Implement text analysis solutions 4.2 Implement speech solutions

A voice agent built on Azure Speech real-time transcription keeps mishearing a few dozen…

Study guide skill: Integrate speech as an agent modality, including custom speech models

Question B5-016

A voice agent built on Azure Speech real-time transcription keeps mishearing a few dozen customer names and product acronyms, and the team wants a fix today without training a model. The company also transcribes archived calls with the Batch transcription API. Which two statements are correct? (Choose two.)

Select 2 answers.

  • A Phrase lists work only with custom speech endpoints, so a custom model must be deployed first
  • B Phrase lists must be uploaded and trained like custom speech data, which takes several hours
  • C Set the phrase list weight to 0.0 to give the phrases maximum influence
  • D Add the names and acronyms to a phrase list at runtime; no model training is needed ✓ Correct answer
  • E The phrase list won't help the archived calls, because the Batch transcription API doesn't support phrase lists ✓ Correct answer

Answer: D, E

A phrase list is a runtime recognition feature: you provide it just before recognition starts, it needs no training, and it works with both base and custom speech endpoints. It's supported for real-time and fast transcription but not by the Batch transcription API. In the Speech SDK, a weight of 0.0 disables the phrase list and 2.0 gives it maximum influence.

Deep dive

How it works

A phrase list gives the recognizer a list of words to expect: names, places, homonyms, and words or acronyms unique to your organization. It's the fast fix for a few dozen misheard terms:

  • Just-in-time: you provide it just before recognition starts, so you don't need to train a custom model.
  • Lightweight: no large dataset. A word or phrase is enough to boost recognition.
  • Any endpoint: it works with both base (standard) and custom speech endpoints.

Where it works matters for the second statement. Phrase lists are supported with real-time transcription and the fast transcription API. The Batch transcription API doesn't support phrase lists, so the archived calls won't benefit. If batch accuracy for those terms matters, the usual path is custom speech training.

With the Speech SDK in real-time transcription you can tune the weight, from 0.0 to 2.0. At 0.0 the list is disabled, 1.0 is the default, and 2.0 is maximum influence. Start at the default and raise it only if terms are still missed.

The other options, and when you'd use them

  • Only with custom speech endpoints: Phrase lists work with both base and custom speech endpoints. No custom model is needed.
  • Uploaded and trained like custom speech data: You apply a phrase list at runtime and it doesn't require model training.
  • Weight 0.0 for maximum influence: Backwards. 0.0 disables the phrase list; 2.0 is the maximum weight.

Exam tip

Phrase list = runtime, no training, base or custom endpoints, real-time and fast transcription only (not batch). Weight 0.0 off, 1.0 default, 2.0 max.

References

Last reviewed:

Ready to test yourself? The quiz asks this and every other AI-103 question, with instant feedback. Practice this in a quiz →

All Text analysis questions · AI-103 study guide

Thanks for your feedback!