Real-time voice agents are having a breakout moment in cloud AI, and mastering the AI-102 language solutions domain is the fastest way to prove you can design one. Microsoft’s Build 2026 announcements around the Voice Live API and Foundry Agent Service pushed speech-to-text, LLM reasoning, and text-to-speech into a single low-latency pipeline. Therefore, exam scenarios now lean on picking the right service, not just describing what each one does.
This is the third set in our AI-102 practice series. Questions 1-10 covered generative AI with Microsoft Foundry, and Questions 11-20 walked through computer vision. Today, we drill into Azure AI Language, Azure AI Speech, and the newer voice-agent stack. As a result, you’ll see prompts that feel like real design reviews.
Each of these 10 AI-102 language solutions questions ends with a hidden explanation. Reveal the answer only after you commit. In addition, we mix easy, medium, and hard difficulty so nothing feels like a warmup for the real exam. Every scenario is grounded in current Microsoft Learn guidance, so expect the “why this service, not that one” reasoning the real AI-102 exam rewards. Specifically, watch for the tell-tale constraints: latency, consent, language coverage, and paraphrase requirements. These are what separate the right answer from three plausible imposters.
Question 21: Voice Live API for Real-Time Agents
Contoso Financial’s platform team is building a phone-based support agent. It must respond in under one second and handle Spanish, English, and Portuguese. The current design chains a separate Azure AI Speech STT call, an Azure OpenAI chat completion, and a separate TTS call. However, latency and interruption handling remain issues. Which Azure service best fits the requirement?
A) Continue with the chained pipeline but move each step to the same region.
B) Use Azure AI Content Understanding to reason over speech chunks in real time.
C) Use the Voice Live API integrated with Foundry Agent Service.
D) Use Azure AI Translator with speech translation and skip the LLM.
👁 Reveal Answer
Correct Answer: C
Explanation: Voice Live API bundles STT, LLM reasoning, and TTS into one low-latency pipeline built for voice agents. Moreover, it now plugs directly into Foundry Agent Service. As a result, it handles interruptions and multilingual output without the round-trips that hurt the chained design. In addition, the shared pipeline preserves conversational state, which is hard to do across three separate service calls. Content Understanding is for extracting structured data from documents, video, and audio, not for driving live conversations. Translator alone handles speech translation but has no LLM reasoning to answer questions. Therefore, choice C is the only design that meets sub-second latency, multilingual output, and clean interruption handling at the same time.
Question 22: Detecting Incoming Ticket Language
Northwind Insurance receives support tickets in dozens of languages. The intake team must route each ticket to a queue based on language before any translation happens. Which capability of the Azure AI Language service should the team call?
A) Named Entity Recognition
B) Language detection
C) Sentiment analysis
D) Custom text classification
👁 Reveal Answer
Correct Answer: B
Explanation: Language detection returns the primary language code and a confidence score for a text blob. Therefore, it is exactly what routing needs. Named Entity Recognition extracts entities but does not identify the language itself. Sentiment analysis rates tone, not language. Custom text classification could be trained for this, however, it needs labeling; for a standard task, the built-in detection is the right call. As a result, Northwind’s intake pipeline can call the detection endpoint, read the ISO language code from the response, and route the ticket in a single step. In addition, the detection endpoint is a low-cost call, which matters at high volume.
Question 23: Live Speech Translation for a Global Summit
Fabrikam Media hosts a live investor summit streaming to attendees in six countries. The team needs live captions in each attendee’s language and playback of a synthesized voice reading the translation. Which Azure AI Speech capability meets both needs in one flow?
A) Batch transcription with Translator API for post-processing.
B) Speech translation using speech-to-text plus speech-to-speech.
C) Conversation summarization for real-time captions.
D) Personal Voice for each spoken language.
👁 Reveal Answer
Correct Answer: B
Explanation: Speech translation supports real-time multi-language speech-to-text and speech-to-speech translation of audio streams. Specifically, this produces both captions and spoken translations in one pipeline. Batch transcription cannot run live. Conversation summarization gives a summary, not translations. Personal Voice generates a synthetic voice model but does not translate. In addition, Fabrikam can wire the same audio stream into the captions overlay and into an audio output channel for each language. Therefore, the operations team only manages one Speech resource, not one per language pair.
