AI-103 Practice Question 89: Custom Speech Phrase List
Microsoft Azure AI Apps and Agents Developer Associate (AI-103) · Test 9 · question 9 of 10
A pharmaceutical support line transcribes calls in real time. Everyday conversation transcribes accurately, but drug names and internal SKU codes are consistently misheard; "Xarelto" comes back as unrelated words that sound similar. New drug names are added to the formulary most weeks, and the team wants a fix it can update the same day, without collecting hours of labeled call audio or training and redeploying a model.
What should they use?
Show answer and explanation
Answer:
A phrase list is a real-time accuracy feature that "does not need model training"; you supply a custom vocabulary of expected words or phrases, and the recognizer favors them over acoustically similar alternatives during recognition, even mid-utterance. Because it applies instantly to the next request, Microsoft recommends it specifically "when the audio context is often changing," which matches a formulary that's updated weekly. Its weight can be raised (0.0–2.0) to increase the likelihood the listed terms win over similar-sounding words.
- A is wrong because a trained Custom Speech model needs a labeled audio and transcript dataset and a deploy step every time it's updated, exactly the overhead the team wants to avoid for a same-day vocabulary change. - B is wrong because fast transcription is a latency mode: it changes how quickly a completed audio file's transcript comes back, not which words the model favors during recognition. It does nothing for the drug-name accuracy problem. - C is wrong because silence timeouts control where the recognizer decides an utterance ends; they don't influence which specific words get recognized within an utterance.