Microsoft AI Engineer Certification Sample Questions

The correct answer is SpeechRecognizer.

The SpeechRecognizer is the Azure Speech SDK object that is designed to convert spoken audio into text. It supports streaming and continuous recognition from live audio such as phone calls and provides built in events and callbacks so you can transcribe callers and then have your backend lookup purchase records with minimal custom audio handling code.

The Speech to Text REST API can also transcribe audio but it is a REST service rather than an SDK object. That option requires managing HTTP uploads and responses and is not the specific SDK object the question asks for.

The SpeechSynthesizer performs the opposite task by converting text into spoken audio. It is not used for transcribing caller audio into text so it is not correct.

When a question asks for an SDK object prefer classes in the Speech SDK such as SpeechRecognizer for live or streaming transcription and use REST endpoints for service or batch scenarios.

The correct option is Deployment identifier endpoint address and API key.

The Azure OpenAI SDK requires the resource endpoint address and an API key for authentication and it also requires the deployment identifier to select which deployed model to call. The endpoint and key grant access to the Azure OpenAI resource and the deployment identifier maps to the specific deployment you created, so all three pieces are required when using the Azure SDK.

Azure Active Directory tenant ID client ID and client secret is incorrect because that set describes Azure AD client credentials. That approach represents an alternative authentication flow and is not the minimal configuration the Azure OpenAI SDK examples expect by default, which use the endpoint, key, and deployment identifier.

Endpoint address API key and model name is incorrect because Azure OpenAI requires the deployment identifier rather than a raw model name when calling the Azure endpoint. The deployment name is what you configure and reference in requests, so specifying a model name alone is not sufficient for the Azure SDK.

When you study these questions remember that the Azure SDK targets a specific deployment identifier on a resource and that you must provide the resource endpoint and API key to authenticate and route requests.

The correct option is GeographyV2.

GeographyV2 is the Dialogflow system entity for geographic locations such as cities countries and regions. The example value “Rome” is a city so using GeographyV2 lets the agent recognize the location and return normalized geographic information for fulfillment.

Machine learned is a generic learned entity type that can capture arbitrary values from training phrases but it does not provide the built in geographic normalization that GeographyV2 offers so it is not the best choice for city names.

List requires you to predefine a closed set of values and synonyms and it is not practical to enumerate all possible cities so it is not appropriate for a general city slot like “Rome”.

Regex matches text by pattern and is useful for structured formats but city names do not follow a consistent pattern so a regex is not a reliable solution for recognizing locations.

Email is intended to match email address formats and would be appropriate for the email example in the utterance but it is not suitable for labeling a city name.

When an utterance contains a city or country choose a geography system entity if available because it gives built in recognition and normalization that a generic or manual list will not.

The correct option is Entity Recognition.

The built in Entity Recognition skill is designed to identify named entities in text and it explicitly extracts locations and place names so you can index or enrich them for search scenarios.

The skill labels entities with types such as person organization and location and it returns the entity text along with metadata so you can map place names to fields or external systems.

Key Phrase Extraction identifies salient phrases that summarize content and it does not label entities as locations so it will not reliably extract place names.

Azure Text Analytics is a broader service that provides natural language features and it is not the specific built in skill asked for. Some built in skills may call the service behind the scenes but the correct answer is the specific skill that extracts entities.

Language Detection determines the language of the text and it does not extract named entity types such as place names so it is not the correct skill for extracting locations.

When a question asks about extracting place names look for skills that mention entities or locations in their name and avoid answers that are service names rather than specific built in skills.

The correct answer is Retrain the existing custom model after adding samples of the new contract layout to the current training dataset.

Retraining the existing custom model with additional examples of the new layout allows the model to learn the new document structure while preserving its ability to extract fields from previously supported layouts. This approach typically requires minimal or no application code changes because the service endpoint and extraction logic remain the same and only the model artifact is updated.

Expanding the current training dataset and retraining keeps development effort low because you avoid adding routing logic or managing multiple model endpoints. It also centralizes maintenance and makes future updates simpler since you only need to add representative samples and retrain.

Create a separate training dataset for the new contract layout and train a new custom model is not ideal because it forces you to manage an additional model and to modify the application to choose the correct model at runtime. That increases code changes and ongoing operational overhead.

Use the Azure Document Intelligence prebuilt contracts model is not the best choice when you already have a custom model that handles specific fields and labels. Prebuilt models can be useful for common standard formats but they often do not match organization specific fields and they would require you to validate field mappings and possibly adjust your extraction logic.

