Implement text analysis solutions is worth 10-15% of AI-103, and it covers more than classic text analytics. You need to know when to use Azure Language and Azure Translator in Foundry Tools, when a language model with a good prompt is the better choice, and how to add speech to agents with transcription, speech synthesis, and real-time voice.
This is the sixth post in our AI-103 series (Questions 51-60). Each question is an original practice scenario mapped to the official skills outline, not a copy of a real exam item.
Several questions here turn on one detail, such as whether a feature works in batch mode or needs a storage account. Read each scenario’s requirements carefully before you reveal the answer.
Question 51: Removing Personal Data From Transcripts
A contact center stores chat transcripts and sends them to a language model for weekly trend analysis. Privacy rules require that names, phone numbers, email addresses, and card numbers are removed before the text is stored or analyzed. Which Foundry Tools capability fits best?
A) PII detection in Azure Language, using the redacted text it returns
B) Sentiment analysis in Azure Language
C) Key phrase extraction in Azure Language
D) Content Safety text moderation for hate and violence
👁 Reveal Answer
Correct Answer: A
Explanation: PII detection identifies personal information such as names, phone numbers, and financial details and can return redacted text, which you can store and pass on to other systems. It supports text, conversations, and documents, and Microsoft lists it as a core capability recommended for new development. B scores opinion, not personal data. C may even surface names as key phrases. D looks for harmful content, not personal data.
Reference: What is Azure Language in Foundry Tools?
Question 52: Detecting Tone, Not Just Positive or Negative
A support team wants every incoming email labeled with a tone from a fixed list (frustrated, confused, urgent, satisfied), a one-sentence reason, and a flag if the customer threatens to cancel. The result must be valid JSON that their ticket system can import. What should they build?
A) Sentiment analysis, mapping negative to frustrated and positive to satisfied
B) Key phrase extraction, then a keyword rule for each tone
C) A language model prompt with the tone definitions and examples, plus structured outputs that use a JSON schema with an enum for the tone field
D) JSON mode with no schema, so the model can pick its own field names
👁 Reveal Answer
Correct Answer: C
Explanation: The outline includes detecting tone and producing structured JSON with generative prompting. A model can apply custom labels that classic sentiment analysis does not offer, and a schema with an enum restricts the tone to the allowed values, while a boolean field captures the cancellation flag. A only gives positive, negative, neutral, or mixed, and Microsoft now lists sentiment analysis among Azure Language’s legacy capabilities. B is brittle and misses tone expressed without obvious keywords. D gives valid JSON but no guaranteed field names, so imports would break.
Reference: How to use structured outputs
Question 53: Translating One Uploaded File Right Away
A web app lets a user upload a single Word document and immediately download a Spanish version with the original formatting preserved. The team does not want to manage a storage account for this feature. Which Translator option fits?
A) Asynchronous (batch) document translation
B) Synchronous document translation
C) Text translation of the document’s extracted text
D) Custom Translator model training
👁 Reveal Answer
Correct Answer: B
Explanation: Synchronous document translation translates a single file, optionally with a glossary, preserves its structure and format, returns the translated document directly in the response, and does not need an Azure Blob Storage account. A is designed for batches and complex files, and it requires Blob Storage containers for source and target documents. C loses the formatting. D improves terminology over time but is not how you translate a file on demand.
Reference: What is Azure Translator in Foundry Tools?
Question 54: Teaching Translations a Company’s Terminology
A medical device company has about 2,000 approved English-German sentence pairs that use its exact product terms. It wants language model-based translations to follow this terminology, and it wants the solution ready within hours, not weeks. What should it use?
A) A phrase list in Azure Speech
B) Markup that prevents the product names from being translated
C) Standard translation followed by a find-and-replace script
D) Adaptive custom translation in Azure Translator, uploading the approved sentence pairs
👁 Reveal Answer
Correct Answer: D
Explanation: Adaptive custom translation lets you upload between 5 and 10,000 prealigned source and target segment pairs. The service builds a custom bilingual index in minutes and uses it to adapt supported language models, so translations follow the approved wording. A is a speech recognition feature and has nothing to do with translation. B keeps names in English, but the company needs correct German terms, not untranslated ones. C is fragile because grammar changes around replaced words.
Reference: What is Azure Translator in Foundry Tools?
Question 55: Fast Transcripts for Voicemail
An insurance company receives voicemails as MP3 files, each a few minutes long. An agent must read the transcript within seconds of each file arriving so it can create a claim. Which Azure Speech option fits best?
A) Batch transcription, checking for results every hour
B) Fast transcription, which returns results synchronously and faster than real time
C) The Voice Live API
D) Text to speech
👁 Reveal Answer
Correct Answer: B
Explanation: The fast transcription API transcribes audio files and returns the result synchronously, faster than real time and with predictable latency. Microsoft lists voicemail among its typical uses. Files must be under 5 hours and 500 MB, and MP3 is supported. A is built for large volumes of recordings where results can arrive later. C is for live, two-way voice conversations, not recorded files. D does the opposite job.
Reference: Use the fast transcription API
Question 56: Product Names Misheard in Batch Transcription
A company transcribes thousands of hours of recorded sales calls every week with batch transcription. Product names such as “Zentrix” and “Q-Flow” are often transcribed incorrectly. A developer suggests adding a phrase list. What should the team do?
A) Add the phrase list to the batch transcription request
B) Add the product names to an SSML lexicon
C) Train a custom speech model with text data that includes the product names, and use it in batch transcription
D) Convert the recordings to a higher bit rate
👁 Reveal Answer
Correct Answer: C
Explanation: Phrase lists are a quick, no-training fix, but Microsoft’s documentation states that the batch transcription API does not support them, so A fails. A custom speech model trained with plain or structured text that contains the domain vocabulary improves recognition of those terms. For batch transcription you do not even need a hosted custom endpoint, which saves cost. You can measure the improvement with word error rate (WER). B affects text to speech pronunciation, not recognition. D does not teach the model new words.
Reference: What is custom speech?
Question 57: A Voice Agent That Mispronounces Things
A voice assistant uses text to speech. It says the product name “Xyra” the wrong way, and it reads the delivery date “11/04” as a fraction. How should the developer fix both problems?
A) Use SSML: a phoneme element or sub alias for the product name, and say-as with a date format for the delivery date
B) Train a custom speech model with the product name
C) Add the product name to a speech recognition phrase list
D) Switch to a different neural voice
👁 Reveal Answer
Correct Answer: A
Explanation: Speech Synthesis Markup Language (SSML) controls how text to speech reads your text. Phonemes and custom lexicons fix pronunciation, a sub alias speaks replacement text, and say-as tells the voice to read a value as a date, telephone number, and so on. B and C improve speech recognition (speech to text), which is the opposite direction. D changes the voice but not how the text is interpreted.
Question 58: A Natural Real-Time Voice Agent
A hotel wants a phone concierge that talks with guests in real time. It must handle background noise, avoid hearing its own voice, let guests interrupt naturally, and call booking functions. The team does not want to wire together separate speech-to-text, model, and text-to-speech services. What should they use?
A) Batch transcription, a chat model, and text to speech, called one after another for each turn
B) Fast transcription for each sentence the guest says
C) Azure Translator text translation
D) The Voice Live API, connected from the hotel’s backend over WebSockets
👁 Reveal Answer
Correct Answer: D
Explanation: The Voice Live API combines speech recognition, a generative model, and text to speech in one fully managed interface for low-latency voice agents. It includes noise suppression, echo cancellation, interruption detection, advanced end-of-turn detection, optional avatars and custom voices, and function calling. Backends connect over WebSockets. A adds delay at every step and is exactly the integration work the team wants to avoid. B handles recorded files, not live conversation. C translates text and does not provide a voice experience.
Reference: Voice Live API overview
Question 59: Translated Subtitles From Recorded Training
A company records English training sessions and needs Korean text for subtitles. It also wants to guide the output style with a short instruction, such as keeping product names in English. Which option fits best?
A) Fast transcription with a translation option
B) Document translation of the audio files
C) LLM Speech with the translate task, a target language of Korean, and a prompt that guides the output
D) Custom voice training in Korean
👁 Reveal Answer
Correct Answer: C
Explanation: LLM Speech enhances a speech model with a large language model. It supports both transcription and translation of audio, plus custom prompting to guide the output style, and Korean is one of its translation targets. Fast transcription (A) does not support translation or custom prompting, which is why Microsoft points to LLM Speech for those needs. B works on documents, not audio. D creates a synthetic voice and does not produce subtitles.
Reference: LLM Speech
Question 60: When the Words Are Not Enough
A quality team reviews short call recordings to spot sarcasm and rising frustration. Text transcripts of these calls often look polite even when the caller clearly sounds annoyed. Which approach can reason over how things were said, not only what was said?
A) Send the audio directly to an audio-capable model, such as one that accepts audio input in the chat completions API, and ask it to assess tone
B) Transcribe the calls and run text sentiment analysis on the transcript
C) Translate the transcripts into English before analysis
D) Run key phrase extraction on the transcripts
👁 Reveal Answer
Correct Answer: A
Explanation: The outline includes enabling multimodal reasoning from audio inputs. Audio-capable models in Foundry bring the audio modality into the chat completions API, so the model hears the recording itself, including tone of voice, emphasis, and pauses. B, C, and D all work from a transcript, which is exactly where the sarcasm and frustration were lost. Note that Microsoft documents that the content filtering system is not applied to prompts and completions handled by some audio models, such as Whisper, so check the safety controls for the model you choose.
Reference: Audio models in Azure OpenAI
Study Tips for Text Analysis and Speech
- Know what is core and what is legacy. PII detection, NER, and language detection are core in Azure Language; many custom labeling tasks now fit language models with structured outputs better.
- Translator has two document modes. Synchronous for one file with no storage account; asynchronous for batches with Blob Storage.
- Pick the speech mode by timing. Live conversation: Voice Live or real time. Recorded file needed now: fast transcription. Large backlog: batch transcription.
- Recognition vs synthesis. Phrase lists and custom speech fix what the service hears; SSML fixes what it says.
- Audio carries meaning. When tone matters, let an audio-capable model listen instead of reading a transcript.
Keep Practicing for AI-103
Review vision in AI-103 Questions 41-50, then finish this batch with Questions 61-70 on retrieval pipelines and Content Understanding. The official AI-103 study guide lists every skill measured.