Microsoft AI-103 (Developing AI Apps and Agents on Azure) Practice Test 44

Question 430

An architecture board is assessing 2 separate deliverables: (1) transcribe thousands of stored recordings without needing low-latency partial output; and (2) improve recognition of a small, frequently misheard set of product names without a full retraining cycle. Which distinct implementations meet these deliverables? Choose TWO.

  1. Handle noisy input with confidence-aware fallbacks
  2. Use speech-to-text recognition for voice input
  3. Use low-latency streaming recognition
  4. Adapt speech recognition to domain vocabulary
  5. Use batch transcription for recorded archives
  6. Provide phrase hints where supported

Explanation

Correct Answers

E:

For the stated requirement to transcribe thousands of stored recordings without needing low-latency partial output, batch transcription processes prerecorded audio asynchronously and supports large backlogs more naturally than interactive streaming; this is the mechanism that the system must implement. This choice satisfies the specific deliverable to transcribe thousands of stored recordings without needing low-latency partial output, while the other deliverables require separate capabilities.

F:

For the stated requirement to improve recognition of a small, frequently misheard set of product names without a full retraining cycle, phrase hints bias decoding toward supplied vocabulary in compatible recognition modes; they cannot guarantee correct recognition; this is the mechanism that the system must implement. This choice satisfies the specific deliverable to improve recognition of a small, frequently misheard set of product names without a full retraining cycle, while the other deliverables require separate capabilities.

Incorrect Answers

A:

For a system whose objective is to avoid acting on a guessed spoken account number when audio quality is poor, confidence-aware confirmation and acoustic preprocessing can reduce high-impact recognition errors; guessing a missing digit is unsafe; this does not directly meet the required outcome to transcribe thousands of stored recordings without needing low-latency partial output.

B:

For a system whose objective is to convert customer spoken questions into text that an agent can route and answer, automatic speech recognition transcribes spoken audio to text; it does not itself understand the full business intent; this does not directly meet the required outcome to transcribe thousands of stored recordings without needing low-latency partial output.

C:

For a system whose objective is to display live captions as someone speaks rather than waiting until the recording ends, streaming recognition provides interim and final text segments; handling revisions to interim hypotheses is part of the UI design; this does not directly meet the required outcome to transcribe thousands of stored recordings without needing low-latency partial output.

D:

For a system whose objective is to recognize specialized clinical terms consistently in dictated notes, custom Speech can adapt recognition to domain-specific vocabulary and speech conditions using appropriate training data; this does not directly meet the required outcome to transcribe thousands of stored recordings without needing low-latency partial output.

Question 431

The engineering change request is to use text-to-speech synthesis for spoken replies. What is the most direct acceptance criterion?

  1. Measure translation accuracy and end-to-end delay during live speech
  2. Track confirmation rate and incorrect-action prevention on noisy test calls
  3. Evaluate intelligibility, pronunciation, and audio generation latency
  4. Measure word error rate on target accents and noisy telephone audio

Explanation

Correct Answer

C:

The accepted change is to use text-to-speech synthesis for spoken replies; this evaluation checks whether it can read an approved agent response aloud with a supported voice instead of returning only text, since text-to-speech converts response text into audio with chosen language and voice parameters; generation of the answer happens separately.

Incorrect Answers

A:

This test concerns the ability to allow participants speaking different languages to receive translated speech or text during a live call, for which speech translation transcribes and translates streaming audio with supported target languages; post-call batch translation misses the live requirement; it does not establish whether the change to use text-to-speech synthesis for spoken replies works for the current release.

B:

This test concerns the ability to avoid acting on a guessed spoken account number when audio quality is poor, for which confidence-aware confirmation and acoustic preprocessing can reduce high-impact recognition errors; guessing a missing digit is unsafe; it does not establish whether the change to use text-to-speech synthesis for spoken replies works for the current release.

D:

This test concerns the ability to convert customer spoken questions into text that an agent can route and answer, for which automatic speech recognition transcribes spoken audio to text; it does not itself understand the full business intent; it does not establish whether the change to use text-to-speech synthesis for spoken replies works for the current release.

Question 432

A system now includes a component to enable speaker diarization for a meeting transcript. Which evaluation should accompany the rollout?

  1. Score speaker attribution error and segment boundary accuracy on annotated meetings
  2. Evaluate intelligibility, pronunciation, and audio generation latency
  3. Measure translation accuracy and end-to-end delay during live speech
  4. Evaluate acoustic-event recognition on annotated recordings with few or no spoken words

Explanation

Correct Answer

A:

The accepted change is to enable speaker diarization for a meeting transcript; this evaluation checks whether it can attribute remarks to different speakers in a shared audio recording, since diarization associates recognized speech segments with speaker identities or labels; plain transcription alone may merge speakers.