Question 24: Redacting Client Identifiers from Transcripts
Woodgrove Bank stores call transcripts that include internal case IDs formatted as “WB-YYYY-NNNNN”. Compliance requires redacting these before analytics. The built-in PII detection does not recognize the pattern. Which Azure AI Language feature should the team use?
A) Prebuilt PII detection with a regex post-processor.
B) Custom Named Entity Recognition trained on labeled samples.
C) Custom text classification with a “sensitive” label.
D) Key phrase extraction filtered to numeric strings.
👁 Reveal Answer
Correct Answer: B
Explanation: Custom Named Entity Recognition lets teams label examples of the specific pattern and train a model that returns those spans. Because the built-in PII service does not know a client-specific ID format, a regex is fragile against near-matches. Moreover, custom text classification returns a document label, not entity spans, so it cannot mark which characters to redact. Key phrase extraction returns semantically important phrases, not IDs. As a result, only Custom NER gives Woodgrove the character offsets it needs for reliable redaction. In addition, Custom NER can be retrained as the ID format evolves.
Question 25: Batch Sentiment for Product Reviews
The customer success team at Litware imports 40,000 product reviews each night. They need an overall positive, neutral, or negative label plus a score for each review. The team wants the simplest managed option. Which service call fits?
A) Azure AI Language sentiment analysis in batch mode.
B) Azure OpenAI GPT-4 with a custom prompt per review.
C) Azure AI Content Understanding.
D) Azure Databricks with a fine-tuned open source model.
👁 Reveal Answer
Correct Answer: A
Explanation: The Language service sentiment analysis endpoint is the managed, no-code option. In addition, it supports batch documents and returns both a label and confidence scores. Azure OpenAI can classify sentiment but adds cost and prompt design overhead. Specifically, at 40,000 reviews a night, prompt tokens add up quickly. Content Understanding is aimed at multimodal document extraction. A Databricks pipeline is powerful, however, it is far more work than needed for a simple label. Therefore, choice A is both the simplest and the cheapest.
Question 26: Replacing a Retired LUIS Bot (Language Solutions Migration)
The engineering group at Adventure Works Cycles built a chatbot on LUIS three years ago. LUIS is retired. Leadership wants to preserve the training data and continue with a modern service. Which service should the team migrate to?
A) Rebuild in Azure OpenAI with function calling.
B) Import the LUIS app into Conversational Language Understanding.
C) Retrain with Azure AI Bot Service intent classifier.
D) Move to Azure AI Content Understanding.
👁 Reveal Answer
Correct Answer: B
Explanation: Conversational Language Understanding (CLU) is the direct successor to LUIS. Specifically, Language Studio provides an import wizard for existing LUIS apps that preserves intents, utterances, and entities. As a result, three years of tuned training data survives the move. Azure OpenAI is a valid design choice, however, it throws away the LUIS training data. Bot Service supplies the bot channel, not intent classification. Content Understanding extracts fields from documents. In addition, CLU supports multiple projects in one Language resource, which simplifies governance across bots.
Question 27: Executive Summaries of Long Reports
Northwind Traders publishes 30-page quarterly reports. The board wants a two-paragraph “executive summary” generated automatically each quarter. Legal requires paraphrased content, not verbatim sentence lifting. Which Language service feature best fits?
A) Extractive document summarization.
B) Abstractive document summarization.
C) Key phrase extraction.
D) Custom text classification.
👁 Reveal Answer
Correct Answer: B
Explanation: Abstractive summarization generates new sentences that paraphrase the source. As a result, it meets the legal requirement for non-verbatim output. Extractive summarization returns key sentences copied from the source, which is exactly what legal wants to avoid. Key phrase extraction returns terms, not paragraphs. Custom classification returns labels, not text. In addition, the abstractive endpoint lets Northwind request a target length, so producing a two-paragraph output is a single parameter. Therefore, choice B is the shortest path from a 30-page report to a paraphrased executive summary.
Question 28: Expressive Narration with Neural TTS
Tailwind Traders is producing an audiobook. Characters need distinct emotional tones using Azure AI Speech neural voices. Engineers need fine-grained control over prosody, pitch, and speaking style within a single TTS call. Which approach fits best?
A) Use plain text with Personal Voice and no markup.
B) Use SSML with mstts:express-as style attributes.
C) Call the LLM Speech API and let the model choose emotions.
D) Send one TTS request per emotional beat with different voices.
