Voice Agent Latency and Turn-Taking Acceptance Test (QA)

Voice Agent Latency and Turn-Taking Acceptance Test (QA)

1. 01

   Freeze the agent build and record the exact model, voice, and prompt version under test

   Any change to the persona prompt or voice model invalidates prior results, so pin the build hash in the test log before the first call.

2. 02

   Assemble a 40-utterance test set that mirrors real caller speech, not scripted demo lines

   Pull 20 utterances from client call transcripts and write 20 adversarial ones: mid-sentence interruptions, background noise, accented speech, and two questions in one breath.

3. 03

   Measure end-of-speech to first-audio latency across 30 consecutive turns

   Log p50 and p95 separately. A p95 above 1.2 seconds reads as hesitation to callers even when the p50 looks healthy.

4. 04

   Score interruption handling by talking over the agent at least ten times per session

   The agent should yield within 300ms and resume the original thread without repeating the full prior sentence. Sindarin exposes turn-taking behavior directly in its API, so its logs can serve as the reference trace.

5. 05

   Verify the agent recovers context after a caller changes topic and returns

   Ask about a billing issue, pivot to hours of operation, then return to billing. If the agent restarts the billing flow from scratch, memory handling fails the gate.

6. 06

   Confirm the escalation trigger fires on the exact phrases and sentiment signals defined in the handoff spec

   Test both explicit requests ('let me talk to a person') and implicit ones (repeated frustration, raised volume). Record false positives as failures, not acceptable noise.

7. 07

   Run the same test set against the client's second channel before signing off

   A voice agent that passes on web chat but fails on phone is a partial deployment. Score each channel independently and report the gap to the client.

8. 08

   Publish a one-page scorecard with pass/fail per criterion and a go-live recommendation

   Include the raw latency numbers and the failing utterances verbatim. Clients approve launches faster when they can see the exact calls that broke.