Lower the application confidence threshold for extracted fields is incorrect because changing the threshold does not teach the system how to handle a new layout. Lowering the threshold may increase recall but it will also raise false positives and reduce overall extraction quality, so it does not solve the layout support issue.

When a question emphasizes minimal code changes look for answers that update training data or retrain an existing custom model rather than replacing models or changing application thresholds. Pay attention to words like retrain and existing model.

Yes is correct because developer documentation commonly treats the terms “subscription key” and “API key” as equivalent when they refer to a simple access token used to call an API.

Both names often describe the same practical artifact which is an opaque string issued to a client to identify and authorize requests and to apply quotas and billing. Documentation and examples across cloud providers frequently use one term or the other depending on product naming, but the underlying concept and usage are the same in most cases.

There can be a nuance where a provider uses the term subscription key to emphasize that the key is tied to a subscription or a specific plan and uses API key as a more generic term. Those are naming differences rather than fundamental technical differences and they do not contradict the common practice of treating the terms as equivalent in developer docs.

No is incorrect because it asserts that the terms are not commonly treated as equivalent, and that does not match how many vendor guides and examples present these tokens in practice.

When a question asks about two terms check how they are used in practice and focus on the function rather than the exact name. If both names describe the same access token choose the answer that reflects common usage.

The correct answer is Person.

Person is the PII category used to identify human names so Priya should choose it to flag employee names in loan files when using Azure AI Language PII detection.

PhoneNumber is incorrect because that category only detects telephone numbers and not personal names.

DateTime is incorrect because that category detects dates and times rather than staff names.

Age is incorrect because that category matches age expressions and not employee names.

Pick the PII category that most closely matches the data type and test against real documents to verify detection and reduce false positives.

The correct answer is Personally identifiable information.

This term refers to information that can be used to identify an individual either on its own or when combined with other data. Names and email addresses and mailing addresses and IP addresses and phone numbers and government identifiers are all examples of information that fit this description because they can be linked back to a specific person.

Azure Purview is incorrect because it is a Microsoft data governance and classification service rather than the name of a type of sensitive data. It helps find and label sensitive data but it is not the general category for those identifiers.

Protected health information is incorrect because it applies specifically to health related information protected under regulations such as HIPAA. It is a subset of personal data and would only be the right choice if the listed identifiers were tied to health information.

When a question lists identifiers such as names or IP addresses think of PII and distinguish it from product names or domain specific terms like PHI.

The correct answer is Azure Computer Vision.

Azure Computer Vision provides optical character recognition and the Read API which are designed to extract printed text from photographs. The service returns detected text with bounding boxes and supports a range of image types and languages, so it is well suited to finding and reading promotional stickers and printed sale prices in distributor photos.

Azure AI Studio is a development and management environment for building and deploying AI models and experiences and it is not the specific OCR engine that reads text from images. You would use a service such as Azure Computer Vision to perform the text extraction itself.

Form Recognizer focuses on structured documents and extracting key value pairs and tables from forms, receipts, and invoices. It can be useful for parsed documents but it is not the general purpose OCR tool for arbitrary photographs of stickers and price tags.

Custom Vision is intended for image classification and object detection. It can be trained to recognize whether an image contains a sticker or label but it does not transcribe or extract the printed text itself so it cannot read sale prices.

When a question asks about extracting printed text from photos look for services that mention OCR or a Read API and choose the computer vision OCR capability rather than image classification tools.

The correct option is Add a new question and answer pair.

Add a new question and answer pair is correct because the knowledge base already contains answers for price questions but lacks information about product colors. Adding a targeted question and answer pair injects the missing factual content directly into the knowledge base so the service can return precise answers for color queries without broader changes.

Enable semantic ranking and answer synthesis is not the best choice because semantic ranking and synthesis improve how results are retrieved and composed but they do not create the missing factual content about colors if that information is not present in the knowledge base.

Update source documents and retrain is heavier and usually unnecessary for a simple missing Q and A. Updating source documents and retraining can work if the color data lives only in external files, but adding a new question and answer pair is faster and supports incremental updates.

When a knowledge base answers some topics correctly but not others try adding a specific question and answer pair first before changing retrieval settings or retraining the project.

The correct option is Billing.

The Billing parameter is required because Azure Cognitive Services containers must receive billing or subscription information when they start so usage can be attributed to your Azure subscription. When you run the Anomaly Detector container on a local Docker host you include the billing parameter in the docker run command so the container can validate and report usage to the service.

Mounts is incorrect because mounting volumes provides filesystem access for data and configuration files and it does not convey billing or subscription information to the container.