Incorrect Answers

B:

This test concerns the ability to read an approved agent response aloud with a supported voice instead of returning only text, for which text-to-speech converts response text into audio with chosen language and voice parameters; generation of the answer happens separately; it does not establish whether the change to enable speaker diarization for a meeting transcript works for the current release.

C:

This test concerns the ability to allow participants speaking different languages to receive translated speech or text during a live call, for which speech translation transcribes and translates streaming audio with supported target languages; post-call batch translation misses the live requirement; it does not establish whether the change to enable speaker diarization for a meeting transcript works for the current release.

D:

This test concerns the ability to answer a question about the tone and acoustic events of a recording rather than only its written words, for which audio understanding can use nontextual acoustic information when a supported multimodal model accepts audio; a transcript loses these cues; it does not establish whether the change to enable speaker diarization for a meeting transcript works for the current release.

Question 433

Before approving the proposed action to translate spoken input into multilingual agent output, which measurable acceptance check should be required?

  1. Measure word error rate on target accents and noisy telephone audio
  2. Evaluate acoustic-event recognition on annotated recordings with few or no spoken words
  3. Test semantic equivalence and end-to-end response language on bilingual call scenarios
  4. Measure timestamp offset between retrieved quotes and their audio spans

Explanation

Correct Answer

C:

The accepted change is to translate spoken input into multilingual agent output; this evaluation checks whether it can accept a spoken request in one language and respond aloud in a second language, since an integrated recognition, translation/reasoning, and synthesis workflow handles cross-language conversations; each stage needs compatible language support.

Incorrect Answers

A:

This test concerns the ability to convert customer spoken questions into text that an agent can route and answer, for which automatic speech recognition transcribes spoken audio to text; it does not itself understand the full business intent; it does not establish whether the change to translate spoken input into multilingual agent output works for the current release.

B:

This test concerns the ability to answer a question about the tone and acoustic events of a recording rather than only its written words, for which audio understanding can use nontextual acoustic information when a supported multimodal model accepts audio; a transcript loses these cues; it does not establish whether the change to translate spoken input into multilingual agent output works for the current release.

D:

This test concerns the ability to enable an agent to cite the exact location of a spoken statement in an audio archive, for which time-aligned transcription connects text tokens or segments to audio positions, supporting later playback and source verification; it does not establish whether the change to translate spoken input into multilingual agent output works for the current release.

Question 434

The team plans to preserve timestamps for searchable voice records. Which test provides the clearest evidence that this change works?

  1. Track confirmation rate and incorrect-action prevention on noisy test calls
  2. Evaluate acoustic-event recognition on annotated recordings with few or no spoken words
  3. Measure timestamp offset between retrieved quotes and their audio spans
  4. Measure word error rate on target accents and noisy telephone audio

Explanation

Correct Answer

C:

The accepted change is to preserve timestamps for searchable voice records; this evaluation checks whether it can enable an agent to cite the exact location of a spoken statement in an audio archive, since time-aligned transcription connects text tokens or segments to audio positions, supporting later playback and source verification.

Incorrect Answers

A:

This test concerns the ability to avoid acting on a guessed spoken account number when audio quality is poor, for which confidence-aware confirmation and acoustic preprocessing can reduce high-impact recognition errors; guessing a missing digit is unsafe; it does not establish whether the change to preserve timestamps for searchable voice records works for the current release.

B:

This test concerns the ability to answer a question about the tone and acoustic events of a recording rather than only its written words, for which audio understanding can use nontextual acoustic information when a supported multimodal model accepts audio; a transcript loses these cues; it does not establish whether the change to preserve timestamps for searchable voice records works for the current release.

D:

This test concerns the ability to convert customer spoken questions into text that an agent can route and answer, for which automatic speech recognition transcribes spoken audio to text; it does not itself understand the full business intent; it does not establish whether the change to preserve timestamps for searchable voice records works for the current release.

Question 435

A customer-facing AI service is being scoped for production. The non-negotiable task is to read an approved agent response aloud with a supported voice instead of returning only text. Which implementation fits?

  1. Use text-to-speech synthesis for spoken replies
  2. Use speech-to-text recognition for voice input
  3. Use low-latency streaming recognition
  4. Use real-time speech translation

Explanation

Correct Answer

A:

For the stated requirement to read an approved agent response aloud with a supported voice instead of returning only text, text-to-speech converts response text into audio with chosen language and voice parameters; generation of the answer happens separately; this is the mechanism that the system must implement.