👁 Reveal Answer
Correct Answer: B
Explanation: SSML with the mstts:express-as element gives precise control over speaking style, pitch, rate, and emphasis inside a single request. In contrast, plain text triggers the neural voice default and cannot vary emotion. LLM Speech API focuses on transcription and translation, not TTS. Firing one call per beat works but explodes cost, complexity, and stitching artifacts. Therefore, SSML is the exam-safe answer whenever a scenario mentions “expressive,” “prosody,” or “style.” In addition, SSML travels well across voices, which matters for a multi-character audiobook.
Question 29: Post-Call Summaries for Support QA
The call-center analytics group at Contoso wants a summary of each finished call. Specifically, they need the problem statement and the resolution. Transcripts already exist in JSON. Which Language service capability delivers this out of the box?
A) Question Answering with a custom knowledge base.
B) Conversation summarization with “issue” and “resolution” aspects.
C) Custom NER trained on 100 labeled calls.
D) Abstractive document summarization on the full transcript.
👁 Reveal Answer
Correct Answer: B
Explanation: Conversation summarization is purpose-built for chat and call transcripts. Moreover, it exposes aspect summaries such as “issue” and “resolution” that map directly to what QA wants. Question Answering serves knowledge base questions. Custom NER only extracts entity spans. Document summarization would work, however, it ignores the turn-taking structure and does not offer the aspect split. As a result, Contoso’s QA reviewers get consistent short summaries without building a custom prompt. In addition, the aspect labels feed cleanly into downstream analytics dashboards.
Question 30: Branded Voice for a Voice Live Agent
Midway through this set, Contoso’s marketing arm wants a consistent brand voice for its Voice Live agent. The voice must sound like a recorded executive. However, compliance requires speaker consent and audio watermarking. Which Azure AI Speech feature meets both needs?
A) Custom Neural Voice standard tier.
B) Personal Voice with the required consent workflow.
C) A prebuilt neural voice matched to the executive’s tone.
D) Voice Live API’s default voice with SSML.
👁 Reveal Answer
Correct Answer: B
Explanation: Personal Voice creates a voice model from a small audio sample. In addition, it includes the consent workflow and audio watermarking that Microsoft mandates for synthetic voice deployments. Custom Neural Voice needs a larger data set and a stricter application process; it is overkill when Personal Voice is available. A prebuilt voice cannot match a specific speaker. The default Voice Live voice ignores the branding requirement. Therefore, Personal Voice satisfies both the sound-alike and compliance sides of the ask. As a result, Contoso can stand up the branded voice quickly without a lengthy onboarding cycle.
Study Tips for AI-102 Language Solutions
- Memorize which Language service feature returns entities, labels, or summaries. For example, NER returns spans; classification returns labels; summarization returns text. Because AI-102 leans on distractors that swap these outputs, a fast mental map saves seconds per question.
- Practice picking Voice Live API versus a chained STT-LLM-TTS design. In addition, remember that Foundry Agent Service is the natural home for voice agents. Therefore, whenever a scenario mentions sub-second latency and interruption handling, Voice Live is usually the exam-safe pick.
- Know when SSML is required. Specifically, prosody, style, and multi-voice control all live in SSML, not plain text. Moreover, mstts:express-as is the tag most often referenced in scenario prompts.
- Remember Personal Voice consent and watermarking rules. Therefore, expect governance-flavored answers on any branded voice question. As a result, “requires consent” is a strong signal for Personal Voice rather than Custom Neural Voice.
- CLU is the direct LUIS successor. As a result, LUIS-migration questions almost always point to CLU with the import wizard. However, if the scenario has no LUIS baggage, an Azure OpenAI function-calling design can also be defensible.
Keep Practicing These AI-102 Language Solutions
These AI-102 language solutions questions cover the meat of the exam’s language and speech domain. For extra practice, review the earlier sets in this series and the neighboring AZ-305 review to reinforce cross-service design thinking. Therefore, keep going, one question at a time. In addition, revisit any AI-102 language solutions question you missed after a day or two.
- AI-102 Questions 1-10: Generative AI Solutions with Microsoft Foundry
- AI-102 Questions 11-20: Computer Vision Solutions
- AZ-305 Questions 41-50: Mixed Domain Review
For the deepest reference, browse the official Azure AI Speech overview on Microsoft Learn. Every service in these AI-102 language solutions questions has a page there with quickstarts and pricing.