Fluentd is incorrect because Fluentd relates to log collection and forwarding and it is not used to provide billing credentials or subscription details to the service.

Http Proxy is incorrect because an HTTP proxy setting controls network routing for outbound requests and it does not supply the billing information the container needs to associate usage with a subscription.

When a question asks which Docker parameter gives the service billing or subscription details look for a parameter name or environment variable that mentions Billing or subscription. That is usually the parameter the container uses to attribute usage to an account.

Built in currency entity is correct because a built in currency entity is specifically designed to capture and parse monetary amounts including currency symbols, numeric values, and currency codes.

Built in currency entity normalizes money into separate amount and currency components and it recognizes locale specific formats so you do not need to write fragile regular expressions or custom parsing logic to handle common currency variations.

Numbers prebuilt entity is incorrect because numeric entities capture raw numeric values only and they do not reliably capture currency symbols or currency codes or normalize values as monetary amounts.

Custom entity enforcing two decimal places is incorrect because enforcing two decimal places is brittle and it fails to account for currencies that use different minor unit rules or formats and it still requires extra parsing to identify the currency type which built in entities already handle.

When possible prefer the built in system entities for common types like currency and then test with multiple locales and currency symbols to ensure correct parsing.

The correct answer is No.

Activate customer managed key encryption would not reduce the chance of query throttling because customer managed keys only affect how data is encrypted at rest and how keys are managed and not how the search service allocates compute or query throughput. Throttling occurs when the service reaches its capacity limits for queries and the solution to that problem is to increase capacity rather than change encryption settings.

Add more replicas to the search service is not the correct answer to the specific question about enabling customer managed keys. Adding replicas is in fact a way to reduce throttling because replicas increase query throughput but it does not address the proposed change in the scenario which is encryption key management.

Scale the service to a higher pricing tier is also not the correct choice for this question because the scenario asked whether enabling customer managed keys will fix throttling. Scaling to a higher tier can reduce throttling by providing more resources and higher limits, but it is a different action than the proposed encryption change and so it is not the right answer here.

When a question ties a configuration change to performance first confirm what the feature actually changes and then match that to capacity causes of throttling. Pay attention to words like encryption and throughput because they point to different solution areas.

The correct answer is Upload and index the clip in the portal.

Uploading the clip in the portal is the initial step because the media analyzer needs the file to be ingested and indexed before it can extract insights and metadata. The portal provides the user interface to submit the file and start the indexing job so the service can run speech to text, face and object detection, and other analysis on the content.

Store the video in Blob Storage and grant analyzer access is incorrect because simply placing the file in blob storage does not by itself start the analysis in the portal. Some analyzers can import from storage if configured, but the immediate action to begin processing is to upload or index the clip through the portal or the analyzer API.

Provision a Media Insights resource in the subscription is incorrect because provisioning backend resources is not the first user action to analyze a single clip in the portal. That option refers to setting up infrastructure which may be required in some architectures but it does not replace the need to upload and index the clip to start the analysis.

When a question asks for the initial action focus on the first user step such as uploading or linking the file to the analyzer and not on background provisioning. Remember that uploading and indexing usually initiates media analysis in portal driven services.

the style attribute on the mstts:express-as element is correct.

The the style attribute on the mstts:express-as element selects an expressive speaking style for the voice and enables behaviors that improve clarity and presence in challenging acoustic environments such as road noise and vehicle cabin reverberation. This attribute changes the overall voice delivery and timbre which is more effective for making speech sound natural in a car than only tweaking basic prosody settings.

The the pitch attribute on the prosody element only changes the pitch or fundamental frequency of synthesized speech and does not apply the high level expressive styles needed to compensate for background noise.

The the level attribute on the emphasis element only controls emphasis strength on specific words and it is not intended to set an overall voice style that improves intelligibility in noisy cabins.

The the rate attribute on the prosody element only adjusts speaking speed and can affect timing but it does not produce the expressive voice characteristics that help the assistant sound natural over automotive road noise.

When a question asks about naturalness or robustness in noisy environments prefer SSML elements that set speaking style such as mstts:express-as rather than only changing basic prosody attributes.

Deployment name is correct because each Azure OpenAI REST API request must include the deployment identifier that selects which deployed instance will handle the call.

Azure OpenAI routes requests to a specific deployment and the Deployment name is provided in the request path to identify that deployment. Including the deployment name tells the service which configured model and settings to use for the request.

API key is used for authentication and must be sent in the Authorization header but it does not choose which deployment will be used.

Managed identity credentials can serve as an authentication method for Azure resources in some integrations but they are not a request parameter that specifies a deployment.

