AI-901 sample exam questions

SpeechConfig holds the Speech resource key, region or endpoint, and language settings that every recognizer or synthesizer uses to connect to the service. AudioConfig and AudioOutputConfig only choose the microphone, speaker or file, and SpeechSynthesizer is the class that performs text to speech.

System instructions (the system message or agent instructions) define the agent's persona, role, tone and rules for how it responds. Temperature only controls randomness, TPM is a deployment rate limit, and max completion tokens just caps response length.

A system prompt sets the model's role, tone and rules. It has no security power (authorization is done by identity and access control) and has nothing to do with the TPM rate limit.

An Azure Content Understanding analyzer lets you define a field schema, including nested objects and arrays, and returns structured values extracted from documents such as invoices. OCR alone returns only raw text, Speech transcription handles audio, and Azure AI Search indexes and searches content rather than extracting schema fields.

Instructions (the system prompt) tell the model what role to play and how to behave, and they set constraints such as tone, format, scope and topics to avoid. Model choice, the Azure region and the TPM quota are deployment settings, not something a prompt instruction controls.

Models are trained once and then used as-is. Agents generate answers rather than copying stored ones, and use the model to decide which steps or tools to use.

Fairness is about treating all groups equally, so comparing results across demographic groups to find and reduce bias is a fairness practice.

Inclusiveness means designing AI so that people of all abilities and backgrounds can use it, for example multiple languages and screen-reader support. Encryption and access control belong to privacy and security, testing across demographic groups is fairness, and explaining predictions is transparency.

With a vision-enabled chat model you send one user message whose content array holds both a text part and an image_url part, so the model handles both in a single call. Splitting them into separate requests adds round trips, and the system message is meant for instructions, not user-supplied images.

Content Understanding processes audio by first running transcription (speech to text, with options such as speaker diarization) and then extracting fields from the transcript. OCR reads text from images and documents, key phrase extraction works on text that already exists, and Voice Live is a real-time voice agent API.

Content Understanding analysis is a long-running operation, so the Python SDK's begin_analyze() returns a poller and poller.result() waits for the operation to finish and returns the extracted fields. Results are not returned synchronously or in the response headers, and plain OCR with hand-written mapping skips the analyzer.

A multimodal generative model works across text and images, for example generating an image from a text description. Object detection is classic computer vision and categories from numbers is classification.

Turning audio into text is speech recognition (speech to text). The text-based workloads only start after the speech is transcribed.

The community answer A is wrong because thumbnails and keyframes for videos come from Azure AI Video Indexer, and Vision's smart thumbnails work only on still images. Azure Vision's OCR (Read) feature pulls printed and handwritten text out of images such as scanned letters. Key phrases come from Azure Language, and breed identification in live video would need a custom model.

PII detection in Azure Language finds personal data such as names, phone numbers and email addresses and returns redacted text with those values masked. Entity linking maps entities to Wikipedia entries, while custom NER and custom text classification need you to label and train your own models and do not redact anything.