Text and Speech Solutions
Objective 2.2, inside the Microsoft Foundry implementation skill area that carries 55–60% of the exam. Two services do most of the work here — Azure Language and Azure Speech in Foundry Tools — and the exam keeps asking the same underlying question: is this a job for a purpose-built feature, or for a model?
Azure Language in Foundry Tools
Azure Language is the text service. You send it text, it sends back structured findings — typed, scored, and consistent. It does not write prose, and that predictability is exactly why it still exists alongside generative models.
Microsoft now splits its features into two tiers, and the distinction confuses people, so be precise about it.
| Tier | Features | What the tier means |
|---|---|---|
| Core | PII detection, language detection, named entity recognition (prebuilt and custom), text analytics for health | Actively evolving. New capability lands here. |
| Legacy | Sentiment analysis and opinion mining, key phrase extraction, summarization, entity linking, conversational language understanding, question answering, custom text classification | Fully supported and working. Not receiving new investment. |
This trips people up badly. “Legacy” here means Microsoft is not adding new capability — it does not mean deprecated, retired, or wrong. The exam objective names “keyword extraction, entity detection, sentiment analysis, and summarization” explicitly, and three of those four are in the legacy tier. They are still correct answers. Do not eliminate a choice because you remember the word legacy.
What each feature actually returns
| Feature | Returns | Example scenario |
|---|---|---|
| Key phrase extraction | A flat list of the main talking points, as phrases | Summarize what a thousand survey comments are about, without reading them |
| Named entity recognition | Typed spans: person, location, organization, datetime, quantity, and more | Pull every company and date mentioned across a set of news articles |
| PII detection | Detected personal information, with the option to redact it | Strip customer identifiers out of support transcripts before analysis |
| Sentiment analysis | Positive, negative, neutral, or mixed, with confidence scores | Track how review tone shifts after a product change |
| Opinion mining | Sentiment tied to a specific aspect, not just the whole document | Learn that guests like the room but dislike the check-in process |
| Language detection | The detected language and a confidence score | Route incoming messages to the right regional team |
| Summarization | A shorter version of the text, extractive or abstractive | Give agents a two-line recap at the top of a long ticket |
| Text analytics for health | Medical entities, relations, and assertions | Extract medications and dosages from clinical notes |
| Conversational language understanding | An intent plus entities from a user utterance | Work out that “book me a flight to Seattle Friday” means BookFlight |
| Question answering | An answer drawn from a knowledge base of question-and-answer pairs | Answer common questions from an existing FAQ page |
Two distinctions the exam leans on
Key phrases versus entities
Key phrases are the main talking points, returned as a plain list. Nothing is typed or categorized.
Entities are typed things — this span is a person, that one is a date. If the question mentions categories, types, or “identify all the organizations,” it wants entity recognition.
Extractive versus abstractive
Extractive summarization selects the most important sentences that already exist, unchanged. Nothing is invented, and every word traces back to the source.
Abstractive summarization writes new sentences that capture the meaning. It can paraphrase and condense across the whole document. If a scenario needs a summary in the reader’s own words, it is abstractive.
Azure Speech in Foundry Tools
Azure Speech handles audio in both directions: turning speech into text, and turning text into speech. It is deterministic, low latency, and built for production audio pipelines.
| Capability | Input | Output | Example scenario |
|---|---|---|---|
| Speech to text (real-time) | Streaming audio | Text, as it is spoken | Live captions during a meeting |
| Batch transcription | Many stored audio files | Text, asynchronously | Transcribe last month’s call recordings overnight |
| Fast transcription | A pre-recorded file | Text, quickly | Turn an uploaded voicemail into text while the user waits |
| Custom speech | Audio plus your own training data | Text, tuned to your vocabulary | Recognize product names and industry jargon reliably |
| Text to speech | Text | Natural neural-voice audio | Read confirmation details back to a caller |
| SSML | Marked-up text | Audio with controlled pronunciation, pace, and emphasis | Make the system say a product name correctly every time |
| Custom neural voice | Recorded voice samples (gated) | A brand-specific synthetic voice | One consistent voice across every customer touchpoint |
| Speech translation | Spoken audio in one language | Text or speech in another | Let a support agent and a caller speak different languages |
| Speaker recognition | Audio plus enrolled voice profiles | Who is speaking, or whether they are who they claim | Verify a caller by voice before discussing an account |
| Pronunciation assessment | Learner audio | Accuracy and fluency scoring | Give feedback in a language-learning app |
| Language identification | Audio | Which language is being spoken | Pick the right transcription model automatically |
| Voice Live | Streaming conversation | Low-latency spoken interaction | A natural, interruptible voice assistant |
These sound almost identical and mean completely different things. Speech recognition answers what was said and produces a transcript. Speaker recognition answers who said it and is used for identification and verification. Any scenario about confirming a caller’s identity by voice is speaker recognition, no matter how much transcription is also happening.
