Microsoft Certified: Azure AI Engineer AssociateImplement natural language processing solutionsHard

A company is developing a new voice assistant for smart home devices. The assistant needs to recognize and respond to a wide variety of commands and questions, including custom phrases specific to their device ecosystem (e.g., 'dim the living room ambiance', 'set the coffee maker to brew'). The solution must be highly accurate and provide low-latency responses. Which Azure AI service feature is critical for improving the accuracy of speech recognition for these custom phrases?

  1. AAzure OpenAI Service - Generative AI
  2. BAzure AI Language - Key Phrase Extraction
  3. CAzure AI Speech - Custom Speech
  4. DAzure AI Language - Named Entity Recognition (NER)
Show answer & explanation

Correct answer: C. Azure AI Speech - Custom Speech

For improving speech recognition accuracy for domain-specific vocabulary and custom phrases, Azure AI Speech's Custom Speech feature is critical. It allows training a custom speech model with specific text data (like a list of commands or product names) to better recognize those terms compared to a general-purpose model.

Why the other options are wrong

  • A. Generative AI creates text, it does not enhance speech recognition accuracy for custom terms.
  • B. Key Phrase Extraction works on text, not for improving speech-to-text accuracy.
  • D. NER works on text after transcription to extract entities, it does not improve the transcription itself.

Custom Speech

A feature of Azure AI Speech that allows users to improve speech recognition accuracy for domain-specific vocabulary, accents, or unique phrases by providing custom training data.

  • Trains a speech model with specific text or audio data.
  • Crucial for applications with unique terminology (e.g., medical, technical, product names).
  • Enhances both transcription accuracy and entity recognition from speech.
  • Can use plain text, structured text, or audio + transcript pairs.

Memory trick: Teach the AI your unique words, and it will listen perfectly.

More Implement natural language processing solutions questions