you might be asked to recognize as capabilities, not to configure in a portal in this objective.
Mira Cole
Speech synthesis is the opposite. Text goes in, audio comes out. Voices, languages, speed, neural voices that sound less robotic. Accessibility, hands-free apps, a bot that speaks.
Owen Park
Translation can sit on top: recognize in one language, translate, synthesize in another. If the story is 'the kiosk talks back in the visitor's language,' synthesis is in the mix. If the story is 'we need a searchable transcript,' recognition is the star.
Mira Cole
Later, in Foundry, a multimodal model can take spoken audio straight in a prompt. That does not retire Azure Speech in Foundry Tools. Speech is still the dedicated service when you want a transcript, a voice, or real-time audio plumbing.
Mira Cole
Question one. You want every product name and date mentioned in warranty emails. Technique?
Owen Park
Entity detection. Product names and dates are typed entities. Key-phrase extraction might return 'warranty claim' as a topic and miss the S.K.U.
Mira Cole
Question two. Leadership wants a weekly digest of five-thousand tickets, half a page. Technique?
Owen Park
Summarization. Not sentiment, unless they asked for a mood chart. They asked for shorter.
Mira Cole
Question three. A support dashboard should glow red when customers are angry. Technique?
Owen Park
Sentiment analysis. Keywords might show 'refund' without telling you if the refund was praised or demanded.
Mira Cole
Question four. A museum app should read exhibit text aloud for visitors who cannot look at the screen. Capability?
Owen Park
Speech synthesis. Text to speech. Recognition would be the visitor talking to the app.
Mira Cole
Question five. Agents need a searchable archive of yesterday's calls. Capability?
Owen Park
Speech recognition, then you can run text analysis on the transcript. The unique speech step is getting text out of audio.
Mira Cole
Episode six finishes the concept floor with computer vision, image generation, and the idea of extracting from images, audio, and video. Then we move into Foundry for the rest of the season.
Owen Park
Bring your eyes. And maybe an invoice photo.