Implement text analysis solutions

This domain turns natural language into structured, machine-usable value, and two questions route every task in it. The first is the medium the language arrives in: text that already exists as characters, or audio that starts or ends as sound. That split is the line between the domain's two subtopics. The second question, asked inside either medium, is whose vocabulary defines the answer. A prebuilt task ships a published label set and a fixed response shape, so an Azure AI Language task or a speech-to-text mode hands back a known object you can consume without guessing. A Microsoft Foundry model steered by a JSON Schema, a machine-readable contract for the reply's fields, instead returns exactly the fields you declare. The trap this model helps you dodge is reaching for a general chat model when a prebuilt task already returns the object you need, or forcing a prebuilt task's fixed categories onto a taxonomy it never published.

Walk the two subtopics in order. Language model text analysis: extraction, sentiment, and translation owns text you already hold as characters. It covers the Azure AI Language tasks (named entity recognition (NER) and personally identifiable information (PII) detection, sentiment and opinion mining, key-phrase extraction, and summarization), Azure Translator for turning text into another language, and schema-steered Foundry extraction for the categories no published task covers. Speech solutions for agentic and analytics workloads owns anything that starts or ends as audio. It covers the three speech-to-text transcription modes (real-time, fast, and batch), synthesis and its Speech Synthesis Markup Language (SSML) control surface, speech translation, the managed Voice Live speech-to-speech API for voice agents, and audio-native analysis that keeps the speaker structure a flat transcript throws away. Reach for the text page when characters are all you have; reach for the speech page the moment audio enters or leaves the pipeline.

Across both subtopics the exam rewards restraint: let a prebuilt task with a fixed response shape carry the work before you fall back to a general model. A Language task returns typed spans with offsets and confidence and, for PII, a ready-made redacted copy; a speech-to-text mode returns a transcript keyed to how the audio arrived. Those contracts are cheaper to consume and easier to trace than free-form prompting. Climb to a schema-steered Foundry model, using strict structured outputs, only when your categories or fields are not in a published list. Climb to a managed collapse-the-stages surface, Voice Live for a voice agent or a Content Understanding audio analyzer for audio-native fields, only when wiring the separate services by hand would buy you nothing. And keep each medium honest: speech translation re-renders what a speaker said and never answers on their behalf.