Incorrect Answers

B:

For a system whose objective is to convert customer spoken questions into text that an agent can route and answer, automatic speech recognition transcribes spoken audio to text; it does not itself understand the full business intent; this does not directly meet the required outcome to read an approved agent response aloud with a supported voice instead of returning only text.

C:

For a system whose objective is to display live captions as someone speaks rather than waiting until the recording ends, streaming recognition provides interim and final text segments; handling revisions to interim hypotheses is part of the UI design; this does not directly meet the required outcome to read an approved agent response aloud with a supported voice instead of returning only text.

D:

For a system whose objective is to allow participants speaking different languages to receive translated speech or text during a live call, speech translation transcribes and translates streaming audio with supported target languages; post-call batch translation misses the live requirement; this does not directly meet the required outcome to read an approved agent response aloud with a supported voice instead of returning only text.

Question 436

An engineering team has a requirement to apply a proprietary extraction algorithm that the built-in search skillset cannot express. Which capability should be in the initial design?

  1. Project chunks to a secondary retrieval index
  2. Use a custom skill for specialized enrichment
  3. Use layout-aware extraction for tables and headings
  4. Add an image-description enrichment stage

Explanation

Correct Answer

B:

For the stated requirement to apply a proprietary extraction algorithm that the built-in search skillset cannot express, custom Web API skills can call external enrichment code and return structured outputs to an indexer pipeline; this is the mechanism that the system must implement.

Incorrect Answers

A:

For a system whose objective is to search fine-grained passages while preserving links to complete source documents, index projections can place chunk records in a retrieval index with parent keys so answers can link back to source files; this does not directly meet the required outcome to apply a proprietary extraction algorithm that the built-in search skillset cannot express.

C:

For a system whose objective is to preserve relationships between headings, cells and surrounding text during document ingestion, layout processing detects document structure and table boundaries beyond simple plain-text OCR; this does not directly meet the required outcome to apply a proprietary extraction algorithm that the built-in search skillset cannot express.

D:

For a system whose objective is to index meaningful visual content of diagrams that contains no searchable text, image captioning or multimodal understanding produces descriptors for visual semantics; OCR alone reads only visible characters; this does not directly meet the required outcome to apply a proprietary extraction algorithm that the built-in search skillset cannot express.

Question 437

A system now includes a component to add an image-description enrichment stage. Which evaluation should accompany the rollout?

  1. Check table-cell relationships and heading attribution on representative document queries
  2. Measure retrieval accuracy on visually described equipment and diagram questions
  3. Check indexed vector coverage and embedding failures by document batch
  4. Test normalized-field accuracy and failed-enrichment retry behavior

Explanation

Correct Answer

B:

The accepted change is to add an image-description enrichment stage; this evaluation checks whether it can index meaningful visual content of diagrams that contains no searchable text, since image captioning or multimodal understanding produces descriptors for visual semantics; OCR alone reads only visible characters.

Incorrect Answers

A:

This test concerns the ability to preserve relationships between headings, cells and surrounding text during document ingestion, for which layout processing detects document structure and table boundaries beyond simple plain-text OCR; it does not establish whether the change to add an image-description enrichment stage works for the current release.

C:

This test concerns the ability to avoid writing a custom service for every supported text-to-vector conversion, for which integrated vectorization uses an indexer and enrichment pipeline to chunk and embed content; query-side vectorization still needs compatible settings; it does not establish whether the change to add an image-description enrichment stage works for the current release.

D:

This test concerns the ability to apply a proprietary extraction algorithm that the built-in search skillset cannot express, for which custom Web API skills can call external enrichment code and return structured outputs to an indexer pipeline; it does not establish whether the change to add an image-description enrichment stage works for the current release.

Question 438

An operator can reproduce this error: Entire manuals are embedded as one giant field and answer retrieval frequently returns irrelevant sections. Which targeted repair is justified?

  1. Configure an indexer and scheduled refresh for the supported source connection
  2. Add a supported OCR or document-reading enrichment step before chunking
  3. Use an appropriate text split strategy with contextual metadata and modest overlap
  4. Configure chunk-level index projections with parent identifiers and source metadata

Explanation

Correct Answer

C:

The fault signal is that entire manuals are embedded as one giant field and answer retrieval frequently returns irrelevant sections; the targeted fix works because chunking creates smaller units for indexing and grounding; titles and parent identifiers should accompany chunks, restoring the ability to keep retrieval passages within embedding and prompt size limits without losing section context.

Incorrect Answers

A:

If the failing function were to keep search content synchronized with periodically changing files in a connected repository, this change could be relevant because Azure AI Search indexers read supported sources and drive enrichment and indexing on a schedule; external custom sources may need a push path; the actual symptom is that entire manuals are embedded as one giant field and answer retrieval frequently returns irrelevant sections, which calls for a different repair.