Model name refers to the underlying model family but Azure OpenAI requires addressing a deployment resource. The mapping from a model to a deployed instance is configured in Azure and you select the deployment by its name rather than supplying a separate model name in the REST call.

Read the question to see if it asks about selecting the target resource or authenticating the call. In Azure OpenAI the deployment identifies the target while the API key or managed identity provide authentication.

The correct answer is JSON response and an Index value.

The Text Moderation endpoint returns a structured JSON response that lists detected words as terms and it provides an Index value for each term to mark where the match begins in the submitted text. This allows clients to locate and process each detected term precisely.

YAML response and a Confidence score between 0 and 1 is incorrect because the service uses a structured JSON response rather than YAML and the location of a match is indicated by an Index value rather than by a single confidence number used as a position marker.

XML response and a quoted excerpt of the matched text is incorrect because the endpoint does not return XML as its primary format and it reports match positions with an Index value rather than relying solely on quoted excerpts for location data.

Plain text response and a ListId identifying a custom list entry is incorrect because the API returns structured JSON response data instead of plain text and although custom list matches may include identifiers the position is still conveyed with an Index value rather than using a ListId as the location marker.

Binary blob response and a byte offset pointer is incorrect because the moderation service does not return an opaque binary blob for text results and it uses an Index value measured in character positions rather than a byte offset pointer to indicate where a term appears.

When a question asks about API responses look for clues about the response format and field names in the official docs. Pay attention to words like JSON and Index as they often point to the exact field you must choose on the exam.

The correct options are Inclusiveness and Fairness.

Fairness is central to monitoring for equitable outcomes because it focuses on detecting and reducing bias across regions and demographic groups. Monitoring programs that measure disparate impact and track fairness metrics across segments help identify where models produce unequal outcomes and guide corrective actions.

Inclusiveness supports equitable outcomes by ensuring that diverse populations are represented in data collection evaluation and feedback loops. Monitoring that applies inclusive practices checks that underrepresented groups are covered and that models perform acceptably for those groups as well as for the majority.

Reliability and safety is important for overall model behavior and robustness but it does not specifically target equity across demographic groups. Reliability monitoring looks at correctness stability and failure modes rather than group level parity.

Privacy and security are essential for protecting user data and maintaining confidentiality but they do not directly ensure equitable outcomes. Privacy techniques can interact with fairness efforts and require attention when designing monitoring, however the principle itself is not the primary guide for fairness monitoring.

When a question asks about ensuring equitable outcomes look for principles that explicitly address representation and bias, such as fairness and inclusiveness.

The correct option is Custom Brands model.

The Custom Brands model is designed to detect and index brand imagery and to find spoken mentions of companies within video content, which makes it the appropriate choice for analyzing internal virtual meeting recordings to search for images or verbal mentions of rival firms.

The Custom Brands model can be trained with specific logos and brand assets so it can recognize visual appearances of a company and it also supports identifying when brand names are spoken in audio, so it covers both visual and verbal searches required by this scenario.

Custom Slate model is incorrect because it does not focus on identifying brands or spoken brand mentions and it is not intended for brand recognition tasks.

Custom Language model is incorrect because it concentrates on natural language processing tasks and text understanding rather than detecting logos or indexing spoken brand mentions within video and image content.

Custom People model is incorrect because it is intended for recognizing and distinguishing individuals or faces and not for detecting company logos or tracking verbal references to brands.

When a question asks about finding logos or verbal mentions of companies in video choose the option that explicitly references brands or brand detection because those models are built to handle both visual logos and spoken brand names.

All exam questions are from certificationexams.pro and my Azure AI Udemy course.

The correct option is SpeechRecognizer. It is the Azure Speech SDK object that converts caller speech into text with minimal development work.

The SpeechRecognizer class provides ready to use speech to text functionality and simple APIs for single shot and continuous recognition. It handles audio capture and sends audio to the service and returns transcribed text, so developers can get accurate speech to text results with only a few lines of setup code.

TranslationRecognizer is focused on speech translation and delivers translated text in a target language rather than serving as the straightforward speech to text class you would pick for simple transcription.

DialogServiceConnector is intended for connecting to conversational and dialog scenarios and integrates with bot and dialog services. It adds complexity and dialog management features that are not required for basic speech to text.

SpeechSynthesizer performs the opposite task because it converts text into spoken audio. It is the text to speech component and does not transcribe caller speech.

Focus on the core capability phrased in the question. If it asks for speech into text then look for the SDK class that directly implements speech to text and not translation or synthesis. SpeechRecognizer is the simplest choice for quick transcription.