Responding to spoken prompts with a multimodal model
The objective says “respond to spoken prompts by using a deployed multimodal model,” and that is a different thing from the Speech service. The realtime and audio model families in Foundry Models take speech in and produce speech out, with a language model reasoning in between. There is no separate transcribe step you manage, no separate synthesis step, and crucially the model understands and answers rather than just converting formats.
| Azure Speech | Multimodal audio model | |
|---|---|---|
| What it does | Converts between audio and text | Understands spoken input and answers it |
| Reasoning | None — it is a converter | Full language model reasoning between input and output |
| Output | A transcript, or synthesized audio you supplied the words for | A spoken answer the model composed itself |
| Reach for it when | You need an accurate transcript, a branded voice, or speaker verification | You want a conversational assistant that listens and replies |
Whisper also appears in Foundry Models as a transcription model. It converts speech to text; it does not hold a conversation.
Purpose-built feature or generative model?
This is the decision the exam returns to again and again, in both directions. Neither answer is always right, and the scenario always contains the deciding constraint.
| Situation | Better choice | Why |
|---|---|---|
| Output must be consistently structured and typed | Purpose-built feature | It returns the same shape every time. A model may phrase things differently on each call. |
| High volume, latency and cost matter | Purpose-built feature | Narrow models are faster and cheaper per call than general-purpose ones. |
| Regulated domain needing predictable behavior | Purpose-built feature | Documented, auditable behavior is easier to defend than free-form generation. |
| Clinical or medical text | Text analytics for health | Purpose-built for medical entities, relations, and assertions. |
| Detecting and redacting personal information | PII detection | Built for exactly this, with redaction included. |
| Verbatim transcription, or a branded voice | Azure Speech | Deterministic transcription, custom neural voice, and SSML control live here. |
| Confirming who is speaking | Azure Speech speaker recognition | No generative model does voice identity. |
| The task is novel, or combines several steps in one call | Generative model | One prompt can summarize, classify, and rewrite together. |
| The task needs reasoning about the content | Generative model | Purpose-built features detect and label; they do not reason. |
| Input mixes text with images or audio | Multimodal model | Purpose-built text features only take text. |
| A spoken, conversational assistant | Multimodal audio model | Speech in, reasoning, speech out, in one model. |
Azure Translator in Foundry Tools
Translator is the third service in this objective and the one people misplace most often. It does text translation between languages, document translation that preserves formatting, and transliteration.
Text translation lives in Translator; spoken translation lives in Azure Speech. If the input is audio, the answer is Speech, even though the task is translation.
Transliteration is not translation. It rewrites text from one script into another — Hindi written in Latin characters, for example. The language does not change, only the alphabet. A question about making a name readable in another script wants transliteration.
Where people lose points
| The mistake | What is actually true |
|---|---|
| Eliminating sentiment analysis or key phrase extraction because they are “legacy” | Legacy means no new investment, not retired. The objective names these features explicitly and they remain correct answers. |
| Answering key phrase extraction when the scenario asks for types | Categories such as person, location, or organization mean named entity recognition. |
| Treating extractive and abstractive summarization as interchangeable | Extractive selects existing sentences; abstractive writes new ones. Scenarios that need paraphrasing are abstractive. |
| Choosing speech recognition for a caller-identity scenario | Identity by voice is speaker recognition. Speech recognition only produces the transcript. |
| Sending audio to Azure Translator | Translator takes text. Spoken translation is Azure Speech. |
| Calling transliteration a kind of translation | Transliteration changes the script, not the language. |
| Reaching for a generative model for everything | When the scenario stresses consistent structure, latency, cost, or compliance, the purpose-built feature is the better answer. |
| Reaching for a purpose-built feature for everything | When the task is novel, needs reasoning, or combines steps, a model wins. |
| Confusing question answering with conversational language understanding | A knowledge base of question-and-answer pairs means question answering. Working out intent from a free-form utterance means conversational language understanding. |
The night-before cheat sheet
Azure Language, one line each
- Key phrases — the main talking points, untyped
- Entities — typed things: person, place, org, date
- PII detection — find and redact personal data
- Sentiment — positive, negative, neutral, mixed
- Opinion mining — sentiment about a specific aspect
- Language detection — which language is this
- Summarization — extractive selects, abstractive rewrites
- Health — medical entities and relations
- CLU — intent plus entities from an utterance
- Question answering — answers from a knowledge base
Azure Speech, one line each
- Speech to text — real-time, batch, or fast
- Custom speech — tuned to your vocabulary
- Text to speech — neural voices
- SSML — control pronunciation and pace
- Custom neural voice — brand voice, gated
- Speech translation — spoken, across languages
- Speaker recognition — who is talking
- Voice Live — low-latency conversation
The decisions, in one line each
- Core versus legacy — legacy still works and is still the right answer when the objective names it.
- Key phrases versus entities — types and categories mean entities.
- Extractive versus abstractive — selects existing sentences versus writes new ones.
- Speech versus speaker — what was said versus who said it.
- Translator versus Speech — text in, or audio in.
- Transliteration — new script, same language.
- Purpose-built versus generative — consistency, latency, cost, and compliance point to purpose-built; novelty, reasoning, and combined steps point to a model.
- Speech service versus multimodal audio model — converting formats versus understanding and answering.