Implement text analysis solutions 4.2 Implement speech solutions
A contact center wants each recorded support call turned into a transcript that labels…
Study guide skill: Enable multimodal reasoning from audio inputs
Question B5-018
A contact center wants each recorded support call turned into a transcript that labels the agent and the customer, plus a call summary, the call sentiment, and the companies and people mentioned, with as little setup as possible. What should you use?
Select one answer.
Answer: C
The prebuilt-callCenter analyzer analyzes call recordings to produce transcripts with speaker role detection (agent and customer), a call summary, call sentiment, and lists of mentioned companies and people, without creating an analyzer setup. Batch transcription is documented for transcribing large amounts of stored audio, so the summary, sentiment, role, and entity steps would still have to be built separately. A custom analyzer would work but needs more setup than the prebuilt one, and speech translation is for translating audio streams.
Deep dive
How it works
Content Understanding's audio analyzers transcribe conversations, separate speakers and then pull structured fields such as summaries, sentiment and key topics out of the transcript. Built-in preprocessing (transcription, diarization and role detection) gives the generative models context, so fields like "call summary" come from understanding the whole call.
For contact centers, there's a prebuilt analyzer, which needs no analyzer setup: prebuilt-callCenter (post-call analysis). For each call recording it generates:
- A conversation transcript with speaker role detection (agent and customer)
- A call summary
- Call sentiment
- The top five articles mentioned
- Lists of the companies and people (name and title or role) mentioned
- A list of relevant call categories
That's every output the question asks for, with as little setup as possible. Prebuilt analyzers use multilingual transcription and have returnDetails enabled, and transcripts come out in WEBVTT format separated by speaker. If you need extra fields later, you can customize a suggested analyzer template instead of starting from scratch.
The other options, and when you'd use them
- A custom audio analyzer built from scratch: Field extraction lets you define any schema, so this would work, but you'd design every field yourself. The prebuilt analyzer already covers these outputs.
- Speech batch transcription with diarization: Batch transcription is for transcribing large amounts of audio in storage, asynchronously. You'd get transcripts, but the summary, sentiment, role labels and entity lists would still have to be built separately.
- Real-time speech translation in the Speech SDK: Speech translation does real-time, multi-language translation of audio streams. It doesn't summarize calls or detect roles.
Exam tip
"Recorded calls plus summary, sentiment, roles and entities with minimal setup" means Content Understanding prebuilt-callCenter. "Just transcripts of lots of stored audio" means batch transcription.
References
Last reviewed: