Text Analysis Solutions — Azure AI Apps and Agents Developer AI-103 practice questions

A media archive holds 40,000 hours of recorded audio that must be transcribed overnight, with no user waiting on any single result. Which Azure Speech in Foundry Tools mode is MOST appropriate?

1. Real-time speech to text over a separate streaming connection per file
2. Speech translation, since it also emits a transcript of the source audio
3. Batch transcription against the audio files already held in storage
4. Text to speech round-tripped to validate the archive before transcription

Answer: C — Batch transcription against the audio files already held in storage

C) Correct — batch transcription is the asynchronous mode designed for large volumes of stored audio, processing files without a live connection. A) tempts because it uses the same recognition models, but streaming is engineered for low latency on live input and scales poorly across tens of thousands of hours. B) does produce source text, but it adds an unnecessary translation step and cost. D) is the reverse capability and does nothing for transcription.

A brand wants a synthetic voice that reproduces a specific named spokesperson for its automated announcements. Which statement about custom neural voice is MOST accurate?

1. It is generally available to any subscription once the resource tier is upgraded
2. It is a limited access capability requiring registration, approval, and voice talent consent
3. It requires no consent recording provided the spokesperson is an employee of the brand
4. It is unrestricted for internal use and gated only when the audio is published publicly

Answer: B — It is a limited access capability requiring registration, approval, and voice talent consent

B) Correct — custom neural voice sits behind limited access: you apply for approval, and you must supply recorded consent from the voice talent plus abide by the responsible AI terms. A) tempts because most speech features do unlock with tier, but this one is gated by review rather than pricing. C) is precisely the misconception the consent requirement exists to prevent; employment status is irrelevant. D) invents an internal/external distinction — the gate applies to creating the voice, not to where the audio is played.

Speaker verification is best described as:

1. Transcribing an audio stream into time-stamped lines of text
2. Matching a voice against many enrolled profiles, one to many
3. Confirming a claimed identity against one enrolled profile
4. Synthesising speech in a voice built from enrolment audio

Answer: C — Confirming a claimed identity against one enrolled profile

C) Verification is a one-to-one check: the caller asserts an identity and the audio is compared with that single enrolled profile. A) Transcription produces text and asserts nothing about identity. B) One-to-many search is identification. D) Building a voice is synthesis.

Document translation in Azure Translator in Foundry Tools provides:

1. Asynchronous translation of whole documents in Blob Storage
2. Synchronous requests only, with a one-page limit per call
3. Translation that discards the original formatting of the document
4. A replacement for speech translation in call centre apps

Answer: A — Asynchronous translation of whole documents in Blob Storage

A) You point the job at a source container and a target container and poll the long-running operation, and the layout of each document is preserved. B) The batch route is asynchronous by design. C) Preserving formatting is the point of the feature. D) Spoken audio is handled by the speech service.

An inbound-mail router must place each message into exactly one of eight internal departments based on its wording. Volume is high, the eight categories are stable, and cost per message matters. Which approach is MOST appropriate?

1. Send each message to a large generative model with the eight departments described in the prompt
2. Train a custom text classification project with a single-label configuration
3. Train a custom NER project and route on whichever entity type appears most often
4. Use key phrase extraction and route on a hand-maintained phrase-to-department table

Answer: B — Train a custom text classification project with a single-label configuration

B) Correct — a stable, closed label set at high volume is the textbook case for single-label custom text classification: a small trained model, low per-call cost, measurable accuracy. A) tempts because a generative model needs no labeled data and would work, but it costs far more per message and gives no per-class evaluation metrics for a task this narrow. C) misuses an extraction model as a classifier. D) is brittle and needs constant manual curation as vocabulary drifts.

Training a custom neural voice requires:

1. English-only training text, with no other language supported
2. No approval at all; any subscription may train a voice
3. Limited Access approval and recorded voice talent consent
4. A GPU on the device where the voice will be synthesised

Answer: C — Limited Access approval and recorded voice talent consent

C) Custom neural voice is gated by responsible AI review, and you must show recorded consent from the voice talent before training. A) The feature is not limited to one language. B) Open self-service training is exactly what the gating prevents. D) Synthesis is a managed cloud service.

Which Azure Language in Foundry Tools capability best redacts national IDs, phone numbers and addresses?

1. Sentiment analysis with opinion mining on each sentence
2. General named entity recognition over the whole document
3. PII detection, which returns categories and text offsets
4. Key phrase extraction ranked by relevance to the text

Answer: C — PII detection, which returns categories and text offsets

C) The PII capability is tuned for sensitive categories such as government identifiers and phone numbers, and returns offsets plus an optional redacted copy of the text. A) Sentiment scores tone, not identifiers. B) General entity recognition finds people and places but misses government IDs and card numbers. D) Key phrases summarize topics.

Sentiment analysis in Azure Language in Foundry Tools returns at minimum:

1. Sentence and document sentiment with confidence scores
2. Translated text in each of the requested target languages
3. Speaker labels aligned to the segments of an audio file
4. Vector embeddings of each sentence in the input document

Answer: A — Sentence and document sentiment with confidence scores

A) Each sentence is scored and the document label is derived from those sentence scores, with confidence values for positive, neutral and negative. B) Translation is a different service. C) Speaker labels come from speech processing. D) Embeddings are produced by a model deployment, not this operation.

Custom text classification in Azure Language in Foundry Tools supports:

1. Single-label and multi-label classification of whole documents
2. Sentiment scoring only, at the sentence and the document level
3. Entity extraction only, using categories that you label yourself
4. Language detection only, returning a script and a confidence

Answer: A — Single-label and multi-label classification of whole documents

A) A project is created as either single-label or multi-label, so a document can carry one category or several. B) Sentiment is a separate prebuilt capability. C) Extraction is custom entity recognition, a different project type. D) Language detection is prebuilt and needs no labelling.

A speech transcription model in Foundry Models is BEST suited to which task?

1. Forecasting future values in tabular time series collected from sensors
2. Synthesizing natural-sounding spoken audio from a written input script
3. Generating illustrative images from a written description of the scene
4. Converting recorded speech to text, including translation into English

Answer: D — Converting recorded speech to text, including translation into English

D) Right — the model turns audio into a text transcript and can also translate the spoken content into English. A) Time series forecasting is not a speech workload. B) That is text-to-speech, the inverse capability. C) Image generation is a different model family entirely.

Real-time vs batch speech-to-text?

Real-time streams audio over a persistent connection and returns partial and final results with low latency. • Batch takes audio already stored at a URL and transcribes asynchronously, which suits high volume and long recordings.

Speaker recognition modes?

Verification is one-to-one: does this audio match the claimed enrolled speaker? • Identification is one-to-many: which enrolled speaker in a group is talking? • Both need enrollment audio first.

What does key phrase extraction return, and what is it good for?

It returns the main talking points of a document as an unranked list of noun phrases, with no scores or categories. It suits tagging, clustering and quick triage of large text volumes; it does not classify or summarize.

Azure Speech in Foundry Tools — capabilities?

Speech-to-text, including real-time and batch transcription. • Text-to-speech with neural voices. • Speech translation. • Speaker recognition. • Pronunciation assessment.