AI-901 lesson 9: text and speech solutions

A text and speech solution is a small application that sends text or audio to a Foundry-hosted capability and reads back a result. There are three building blocks to tell apart: Azure Language for analysing text, a deployed multimodal model that can take audio straight into a prompt, and Azure Speech for converting between audio and text. The exam objective calls all three “lightweight” because the application only needs a client, credentials, an input and code that reads the output.

A text analysis application

Lesson 4 covered what keyword extraction, entity detection, sentiment analysis and summarization do. Here the question is how an application uses them. Azure Language in Foundry Tools is a service that exposes these features. Its pre-built features return structured results, such as a sentiment label with confidence scores, a list of key phrases, or entities with a category and position in the text.

The flow has four steps. A Foundry resource provides an endpoint and credentials. The application creates a client for that endpoint, sends one or more text documents and names the feature wanted. The service returns a structured response. The application then does something with it: store the sentiment on a ticket, show the entities in a sidebar, or route a message.

Microsoft’s current Language documentation separates “core” features from “legacy” ones: sentiment analysis, key phrase extraction and summarization are listed as legacy, meaning established and supported for existing use rather than the focus of new investment. The AI-901 outline still names them, so recognise them. Older material says Text Analytics or Azure AI Language.

A model can do these jobs through a prompt too, but the service wins when output must be consistent and structured.

Two ways to handle spoken input

The figure above shows the choice. Path A sends audio directly to a deployed model that accepts audio input. The model interprets the speech and answers in the same step, and some audio-capable models can also answer with generated speech. Path B splits the work: Azure Speech turns the audio into text (speech recognition), a text model produces a text reply, and Azure Speech turns that reply back into audio (speech synthesis).

Neither is the default. Path A has fewer moving parts and suits a conversational assistant where the model should react to what was said. Path B gives more control: a dedicated recognizer can be tuned with custom vocabulary, results can be logged as text before any model sees them, and a chosen neural voice can read the reply. Azure Speech also offers options a chat model does not centre on, such as real-time transcription of streaming audio, batch transcription of many recordings, and speech translation.

The application reaches Azure Speech through the Speech SDK or REST calls, and SSML lets it adjust pitch, speaking rate and pronunciation.

Worked example: a clinic reminder line

A 60-person physiotherapy clinic wants patients to phone a number, say when they would like to move an appointment, and hear a spoken confirmation. A voice request is short and conversational, with no need for a transcript, so path A fits: the audio goes into a deployed audio-capable multimodal model and the reply comes back as speech or text.

Later the clinic adds two requirements: every call must be stored as a text transcript for compliance, and the reply must use one consistent brand voice. Now path B is more natural. Azure Speech recognition produces the transcript, a text model decides the reply, and speech synthesis reads it in the chosen voice.

Easy to confuse

Key points

• Azure Language returns structured results (sentiment, key phrases, entities, summaries) without prompt writing.
• A lightweight text analysis app creates a client for the resource endpoint, sends documents and reads the structured response.
• A deployed multimodal model can take audio in a prompt and answer in one step.
• Azure Speech provides speech recognition, speech synthesis, translation and batch or real-time transcription.
• Choose the model path for conversational reasoning over audio and the Azure Speech path when conversion, tuning or transcripts matter.
• Speech recognition is audio to text; speech synthesis is text to audio.

Check yourself

This lesson belongs to the Implement AI solutions by using Microsoft Foundry domain, which carries 55-60% of AI-901. The full domain recap is on the Foundry implementation page, and the AI-901 cheat sheet keeps the speech terms side by side.