B:

If the failing function were to make text from photographed forms available for retrieval and citation, this change could be relevant because OCR converts visible characters into machine-readable text; text splitters cannot recover words from image-only pages; the actual symptom is that entire manuals are embedded as one giant field and answer retrieval frequently returns irrelevant sections, which calls for a different repair.

D:

If the failing function were to search fine-grained passages while preserving links to complete source documents, this change could be relevant because index projections can place chunk records in a retrieval index with parent keys so answers can link back to source files; the actual symptom is that entire manuals are embedded as one giant field and answer retrieval frequently returns irrelevant sections, which calls for a different repair.

Question 439

A solution architect is deciding how to make text from photographed forms available for retrieval and citation. Which Azure AI pattern addresses that task?

  1. Use a custom skill for specialized enrichment
  2. Attach retrieved source identifiers to agent tools
  3. Apply OCR to scanned document images
  4. Add an image-description enrichment stage

Explanation

Correct Answer

C:

For the stated requirement to make text from photographed forms available for retrieval and citation, OCR converts visible characters into machine-readable text; text splitters cannot recover words from image-only pages; this is the mechanism that the system must implement.

Incorrect Answers

A:

For a system whose objective is to apply a proprietary extraction algorithm that the built-in search skillset cannot express, custom Web API skills can call external enrichment code and return structured outputs to an indexer pipeline; this does not directly meet the required outcome to make text from photographed forms available for retrieval and citation.

B:

For a system whose objective is to make an agent's grounded answer traceable back to the correct document chunk, a retrieval tool should return source identity and relevant snippet metadata along with content, not text alone; this does not directly meet the required outcome to make text from photographed forms available for retrieval and citation.

D:

For a system whose objective is to index meaningful visual content of diagrams that contains no searchable text, image captioning or multimodal understanding produces descriptors for visual semantics; OCR alone reads only visible characters; this does not directly meet the required outcome to make text from photographed forms available for retrieval and citation.

Question 440

An Azure AI delivery has 3 independently testable requirements: (1) rank both exact internal policy codes and semantically related explanations; (2) support both full-text answers and exact filtering by department and policy year; and (3) promote passages that directly answer questions over textually similar but less useful chunks. Select the most directly specialized implementation for each requirement. Choose THREE.

  1. Match index vector dimensions to embedding model outputs
  2. Attach retrieved source identifiers to agent tools
  3. Use layout-aware extraction for tables and headings
  4. Run a hybrid keyword-plus-vector query
  5. Define searchable and filterable fields appropriately
  6. Add semantic reranking for answer relevance

Explanation

Correct Answers

D:

For the stated requirement to rank both exact internal policy codes and semantically related explanations, hybrid retrieval combines text search and vector retrieval, merging candidates using reciprocal-rank fusion; this is the mechanism that the system must implement. This choice satisfies the specific deliverable to rank both exact internal policy codes and semantically related explanations, while the other deliverables require separate capabilities.

E:

For the stated requirement to support both full-text answers and exact filtering by department and policy year, Azure AI Search index fields must be configured for the operations used; searchable text differs from structured filterable metadata; this is the mechanism that the system must implement. This choice satisfies the specific deliverable to support both full-text answers and exact filtering by department and policy year, while the other deliverables require separate capabilities.

F:

For the stated requirement to promote passages that directly answer questions over textually similar but less useful chunks, semantic ranking reranks a first-stage candidate list using language understanding; it cannot rescue documents never retrieved; this is the mechanism that the system must implement. This choice satisfies the specific deliverable to promote passages that directly answer questions over textually similar but less useful chunks, while the other deliverables require separate capabilities.

Incorrect Answers

A:

For a system whose objective is to ensure the index's vector field accepts embeddings produced by the selected model, an Azure AI Search vector field must have dimensions compatible with the embedding output and a valid vector search configuration; mismatched vectors fail at indexing or query time; this does not directly meet the required outcome to rank both exact internal policy codes and semantically related explanations.

B:

For a system whose objective is to make an agent's grounded answer traceable back to the correct document chunk, a retrieval tool should return source identity and relevant snippet metadata along with content, not text alone; this does not directly meet the required outcome to rank both exact internal policy codes and semantically related explanations.

C:

For a system whose objective is to preserve relationships between headings, cells and surrounding text during document ingestion, layout processing detects document structure and table boundaries beyond simple plain-text OCR; this does not directly meet the required outcome to rank both exact internal policy codes and semantically related explanations.

Developing AI Apps and Agents on Azure (AI-103) Practice Tests